Effective Human-Robot Interaction (HRI) is crucial for enhancing accessibility and usability in real-world robotics applications. However, existing solutions often rely on gesture-only or language-only commands, making interaction inefficient and ambiguous, particularly for users with physical impairments. In this paper, we introduce FAM-HRI, an efficient multimodal framework for HRI that integrates language and gaze inputs via foundation models. By leveraging lightweight Meta ARIA glasses, our system captures real-time multimodal signals and utilizes large language models (LLMs) to fuse user intention with scene context, enabling intuitive and precise robot manipulation. Our method accurately determines the gaze fixation time interval, reducing noise caused by the gaze dynamic nature. Experimental evaluations demonstrate that FAM-HRI achieves a high success rate in task execution while maintaining a low interaction time, providing a practical solution for individuals with limited physical mobility or motor impairments. To support the community, we have released our system design, algorithms, and solutions at https://github.com/laiyuzhi/FAM-HRI.

Open-world 3D semantic occupancy prediction aims to generate a voxelized 3D representation from sensor inputs while recognizing both known and unknown objects. Transferring open-vocabulary knowledge from vision-language models (VLMs) offers a promising direction but remains challenging. However, methods based on VLM-derived 2D pseudo-labels with traditional supervision are limited by a predefined label space and lack general prediction capabilities. Direct alignment with pretrained image embeddings, on the other hand, often fails to achieve reliable performance because of inconsistent image and text representations in VLMs. To address these challenges, we propose AGO, a novel 3D occupancy prediction framework with adaptive grounding to handle diverse open-world scenarios. AGO first encodes surrounding images and class prompts into 3D and text embeddings, respectively, leveraging similarity-based grounding training with 3D pseudo-labels. Additionally, a modality adapter maps 3D embeddings into a space aligned with VLM-derived image embeddings, reducing modality gaps. Experiments on Occ3D-nuScenes show that AGO improves unknown object prediction in zero-shot and few-shot transfer while achieving state-of-the-art closed-world self-supervised performance, surpassing prior methods by 4.09 mIoU. Code is available at: https://github.com/EdwardLeeLPZ/AGO.

Stable and robust robotic grasping is essential for current and future robot applications. In recent works, the use of large datasets and supervised learning has enhanced speed and precision in antipodal grasping. However, these methods struggle with perception and calibration errors due to large planning horizons. To obtain more robust and reactive grasping motions, leveraging reinforcement learning combined with tactile sensing is a promising direction. Yet, there is no systematic evaluation of how the complexity of force-based tactile sensing affects the learning behavior for grasping tasks. This paper compares various tactile and environmental setups using two model-free reinforcement learning approaches for antipodal grasping. Our findings suggest that under imperfect visual perception, various tactile features improve learning outcomes, while complex tactile inputs complicate training.

Neuromorphic Computing (NC) and Spiking Neural Networks (SNNs) in particular are often viewed as the next generation of Neural Networks (NNs). NC is a novel bio-inspired paradigm for energy efficient neural computation, often relying on SNNs in which neurons communicate via spikes in a sparse, event-based manner. This communication via spikes can be exploited by neuromorphic hardware implementations very effectively and results in a drastic reductions of power consumption and latency in contrast to regular GPU-based NNs. In recent years, neuromorphic hardware has become more accessible, and the support of learning frameworks has improved. However, available hardware is partially still experimental, and it is not transparent what these solutions are effectively capable of, how they integrate into real-world robotics applications, and how they realistically benefit energy efficiency and latency. In this work, we provide the robotics research community with an overview of what is possible with SNNs on neuromorphic hardware focusing on real-time processing. We introduce a benchmark of three popular neuromorphic hardware devices for the task of event-based object detection. Moreover, we show that an SNN on a neuromorphic hardware is able to run in a challenging table tennis robot setup in real-time.

Prior knowledge vastly exists in the automation industry, especially for tasks like pick-and-place, where simple programmatic demonstrations with online generation ability can be acquired easily. How to learn a policy faster with higher flexibility and generalization ability based on these demonstrations is a question to be answered. End-to-end target learning and imitation learning are widely discussed in previous works. Here, we focus on the online generation ability of the demonstration and propose a demo injection method based on actor-critic off-policy reinforcement learning (RL) for the interaction and policy optimization phase. We conduct experiments and an ablation study based on four research questions around a two-finger reactive grasping task with a Panda robot. The result shows our proposed injection method increases the training stability, strongly reduces the time to convergence and benefits sim-2-real transfer with smooth motion.

Accurate calibration is crucial for using multiple cameras to triangulate the position of objects precisely. However, it is also a time-consuming process that needs to be repeated for every displacement of the cameras. The standard approach is to use a printed pattern with known geometry to estimate the intrinsic and extrinsic parameters of the cameras. The same idea can be applied to event-based cameras, though it requires extra work. By using frame reconstruction from events, a printed pattern can be detected. A blinking pattern can also be displayed on a screen. Then, the pattern can be directly detected from the events. Such calibration methods can provide accurate intrinsic calibration for both frame- and event-based cameras. However, using 2D patterns has several limitations for multi-camera extrinsic calibration, with cameras possessing highly different points of view and a wide baseline. The 2D pattern can only be detected from one direction and needs to be of significant size to compensate for its distance to the camera. This makes the extrinsic calibration time-consuming and cumbersome. To overcome these limitations, we propose eWand, a new method that uses blinking LEDs inside opaque spheres instead of a printed or displayed pattern. Our method provides a faster, easier-to-use extrinsic calibration approach that maintains high accuracy for both event- and frame-based cameras.

Real-world autonomous driving systems must make safe decisions in the face of rare and diverse traffic scenarios. Current state-of-the-art planners are mostly evaluated on real-world datasets like nuScenes (open-loop) or nuPlan (closed-loop). In particular nuPlan seems to be an expressive evaluation method since it is based on real-world data and closed-loop, yet it mostly covers basic driving scenarios. This makes it difficult to judge a planner’s capabilities to generalize to rarely-seen situations. Therefore, we propose a novel closed-loop benchmark interPlan containing several edge cases and challenging driving scenarios. We assess existing state-of-the-art planners on our benchmark and show that neither rule-based nor learning-based planners can safely navigate the interPlan scenarios.A recently evolving direction is the usage of foundation models like large language models (LLM) to handle generalization. We evaluate an LLM-only planner and introduce a novel hybrid planner that combines an LLM-based behavior planner with a rule-based motion planner that achieves state-of-the-art performance on our benchmark.

The expanding use of automated vision, assistance systems, and augmented reality applications in marine settings calls for reliable and accurate horizon detection and locking. Traditional methods utilizing Inertial Measurement Units (IMU) or feature-based computer vision techniques often yield inconsistent results, particularly when unmanned surface vehicles or boats are subject to high-speed movement or choppy waters. Addressing this, our work introduces a computer vision (CV)-based solution for real-time horizon locking. Employing real-time semantic segmentation, we accurately differentiate between sky, land or water in the frame, enabling computational locking of the horizon’s position. This stable visual reference significantly improves the performance and reliability of on-board systems for autonomous navigation, augmented reality overlays, and multi-object tracking. Supported by a dataset collected under various marine conditions, our method has proven to achieve high accuracy with low computational latency, making it a promising avenue for wide-scale implementation on automated and semi-automated systems.

Predicting the future motion of surrounding vehicles is a crucial enabler for safe autonomous driving. The field of motion prediction has seen large progress recently with State-of-the-Art (SotA) models achieving impressive results on large-scale public benchmarks. However, recent work revealed that learning-based methods are prone to predict off-road trajectories in challenging scenarios. These can be created by perturbing existing scenarios with additional turns in front of the target vehicle while the motion history is left unchanged. We argue that this indicates that SotA models do not consider the map information sufficiently and demonstrate how this can be solved by representing the model inputs and outputs in a Frenet frame defined by lane centreline sequences. To this end, we present a general wrapper that leverages a Frenet representation of the scene, and that can be applied to SotA models without changing their architecture. We demonstrate the effectiveness of this approach in a comprehensive benchmark comprising two SotA motion prediction models. Our experiments show that this reduces the off-road rate in challenging scenarios by more than 90%, without sacrificing average performance. Code and supplementary material are available under: https://mh0797.github.io/stayontrack/.

For successful deployment of robots in multifaceted situations, an understanding of the robot for its environment is indispensable. With advancing performance of state-of-the-art object detectors, the capability of robots to detect objects within their interaction domain is also enhancing. However, it binds the robot to a few trained classes and prevents it from adapting to unfamiliar surroundings beyond predefined scenarios. In such scenarios, humans could assist robots amidst the overwhelming number of interaction entities and impart the requisite expertise by acting as teachers. We propose a novel pipeline that effectively harnesses human gaze and augmented reality in a human-robot collaboration context to teach a robot novel objects in its surrounding environment. By intertwining gaze (to guide the robot's attention to an object of interest) with augmented reality (to convey the respective class information) we enable the robot to quickly acquire a significant amount of automatically labeled training data on its own. Training in a transfer learning fashion, we demonstrate the robot's capability to detect recently learned objects and evaluate the influence of different machine learning models and learning procedures as well as the amount of training data involved. Our multimodal approach proves to be an efficient and natural way to teach the robot novel objects based on a few instances and allows it to detect classes for which no training dataset is available. In addition, we make our dataset publicly available to the research community, which consists of RGB and depth data, intrinsic and extrinsic camera parameters, along with regions of interest.

We develop a graph convolutional neural network for power system contingency analysis. In contrast to other methods, the proposed architecture is purely data-driven and does not require knowledge of the power grid’s underlying topology. Instead, the estimation of multiple correlation-based graphs enables a pinpoint exploitation of various power system intrinsic structures. The architecture is tested on two large real-world type power grids containing over 6000 approximated output variables. The evaluation shows that the proposed method requires only a fraction of the training parameters to still perform significantly better than the baseline methods, especially when only few training samples are available.

Unmanned aerial vehicles assist in maritime search and rescue missions by flying over large search areas to autonomously search for objects or people. Reliably detecting objects of interest requires fast models to employ on embedded hardware. Moreover, with increasing distance to the ground station only part of the video data can be transmitted. In this work, we consider the problem of finding meaningful region of interest proposals in a video stream on an embedded GPU. Current object or anomaly detectors are not suitable due to their slow speed, especially on limited hardware and for large image resolutions. Lastly, objects of interest, such as pieces of wreckage, are often not known a priori. Therefore, we propose an end-to-end future frame prediction model running in real-time on embedded GPUs to generate region proposals. We analyze its performance on large-scale maritime data sets and demonstrate its benefits over traditional and modern methods.

In dynamic environments, robotic manipulators and especially cobots must be able to react to changing circumstances while in motion. This substantiates the need for quick trajectory planning algorithms that are able to cope with arbitrary velocity and acceleration boundary conditions. Apart from dynamic re-planning, being able to seamlessly join trajectories together opens the door for divide-and-conquer-type algorithms to focus on the individual parts of a motion separately. While geodesic motion planning has proven that it can produce very smooth and efficient actuator movement, the problem of incorporating non-zero boundary conditions has not been addressed yet. We show how a set of generalized coordinates can be used to transition between boundary conditions and free movement in an optimal way while still retaining the known advantages of geodesic planners. We also outline, how our approach can be combined with the family of time-scaling algorithms for further improvement of the generated trajectories.

Event cameras are becoming increasingly popular in robotics and computer vision due to their beneficial properties, e.g., high temporal resolution, high bandwidth, almost no motion blur, and low power consumption. However, these cameras remain expensive and scarce in the market, making them inaccessible to the majority. Using event simulators minimizes the need for real event cameras to develop novel algorithms. However, due to the computational complexity of the simulation, the event streams of existing simulators cannot be generated in real-time but rather have to be pre-calculated from existing video sequences or pre-rendered and then simulated from a virtual 3D scene. Although these offline generated event streams can be used as training data for learning tasks, all response time dependent applications cannot benefit from these simulators yet, as they still require an actual event camera. This work proposes simulation methods that improve the performance of event simulation by two orders of magnitude (making them real-time capable) while remaining competitive in the quality assessment.

Recently, end-to-end deep networks based stereo matching methods, mainly because of their performance, have gained popularity. However, this improvement in performance comes at the cost of increased computational and memory bandwidth requirements, thus necessitating specialized hardware (GPUs); even then, these methods have large inference times compared to classical methods. This limits their applicability in real-world applications. Although we desire high accuracy stereo methods albeit with reasonable inference time. To this end, we propose a fast end-to-end stereo matching method. Majority of this speedup comes from integrating a leaner backbone. To recover the performance lost because of a leaner backbone, we propose to use learned attention weights based cost volume combined with LogL1 loss for stereo matching. Using LogL1 loss not only improves the overall performance of the proposed network but also leads to faster convergence. We do a detailed empirical evaluation of different design choices and show that our method requires$4\times$less operations and is also about 9 to$14\times$faster compared to the state of the art methods like ACVNet [1], LEAStereo [2] and CFNet [3] while giving comparable performance11Code: https://github.com/cogsys-tuebingen/LeanStereo

Knowledge distillation has been quite popular in vision for tasks like classification and segmentation however not much work has been done for distilling state-of-the-art stereo matching methods despite their range of applications. One of the reasons for its lack of use in stereo matching networks is due to the inherent complexity of these networks, where a typical network is composed of multiple two- and three-dimensional modules. In this work, we systematically combine the insights from state-of-the-art stereo methods with general knowledge-distillation techniques to develop a joint framework for stereo networks distillation with competitive results and faster inference. Moreover, we show, via a detailed empirical analysis, that distilling knowledge from the stereo network requires careful design of the complete distillation pipeline starting from backbone to the right selection of distillation points and corresponding loss functions. This results in the student networks that are not only leaner and faster but give excellent performance. For instance, our student network while performing better than the performance oriented methods like PSMNet [1], CFNet [2], and LEAStereo [3]) on benchmark SceneFlow dataset, is 8 x, 5 x, and 8 x faster respectively. Furthermore, compared to speed oriented methods having inference time less than 100ms, our student networks perform better than all the tested methods. In addition, our student network also shows better generalization capabilities when tested on unseen datasets like ETH3D and Middlebury11Code: https://github.com/cogsys-tuebingen/Distilling-Stereo-Networks.

As robots become increasingly prevalent amidst diverse environments, their ability to adapt to novel scenarios and objects is essential. Advances in modern object detection have also paved the way for robots to identify interaction entities within their immediate vicinity. One drawback is that the robot's operational domain must be known at the time of training, which hinders the robot's ability to adapt to unexpected environments outside the preselected classes. However, when encountering such challenges a human can provide support to a robot by teaching it about the new, yet unknown objects on an ad hoc basis. In this work, we merge augmented reality and human gaze in the context of multimodal human-robot interaction to compose saliency-aware gaze heatmaps leveraged by a robot to learn emerging objects of interest. Our results show that our proposed method exceeds the capabilities of the current state of the art and outperforms it in terms of commonly used object detection metrics.

Spin plays a considerable role in table tennis, making a shot's trajectory harder to read and predict. However, the spin is challenging to measure because of the ball's high velocity and the magnitude of the spin values. Existing methods either require extremely high framerate cameras or are unreliable because they use the ball's logo, which may not always be visible. Because of this, many table tennis-playing robots ignore the spin, which severely limits their capabilities. This paper proposes an easily implementable and reliable spin estimation method. We developed a dotted-ball orientation estimation (DOE) method, that can then be used to estimate the spin. The dots are first localized on the image using a CNN and then identified using geometric hashing. The spin is finally regressed from the estimated orientations. Using our algorithm, the ball's orientation can be estimated with a mean error of 2.4° and the spin estimation has an relative error lower than 1%. Spins up to 175 rps are measurable with a camera of 350 fps in real time. Using our method, we generated a dataset of table tennis ball trajectories with position and spin, available on our project page. Project page: https://cogsys-tuebingen.github.io/spindoe/.

This paper introduces a novel approach to video object detection detection and tracking on Unmanned Aerial Vehicles (UAVs). By incorporating metadata, the proposed approach creates a memory map of object locations in actual world coordinates, providing a more robust and interpretable representation of object locations in both, image space and the real world. We use this representation to boost confidences, resulting in improved performance for several temporal computer vision tasks, such as video object detection, short and long-term single and multi-object tracking, and video anomaly detection. These findings confirm the benefits of metadata in enhancing the capabilities of UAVs in the field of temporal computer vision and pave the way for further advancements in this area.

Pose estimation of objects in images is an essential problem in virtual and augmented reality and robotics. Traditional solutions use depth cameras, which can be expensive, and working solutions require long processing times. This work focuses on the more difficult task when only RGB information is available. To this end, we predict not only the pose of an object but the complete probability density function (pdf) on the rotation manifold. This is the most general way to approach the pose estimation problem and is particularly useful in analysing object symmetries. In this work, we leverage implicit neural representations for the task of pose estimation and show that hypernetworks can be used to predict the rotational pdf. Furthermore, we analyse the Fourier embedding on SO(3) and evaluate the effectiveness of an initial Fourier embedding that proved successful. Our HyperPosePDF outperforms the current SOTA approaches on the SYMSOL dataset.

Deep Learning could drastically boost the classification accuracy for Hyperspectral Imaging (HSI). Still, the training on the mostly small hyperspectral data sets is not trivial. Two key challenges are the large channel dimension of the recordings and the incompatibility between cameras of different manufacturers.By introducing a suitable model bias and continuously defining the channel dimension, we propose a 2D convolution optimized for these challenges of Hyperspectral Imaging. We evaluate the method based on two different hyperspectral applications (inline inspection and remote sensing). Besides the shown superiority of the model, the modification adds additional explanatory power.In addition, the model learns the necessary camera filters in a data-driven manner. Based on these camera filters, an optimal camera can be designed.

For sensible human-robot interaction, it is crucial for the robot to have an awareness of its physical surroundings. In practical applications, however, the environment is manifold and possible objects for interaction are innumerable. Due to this fact, the use of robots in variable situations surrounded by unknown interaction entities is challenging and the inclusion of pre-trained object-detection neural networks not always feasible. In this work, we propose deploying augmented reality and eye tracking to flexibilize robots in non-predefined scenarios. To this end, we present and evaluate a method for extrinsic calibration of robot sensors, specifically a camera in our case, that is both fast and user-friendly, achieving competitive accuracy compared to classical approaches. By incorporating human gaze into the robot's segmentation process, we enable the 3D detection and localization of unknown objects without any training. Such an approach can facilitate interaction with objects for which training data is not available. At the same time, a visualization of the resulting 3D bounding boxes in the human's augmented reality leads to exceedingly direct feedback, providing insight into the robot's state of knowledge. Our approach thus opens the door to additional interaction possibilities, such as the subsequent initialization of actions like grasping.

Implicit neural representations parameterized by multilayer perceptrons have been shown to be a powerful paradigm that offers many potential advantages over traditional representations. The representation of fine details of signals is approached using a Fourier mapping to the input. However, naively applying a Fourier embedding is a double-edged sword, since gradient signals from the sampled points are incoherent (in terms of both direction and magnitude) and can easily cancel each other out. Hence, we propose an iterative algorithm that is able to gradually adjust a poorly chosen Fourier embedding in a way that it reaches an optimal parametrization after a few iterations.

Acquiring data to train deep learning-based object detectors on Unmanned Aerial Vehicles (UAVs) is expensive, time-consuming and may even be prohibited by law in specific environments. On the other hand, synthetic data is fast and cheap to access. In this work, we explore the potential use of synthetic data in object detection from UAVs across various application environments. For that, we extend the open-source framework DeepGTAV to work for UAV scenarios. We capture various large-scale high-resolution synthetic data sets in several domains to demonstrate their use in real-world object detection from UAVs by analyzing multiple training strategies across several models. Furthermore, we analyze several different data generation and sampling parameters to provide actionable engineering advice for further scientific research. The DeepGTAV framework is available at https://git.io/Jyf5j.

This work introduces a new preprocessing step for object detection applicable to UAV bird’s eye view imagery, which we call Adaptive Resizing. By design, it helps alleviate the challenges coming with the vast variances in objects’ scales, naturally inherent to UAV data sets. Furthermore, it improves inference speed by two to three times on average. We test this extensively on UAVDT, VisDrone, and on a new data set we captured ourselves and achieve consistent improvements while being considerably faster. Moreover, we show how to apply this method to generic UAV object detection tasks. Additionally, we successfully test our approach on a height transfer task where we train on some interval of altitudes and test on a different one. Furthermore, we introduce a small, fast detector meant for deployment to an embedded GPU. Code is available at https://github.com/cogsys-tuebingen/adaptive_resizer.

We introduce a model-free approach to predict the future state of the ball and learn the appropriate stroke accordingly for robotic table tennis. Based on the gated recurrent unit (GRU) and the encoder-decoder (ED), a GRU-ED approach is developed for predicting the future state (position, velocity and acceleration) of the ball when observing a partial trajectory. By taking as input the predicted state at hitting time, we learn an appropriate stroke movement with a model-free reinforcement learning (RL) approach. The experimental results show that the proposed approach outperforms others in trajectory prediction. Acceleration and spin of the ball provide an equivalent effect in learning an accurate stroke motion. An additional experiment conducted with a real table tennis robot shows that the robot can accurately hit the ball and return the ball to the desired target with a pretrained RL model.

Implicit Neural Representations (INR) use multilayer perceptrons to represent high-frequency functions in low-dimensional problem domains. Recently these representations achieved state-of-the-art results on tasks related to complex 3D objects and scenes. A core problem is the representation of highly detailed signals, which is tackled using networks with periodic activation functions (SIRENs) or applying Fourier mappings to the input. This work analyzes the connection between the two methods and shows that a Fourier mapped perceptron is structurally like one hidden layer SIREN. Furthermore, we identify the relationship between the previously proposed Fourier mapping and the general d-dimensional Fourier series, leading to an integer lattice mapping. Moreover, we modify a progressive training strategy to work on arbitrary Fourier mappings and show that it improves the generalization of the interpolation task. Lastly, we compare the different mappings on the image regression and novel view synthesis tasks. We confirm the previous finding that the main contributor to the mapping performance is the size of the embedding and standard deviation of its elements.

Recent methods in stereo matching have continuously improved the accuracy using deep models. This gain, however, is attained with a high increase in computation cost, such that the network may not fit even on a moderate GPU. This issue raises problems when the model needs to be deployed on resource-limited devices. For this, we propose two light models for stereo vision with reduced complexity and without sacrificing accuracy. Depending on the dimension of cost volume, we design a 2D and a 3D model with encoder-decoders built from 2D and 3D convolutions, respectively. To this end, we leverage 2D MobileNet blocks and extend them to 3D for stereo vision application. Besides, a new cost volume is proposed to boost the accuracy of the 2D model, making it performing close to 3D networks. Experiments show that the proposed 2D/3D networks effectively reduce the computational expense (27%/95% and 72%/38% fewer parameters/operations in 2D and 3D models, respectively) while upholding the accuracy. Code: https://github.com/cogsys-tuebingen/mobilestereonet.

Unmanned Aerial Vehicles (UAVs) are of crucial importance in search and rescue missions in maritime environments due to their flexible and fast operation capabilities. Modern computer vision algorithms are of great interest in aiding such missions. However, they are dependent on large amounts of real-case training data from UAVs, which is only available for traffic scenarios on land. Moreover, current object detection and tracking data sets only provide limited environmental information or none at all, neglecting a valuable source of information. Therefore, this paper introduces a large-scaled visual object detection and tracking benchmark (SeaDronesSee) aiming to bridge the gap from land-based vision systems to sea-based ones. We collect and annotate over 54,000 frames with 400,000 instances captured from various altitudes and viewing angles ranging from 5 to 260 meters and 0 to 90° degrees while providing the respective meta information for altitude, viewing angle and other meta data. We evaluate multiple state-of-the-art computer vision algorithms on this newly established benchmark serving as baselines. We provide an evaluation server where researchers can upload their prediction and compare their results on a central leaderboard1.

Bin picking is a core problem in industrial environments and robotics, with its main module as 6D pose estimation. However, industrial depth sensors have a lack of accuracy when it comes to small objects. Therefore, we propose a framework for pose estimation in highly cluttered scenes with small objects, which mainly relies on RGB data and makes use of depth information only for pose refinement. In this work, we compare synthetic data generation approaches for object detection and pose estimation and introduce a pose filtering algorithm that determines the most accurate estimated poses. We will make our real dataset for object detection available with the paper.

Deep learning based 3D stereo networks give superior performance compared to 2D networks and conventional stereo methods. However, this improvement in the performance comes at the cost of increased computational complexity, thus making these networks non-practical for the real-world applications. Specifically, these networks use 3D convolutions as a major work horse to refine and regress disparities. In this work first, we show that these 3D convolutions in stereo networks consume up to 94% of overall network operations and act as a major bottleneck. Next, we propose a set of “plug-&-run” separable convolutions to reduce the number of parameters and operations. When integrated with the existing state of the art stereo networks, these convolutions lead up to $7\times$ reduction in number of operations and up to $3.5\times$ reduction in parameters without compromising their performance. In fact these convolutions lead to improvement in their performance in the majority of cases11This work is part of the project DeepStereoVision (FRE: 01IS18024B) sponsored by the German Ministry of Education & Research (BMBF).

Naturally, finding joint trajectories for robotic manipulators involves competing optimization goals. On the one hand, the end-effector should move along a predictable and short path while on the other hand joint movement and acceleration should be kept to a minimum. Obstacles in the workspace or joint limits complicate the situation even further. Constructing a metric that makes undesired configurations more expensive to travel through, we equip the joint space with a notion of cost. The motion planning task then reduces to the problem of finding cheapest paths connecting two configurations – so-called geodesics. We show how to construct suitable metrics for a variety of typical scenarios and present an efficient algorithm for the computation of the corresponding geodesics. Our approach makes it very easy to balance different optimization goals and produces natural and smooth manipulator movement.

Reinforcement learning (RL) has achieved some impressive recent successes in various computer games and simulations. Most of these successes are based on having large numbers of episodes from which the agent can learn. In typical robotic applications, however, the number of feasible attempts is very limited. In this paper we present a sample-efficient RL algorithm applied to the example of a table tennis robot. In table tennis every stroke is different, with varying placement, speed and spin. An accurate return therefore has to be found depending on a high-dimensional continuous state space. To make learning in few trials possible the method is embedded into our robot system. In this way we can use a one-step environment. The state space depends on the ball at hitting time (position, velocity, spin) and the action is the racket state (orientation, velocity) at hitting. An actor-critic based deterministic policy gradient algorithm was developed for accelerated learning. Our approach performs competitively both in a simulation and on the real robot in a number of challenging scenarios. Accurate results are obtained without pre-training in under 200 episodes of training. The video presenting our experiments is available at https://youtu.be/uRAtdoL6Wpw.

We present a system to measure the ripeness of fruit with a hyperspectral camera and a suitable deep neural network architecture. This architecture did outperform competitive baseline models on the prediction of the ripeness state of fruit. For this, we recorded a data set of ripening avocados and kiwis, which we make public. We also describe the process of data collection in a manner that the adaption for other fruit is easy. The trained network is validated empirically, and we investigate the trained features. Furthermore, a technique is introduced to visualize the ripening process.

Multi-object tracking is a critical component in autonomous navigation, as it provides valuable information for decision-making. Many researchers tackled the 3D multi-object tracking task by filtering out the frame-by-frame 3D detections; however, their focus was mainly on finding useful features or proper matching metrics. Our work focuses on a neglected part of the tracking system: score refinement and tracklet termination. We show that manipulating the scores depending on time consistency while terminating the tracklets depending on the tracklet score improves tracking results. We do this by increasing the matched tracklets’ score with score update functions and decreasing the unmatched tracklets’ score. Compared to count-based methods, our method consistently produces better AMOTA and MOTA scores when utilizing various detectors and filtering algorithms on different datasets. The improvements in AMOTA score went up to 1.83 and 2.96 in MOTA. We also used our method as a late-fusion ensembling method, and it performed better than voting-based ensemble methods by a solid margin. It achieved an AMOTA score of 67.6 on nuScenes test evaluation, which is comparable to other state-of-the-art trackers. Code is publicly available at: https://github.com/cogsys-tuebingen/CBMOT.

We present FourierNet, a single shot, anchor-free, fully convolutional instance segmentation method that predicts a shape vector. Consequently, this shape vector is converted into the masks' contour points using a fast numerical transform. Compared to previous methods, we introduce a new training technique, where we utilize a differentiable shape decoder, which manages the automatic weight balancing of the shape vector's coefficients. We used the Fourier series as a shape encoder because of its coefficient interpretability and fast implementation. FourierNet shows promising results compared to polygon representation methods, achieving 30.6 mAP on the MS COCO 2017 benchmark. At lower image resolutions, it runs at 26.6 FPS with 24.3 mAP. It reaches 23.3 mAP using just eight parameters to represent the mask (note that at least four parameters are needed for bounding box prediction only). Qualitative analysis shows that suppressing a reasonable proportion of higher frequencies of Fourier series, still generates meaningful masks. These results validate our understanding that lower frequency components hold higher information for the segmentation task, and therefore, we can achieve a compressed representation. Code is available at: github.com/cogsys-tuebingen/FourierNet.

In this paper we propose a new deep neural network system, called Yolo+FPN, which fuses both 2D and 3D object detection algorithms to achieve better real-time object detection results and faster inference speed, to be used on real robots. Finding an optimized fusion strategy to efficiently combine 3D object detection with 2D detection information is useful and challenging for both indoor and outdoor robots. In order to satisfy real-time requirements, a trade-off between accuracy and efficiency is needed. We not only have improved training and test accuracies and lower mean losses on the KITTI object detection benchmark comparing with our baseline method, but also achieve competitive average precision on 3D detection of all classes in three levels of difficulty comparing with other state-of-the-art methods. Also, we implemented Yolo+FPN system using an RGB-D camera, and compared the speed of object detection using different GPUs. For the real implementation of both indoor and outdoor scenes, we focus on person detection, which is the most challenging and important among the three classes.

Simultaneous Localization and Mapping (SLAM) is one of the basic problems in mobile robotics. While most approaches are based on occupancy grid maps, Normal Distributions Transforms (NDT) and mixtures like Occupancy Normal Distribution Transforms (ONDT) have been shown to represent sensor measurements more accurately. In this work, we slightly re-formulate the (O)NDT matching function such that it becomes a least squares problem that can be solved with various robust numerical and analytical non-linear optimizers. Further, we propose a novel global (O)NDT scan matcher for loop closure. In our evaluation, our NDT and ONDT methods are able to outperform the occupancy grid map based ones we adopted from Google's Cartographer implementation.

In table tennis, the rotation (spin) of the ball plays a crucial role. A table tennis match will feature a variety of strokes. Each generates different amounts and types of spin. To develop a robot that can compete with a human player, the robot needs to detect spin, so it can plan an appropriate return stroke. In this paper we compare three methods to estimate spin. The first two approaches use a high-speed camera that captures the ball in flight at a frame rate of 380 Hz. This camera allows the movement of the circular brand logo printed on the ball to be seen. The first approach uses background difference to determine the position of the logo. In a second alternative, we train a CNN to predict the orientation of the logo. The third method evaluates the trajectory of the ball and derives the rotation from the effect of the Magnus force. This method gives the highest accuracy and is used for a demonstration. Our robot successfully copes with different spin types in a real table tennis rally against a human opponent.

Successful and meaningful human-robot interaction requires robots to have knowledge about the interaction context - e.g., which objects should be interacted with. Unfortunately, the corpora of interactive objects is - for all practical purposes - infinite. This fact hinders the deployment of robots with pre-trained object-detection neural networks other than in pre-defined scenarios. A more flexible alternative to pre-training is to let a human teach the robot about new objects after deployment. However, doing so manually presents significant usability issues as the user must manipulate the object and communicate the object's boundaries to the robot. In this work, we propose streamlining this process by using automatic object location proposal methods in combination with human gaze to distill pertinent object location proposals. Experiments show that the proposed method 1) increased the precision by a factor of approximately 21 compared to location proposal alone, 2) is able to locate objects sufficiently similar to a state-of-the-art pre-trained deep-learning method (FCOS) without any training, and 3) detected objects that were completely missed by FCOS. Furthermore, the method is able to locate objects for which FCOS was not trained on, which are undetectable for FCOS by definition.

A major challenge in current optimization research for deep learning is to automatically find optimal step sizes for each update step. The optimal step size is closely related to the shape of the loss in the update step direction. However, this shape has not yet been examined in detail. This work shows empirically that the sample loss over lines in negative gradient direction is mostly convex and well suited for one-dimensional parabolic approximations. Exploiting this parabolic property we introduce a simple and robust line search approach, which performs loss-shape dependent update steps. Our approach combines well-known methods such as parabolic approximation, line search and conjugate gradient, to perform efficiently. It successfully competes with common and state-of-the-art optimization methods on a large variety of experiments without the need of hand-designed step size schedules. Thus, it is of interest for objectives where step-size schedules are unknown or do not perform well. Our excessive evaluation includes multiple comprehensive hyperparameter grid searches on several datasets and architectures. We provide proof of convergence for an adapted scenario. Finally, we give a general investigation of exact line searches in the context of sample losses and exact losses, including their relation to our line search approach.

In this article we develop a graph convolutional neural network (GCN) for the approximation of the AC power flow in electrical power grids. The proposed architecture is fully generic and purely data driven, such that no information about the actual underlying physical topology of the power grid is required. This gives the opportunity to apply this approach to a wide range of multivariate regression problems. We test our architecture on 3 datasets of different sizes, two of which are real world type power grids containing up to 5488 nodes. We show that the proposed method allows an accurate approximation of the power flow specifically in the case of large power grids. The GCN architecture implies intrinsic extrapolation, allowing a reasonable reduction of the number of trainable parameters as well as training samples. In contrast, classical approaches based on fully connected networks are shown to face difficulties when fitting such high dimensional functions.

While recent Neural Architecture Search (NAS) algorithms are thousands of times faster than the pioneering works, it is often overlooked that they use fewer candidate operations, resulting in a significantly smaller search space. We present PR-DARTS, a NAS algorithm that discovers strong network configurations in a much larger search space and a single day. A small candidate operation pool is used, from which candidates are progressively pruned and replaced with better performing ones. Experiments on CIFAR-10 and CIFAR-100 achieve 2.51% and 15.53% test error, respectively, despite searching in a space where each cell has 150 times as many possible configurations than in the DARTS baseline. All of our code is available at https://github.com/cogsys-tuebingen/prdarts.

This paper focuses on the semantic segmentation networks of 3D point clouds for indoor scenes. We first reduce the PointNet structure to get a reduced point network (RPN) that achieves the same performance but has less training and evaluation time comparing with PointNet. Secondly, we propose two solutions to get scale invariance and robust test performance: one is modifying RPN to get the robust performance and adding stable multi-scaling layers (MPN); another is introducing a novel point-based network based on Angular coordinates instead of Euclidean coordinates for point representation (APN). The ablation study of our networks (RPN, MPN, APN) is done. Compared to state-of-the-art semantic segmentation networks based on 3D point clouds, the experimental results show that our MPN and APN networks both achieve higher training and evaluation accuracy, as well as mean intersection over union (IoU) and overall accuracy on two benchmarks. We also have better qualitative segmentation results when directly test on another benchmark indoor scenes as well as real corridor scenes from our robots RGB-D mapping.

This work presents a model based traversability analysis method which employs a detailed vehicle model to perform real-time path planning in complex environments. The vehicle model represents the vehicle's wheels and chassis, allowing it to accurately predict the vehicles 3D pose, detailed contact information for each wheel and the occurrence of a chassis collision given a 2D pose on an elevation map. These predictions are weighted, depending on the safety requirements of the vehicle, to provide a scoring function for an A*-like search strategy. The proposed method is designed to run at frame rates of 30Hz on data from a RGB-D sensor to provide reactive planning of safe paths. For evaluation, two wheeled mobile robots in different simulated and real world environment setups were tested to show the reliability and performance of the proposed method.

Neural network architectures found by sophistic search algorithms achieve strikingly good test performance, surpassing most human-crafted network models by significant margins. Although computationally efficient, their design is often very complex, impairing execution speed. Additionally, finding models outside of the search space is not possible by design. While our space is still limited, we implement undiscoverable expert knowledge into the economic search algorithm Efficient Neural Architecture Search (ENAS), guided by the design principles and architecture of ShuffleNet V2. While maintaining baseline-like 2.85% test error on CIFAR-10, our ShuffleNASNets are significantly less complex, require fewer parameters, and are two times faster than the ENAS baseline in a classification task. These models also scale well to a low parameter space, achieving less than 5% test error with little regularization and only 236K parameters.

Recursive state estimation is often used to estimate a probability density function of a specific state, e.g. a robot's pose, over time. Compared to Kalman filters, Sequential Monte Carlo (SMC) methods are less constrained in regard to state propagation and update model definition, which makes it easier to implement any suitable problem. In this work, we present a generic Sequential Monte Carlo framework, which uses abstract formulations for importance weighting, propagation and resampling and provides an independent core algorithm that is usable for any problem instantiation, such that diverse SMC problems can be implemented easily and quickly, since the basic algorithms are already provided. Current applications include 2D localization, 2D tracking in a SLAM system and contact point localization on a manipulator surface. Further, we introduce concepts to deal with data input synchronization and fair execution of different weighting models, which makes it possible to incorporate data from as many update sources, e.g. sensors, as desired. As a typical application scenario, we provide a plugin-based and hence easily extensible instantiation for 2D localization and demonstrate the capabilities of our framework and methods based on a well-known dataset.

Many robotic applications, especially exploration scenarios, benefit from deploying multiple collaborating robots with the aim of parallelizing and therefore accelerating the involved task. One critical part of a multi-robot system is its communication system. Depending on the application scenario, high-bandwidth wireless connections, such as WiFi, may not always be available and suffer from a limited communication range. On the other hand, low-bandwith systems require the application itself to deal with limited information exchange. In this work, we present a novel approach for collaborative mapping using mixtures of occupancy and NDT maps (called ONDT), which provide detailed information at low resolutions and in which the pose uncertainty can be encoded efficiently. Further, we compare the applicability of three different radio frequency modules operating at different frequencies in real world experiments with five robots at large distances.

Detecting and localizing contacts acting on a manipulator is a relevant problem for manipulation tasks like grasping, since contact information can be helpful for recovering from collisions or for improving the grasping performance itself. In this work, we present a solution for contact point localization, which is based on Monte Carlo Localization. Usually, an Articulated Robotic Manipulator (ARM) is not equipped with tactile skin, but with proprioceptive sensors, which we assume as an input for our method. In our experiments, we compare our method with a direct optimization method, machine learning approaches and another particle filter method, both on simulated and real world data from a Kinova Jaco2. While our proposed method clearly outperforms the other optimization approaches, it performs about equally well as Random Forest (RF) classifiers, although both methods have their strengths on different parts of the manipulator, and even achieves better results than multi-layer perceptions (MLPs) on the links farthest from the manipulator base.

Fast and robust person detection is one of the most important tasks for robotic applications involving human interaction. Particularly in mobile robotics this task is still challenging. Though there are already reliable and real-time capable approaches, they are usually computationally expensive. They either require GPUs or multiple CPU cores in order to work properly. Furthermore, some of the approaches are designed for special environments and sensor types, which reduces general applicability. In this work, we present a robust, generic and lightweight solution for real-time 3D person detection. Since our approach requires only a single CPU thread, it can be run as a background process and is suitable for smaller robotic systems. We demonstrate applicability to indoor and outdoor scenarios using different 3D sensor types separately. Moreover, we are able to show that the proposed method outperforms other state-of-the-art approaches, including a DCNN.

A model-based Machine Learning (ML) approach is presented to detect and localize external contacts on a 6 degree of freedom (DoF) serial manipulator. This approach only requires the use of proprioceptive sensors (joint positions, velocities and one-dimensional (ID) joint torques already available in the robot arm). Good results are obtained with Random Forests (RFs) and Multi-Layer-Perceptrons (MLPs) leading to a precise localization of the contact link and its orientation. Apart from the link in contact and the orientation of the force, RFs and MLPs are also able to differentiate between contact points on the same link and orientation but with different distances to the joint axis. We experimentally verify this approach on simulated and real data obtained from the Kinova Jaco 2 manipulator and compare it to an optimization based approach.

RGBD sensors are widely used in robotics and computer vision. Since the Microsoft Kinect was launched, various other sensors, such as the Asus Xtion PRO LIVE and the Intel R200, have been developed. Among other applications, these sensors are useful for 3D reconstruction, which can be achieved with KinFu (an open-source implementation of KinectFusion). Reconstructions of objects and people are useful in multiple fields such as computer graphics and ergonomics. In this paper, we analyze if the 3D reconstructions generated by KinFu depend on the sensor used to record its input. The analysis comprehends qualitative and quantitative evaluations of reconstructions created from point clouds recorded with six well-known sensors. The results suggest that the reconstruction is sensor dependent and the selection of a determined sensor is conditioned to the requirements of each application.

Feature-based simultaneous localization and mapping (SLAM) algorithms with additional semantics can have better feature matching and tracking accuracies than the original SLAM algorithms. Therefore, this paper shows how to improve feature-based SLAM by only matching features from objects of the same semantic class. The basic idea is to use a deep neural network, YOLO (you only look once [1]), to classify objects and to associate features with the objects in whose bounding box they appear, thus giving features the semantic label of these objects. During feature matching of the SLAM algorithms, only features with the same semantic label are matched (e.g. books with books, bottles with bottles etc.), eliminating matches of similar features on different classes of objects. Experiments of classical ORB-SLAM2 with YOLO have been performed on an embedded PC. Additionally, ORB-SLAM2 with different versions of YOLO has also been tested on a powerful desktop GPU as well as on an Nvidia Jetson TX2 board. The experimental results show that using the semantic information given by object recognition methods reduces wrong feature matches in tracking and decreases the tracking lost cases.

Efficient 2D and 3D map representations of both static and dynamic, indoor and outdoor environments are crucial for navigation of driving and flying robots. In this paper, we propose a fast and accurate approach for 2D and 3D Normal Distributions Transform (NDT) mapping based on indexed kd-trees. Similar to other approaches, we also model free space, which allows us to obtain occupancy probabilities. Additionally, we provide optional visibility based updates to enhance map consistency in case of noisy data, e.g. from stereo cameras. Unlike other available implementations, our approach is natively applicable to large-scale environments and in real-time, because our maps are able to grow dynamically. This also offers applicability to exploration tasks. To evaluate our approach, we present experimental results on publicly available datasets and discuss the mapping efficiency in terms of accuracy, runtime and memory management. As an exemplary use case, we apply our maps to Monte Carlo Localization on a well-known large-scale dataset.

In this paper, we propose and experimentally verify a distributed formation control algorithm for a group of multirotor unmanned aerial vehicles (UAVs). The algorithm brings the whole group of UAVs simultaneously to a prescribed submanifold that determines the formation shape in an asymptotically stable fashion in two- and three-dimensional environments. The complete distributed control framework is implemented with the combination of a fast model predictive control method executed at 50 Hz on low-power computers onboard multirotor UAVs and validated via a series of hardware-in-the-loop simulations and real-robot experiments. The experiments are configured to study the control performance in various formation cases of arbitrary time-varying (e.g., expanding, shrinking, or moving) shapes. In the actual experiments, up to four multirotors have been implemented to form arbitrary triangular, rectangular, and circular shapes drawn by the operator via a human-robot interaction device. We also carry out hardware-in-the-loop simulations using up to six onboard computers to achieve spherical formations and a formation moving through obstacles.

A fundamental task of automated guided vehicles is transporting heavy payloads. These payloads are often given in the form of large containers that are mounted onto a cart with four caster wheels. In this paper we investigate a combined detection and control architecture that allows an AGV to reliably detect and acquire such containers at any location, independently of a global localization system. We propose a solution using a horizontally mounted 2D laser scanner and a forward facing 3D time-of-flight camera to detect the payload. Furthermore we demonstrate the algorithm on an asymmetric AGV that consists of a differentially driven base and a trailer with a lifting unit.

A new nonlinear control law for path following with skid-steered mobile robots is proposed. A terrain dependent kinematic model is utilized in path coordinates, and the kinematic parameters are experimentally evaluated. A kinematic path following control is developed using the Lyapunov approach. A separate linear velocity control is then proposed, taking reachable curvatures and actuator saturation into account. The proposed approach is experimentally evaluated in different terrain scenarios, and compared with two other state-of-the-art algorithms. The skid-steered vehicle used for the experiments is the Robotnik Summit XL, a well known commercial mobile robot.

We present LS-ELAS, a line segment extension to the ELAS algorithm, which increases the performance and robustness. LS-ELAS is a binocular dense stereo matching algorithm, which computes the disparities in constant time for most of the pixels in the image and in linear time for a small subset of the pixels (support points). Our approach is based on line segments to determine the support points instead of uniformly selecting them over the image range. This way we find very informative support points which preserve the depth discontinuity. The prior of our Bayesian stereo matching method is based on a set of line segments and a set of support points. Both sets are given to a constrained Delaunay triangulation to generate a triangulation mesh which is aware of possible depth discontinuities. We further increased the accuracy by using an adaptive method to sample candidate points along edge segments. We evaluated our algorithm on the Middlebury benchmark.

A new navigation system for outdoor person following at higher speeds (maximum speed ≈2.5 m/s) is proposed. A combination of global and local path planning, and path following control, allows a robot to follow a jogger in various outdoor scenarios, including highly dynamical environments with pedestrians. Our fast path planning provides smooth and obstacle-free paths, even at higher speeds, in the presence of dynamic obstacles and in narrow passages. Our control algorithm allows precise following of the planned paths on different terrain types, including snow and ice. The mobile robot used in the experiments is the skid-steered robot Robotnik Summit XL.

Fast and accurate obstacle detection is a crucial component for autonomous robot navigation. It becomes even more important for a shared control vehicle like an electric wheeled walker, since the safety of the vehicle and the user depend on the correct classification of obstacles. This work describes a method for pose estimation of four-wheeled vehicles, which utilizes the fixed resolution of digital elevation maps to generate a detailed vehicle model. The vehicle's wheels are also approximated using digital elevation maps, allowing efficient calculation of wheel to ground contact points and therefore fast and accurate estimation of valid vehicle poses. To evaluate the proposed method, pose estimates are compared to three datasets including ground truth poses: one created using an external tracking system and two created by simulations of wheeled robots. It is also shown that the method is fast enough for real time operation.

In this work we present a novel fingerprint similarity sensor model for the purpose of localizing a mobile robot with passive ultra-high frequency (UHF) radio-frequency identification (RFID) through location fingerprinting. We firstly evaluate the performance of probabilistic similarity measures applied to received signal strength (RSS) and compare them with previous results obtained with well known vector similarity measures. We furthermore extend the observation model used in a particle filter to dynamically adapt to the uncertainty of single candidate fingerprints by using their similarity to the current observation. For this purpose, we derive a new likelihood function and introduce an alternative way of selecting candidate fingerprints using a combination of their signal space similarity as well as the distance between the currently estimated pose and reference fingerprints. Results obtained from experiments in two different environments highlight the improved accuracy as well as robustness of the proposed methods.

This paper investigates Recurrent Neural Networks (RNNs), particularly Dynamic Cortex Memories (DCMs), an extension of Long Short Term Memories (LSTMs) for classification of 14 different ground types based on vibration data. Also a simple regularization technique called Sequence Boundary Dropout (SBD) is introduced, which effectively enlarges the training set and improves generalization. The neural networks perform in the time domain without any explicit feature computation, while previous state-of-the-art methods extract features mainly in the frequency domain. The presented approach does not require a time window, is causal, and works just-in-time, such that a classification can be done online at each new time step. Furthermore, we show that the neural networks outperform previous methods significantly in terms of classification accuracy. Finally, we demonstrate that the networks retrained with Random Activation Preservation (RAP) can classify very early - within a fraction of a second - but robustly at the same time in a continuous recognition scenario with varying classes.

Automated guided vehicles (AGVs) are useful for a variety of transportation tasks. They usually detect obstacles on the path they are following using 2D laser scanners. If an AGV should be deployed in a shared space with people, 3D information has to be considered as well to detect unforeseen obstacles. These can be small objects on the floor or overhanging parts of larger objects, which cannot be seen by the standard 2D safety scanners. We propose a generic object detection pipeline using 3D time-of-flight cameras, that can be used in real-time on AGVs of low height and demonstrate its robustness to different measurement artefacts.

During various inflammatory processes circulating cytokines including IL-6, IL-1β, and TNFα elicit a broad and clinically relevant impairment of hepatic detoxification that is based on the simultaneous downregulation of many drug metabolizing enzymes and transporter genes. To address the question whether a common mechanism is involved we treated human primary hepatocytes with IL-6, the major mediator of the acute phase response in liver, and characterized acute phase and detoxification responses in quantitative gene expression and (phospho-)proteomics data sets. Selective inhibitors were used to disentangle the roles of JAK/STAT, MAPK, and PI3K signaling pathways. A prior knowledge-based fuzzy logic model comprising signal transduction and gene regulation was established and trained with perturbation-derived gene expression data from five hepatocyte donors. Our model suggests a greater role of MAPK/PI3K compared to JAK/STAT with the orphan nuclear receptor RXRα playing a central role in mediating transcriptional downregulation. Validation experiments revealed a striking similarity of RXRα gene silencing versus IL-6 induced negative gene regulation (rs = 0.79; P<0.0001). These results concur with RXRα functioning as obligatory heterodimerization partner for several nuclear receptors that regulate drug and lipid metabolism.

The goal of saliency detection is to highlight objects in image data that stand out relative to their surrounding. Therefore, saliency detection aims to capture regions that are perceived as important. The most recent bottom-up approaches for saliency detection measure contrast based on visual features in 2D scenes, ignoring depth value. This work presents an effective method to measure saliency by mapping pixels into foreground and background regions in RGB-D images. Namely, we first segment an image into regions to evaluate the object uniqueness and consistency using graph-based segmentation. Then, we utilize the region color, depth, layout and boundary information to produce robust foreground and background saliency measures. Finally, we combine the two saliency maps based on Gaussian weights. As a result, our approach produces high-quality saliency maps, which may be used for further processing like object detection or recognition. Experimental results on two datasets compare our method with the state of the art and highlight its effectiveness.

The recently introduced Dynamic Cortex Memory (DCM) is an extension of the Long Short Term Memory (LSTM) providing a systematic inter-gate connection infrastructure. In this paper the behavior of DCM networks is studied in more detail and their potential in the field of gradient-based sequence learning is investigated. Hereby, DCM networks are analyzed regarding particular key features of neural signal processing systems, namely, their robustness to noise and their ability of time warping. Throughout all experiments we show that DCMs converge faster and yield better results than LSTMs. Hereby, DCM networks require overall less weights than pure LSTM networks to achieve the same or even better results. Besides, a promising neurally implemented just-in-time online signal filter approach is presented, which is latency-free and still provides an accurate filtering performance much better than conventional low-pass filters. We also show that the neural networks can do explicit time warping even better than the Dynamic Time Warping (DTW) algorithm, which is a specialized method developed for this task.

This paper presents a nonlinear control approach for quadrotor Micro Aerial Vehicles (MAVs), which combines a backstepping-like regulator based on the solution of a certain class of global output regulation problems for the rigid body equations on SO(3), a robust controller for the system with bounded disturbances, as well as a trajectory generator using a model predictive control method. The proposed algorithm is endowed with strong convergence properties so that it allows the quadrotor MAVs to reach almost all the desired attitudes. The control approach is implemented on a high-payload-capable quadcopter with unstructured dynamics and unknown disturbances. The performance of our algorithm is demonstrated through a series of experimental evaluations and comparisons with another control method on normal and aggressive trajectory tracking tasks.

Traversable region detection is important for autonomous visual navigation of mobile robots. Only short range traversable regions can be detected using traditional methods based on stereo vision because of the limited image resolution and baseline of stereo vision. In this paper, we propose a novel method to detect long range traversable regions without using any supervised or self-supervised learning process. Superpixels are clustered using an improved spectral clustering algorithm to segment the image effectively, and after integrating short range traversable region detection based on u-v-disparity, the traversable region can be extended to long range naturally. The experimental results show that the proposed method works well in different outdoor/field environments, and the detecting range can be improved greatly in comparison with traditional methods. Furthermore, the proposed superpixels clustering algorithm can also be applied in other robot vision tasks like road detection and object recognition.

We present a fast and simple method for solving the perspective-three-point (P3P) problem. The problem is specified as finding the position and orientation of a camera from a single image, showing three points, while the position of the corresponding points in the world is known. Our approach is based on a new geometric parametrization, yielding simpler algebraic terms to compute, compared to all state-of-the-art methods. It is therefore faster and more accurate, since calculations are saved and rounding errors appear less. The method also showed to be robust against noise. A special case of this method has been successfully applied to a robotics scenario. However, the method is feasible for all applications that utilize P3P problem solvers.

In this paper, we present a new variant of the well-known Random Sample Consensus (RANSAC) algorithm for robust estimation of model parameters. The idea of our method is based on a kind of volatile memory which is similar to the pheromone evaporation in the ant colony optimization algorithm. Therefore, we call our improved RANSAC like algorithm ANTSAC. We describe our new approach and the influence of its relevant parameters to the achieved performance in detail. ANTSAC is computationally efficient and convincingly easy to implement. It turns out that ANTSAC significantly outperforms RANSAC regarding the number of inliers after a given number of iterations. Further, we show that the advantage of ANTSAC increases with the complexity of the problem, i.e., with the number of model parameters, as well as with the relative number of outliers. ANTSAC is entirely generic, such that no further domain knowledge is required, as it is for many other RANSAC extensions. Nevertheless, we show that it is competitive to state-of-the-art methods even in domain specific scenarios.

In this paper we propose a novel approach to classify detection events from a stream of radio-frequency identification (RFID) measurements for the purpose of mapping RFID transponders. Since raw readings from RFID readers only provide information on positive read attempts, i.e. the detections of a tag, we propose an outlier filter method solely based on the spatial extent of the sensor model that is used for the mapping process. Furthermore, we use this filter to actually classify detections as well as non-detections of tags into valid and invalid positive as well as negative detection events. We incorporate the different classes into our mapping pipeline and introduce several extensions to improve the mapping accuracy. Experimental results including the classification and mapping accuracy are presented to prove the effectiveness of our approach.

A map of occupied and free space in a robot's environment is a common prerequisite for navigational tasks. Although the first methods for occupancy mapping relied on a 2D grid representation, 3D volumetric approaches are becoming increasingly popular. In this paper we present a new volumetric mapping approach that is based on the OctoMap method. We designed this method to be more robust against measurement errors, in particular against high temporally or spatially correlated errors usually received from a stereo vision system. For this purpose, we define a probability measure that a voxel is currently visible. An update of a voxel's occupancy probability then happens with respect to this visibility probability, allowing us to neglect measurements for voxels that are actually unobservable. Finally, we model the depth error of a stereo vision sensor, and take care of this error when performing a map update. By evaluation we show that our method produces maps with far less erroneous artifacts compared to OctoMap. Our maps also require less memory, and due to an optimized update reduction, our method is also faster than OctoMap when processing dense range measurement data.

This paper extends a monocular visual simultaneous localization and mapping (SLAM) system to utilize two cameras with non-overlap in their respective field of views (FOVs). We achieve using it to enable autonomous navigation of a micro aerial vehicle (MAV) in unknown environments. The methodology behind this system can easily be extended to multi-camera rigs, if the onboard computation capability allows this. We analyze the iterative optimizations for pose tracking and map refinement of the SLAM system in multicamera cases. This ensures the soundness and accuracy of each optimization update. Our method is more resistant to tracking failure than conventional monocular visual SLAM systems, especially when MAVs fly in complex environments. It also brings more flexibility to configurations of multiple cameras used onboard of MAVs. We demonstrate its efficiency with both autonomous flight and manual flight of a MAV. The results are evaluated by comparisons with ground truth data provided by an external tracking system.

The correct classification of the surrounding terrain is an important ability of a mobile robot that drives in outdoor environments. Our robot uses a 3D LIDAR and a camera to classify terrain as either asphalt, cobblestones, grass, or gravel. We build on previous work where we modeled the terrain as a Conditional random field to account for spatial dependencies, which improved results substantially. We now show how to speed up the spatial classification by defining a new energy term for neighborhood relations. Moreover, we now also consider temporal dependencies as the robot moves. This not only further improves the results, but makes it possible to build local terrain maps of the environment. We describe how to efficiently integrate the classification results of each time step into the map in a probabilistic manner. By also detecting obstacles with the LIDAR, the robot can build combined terrain and elevation maps. We show that these maps can be used for semantic robot localization.

Recent research deals more and more with the application of ultra high frequency (UHF) radio-frequency identification (RFID) on mobile robots. However, the sensing characteristics between the reader and the tag (i.e. detections and signal strength) are challenging to model due to the influence of environmental effects (e.g. tag density, reflection, diffraction, or absorption). In this paper, we address the problem of dynamic objects tracking with a mobile agent using the signal strength from UHF RFID tags attached to objects. Our solution estimates the positions of RFID tags under a Bayesian framework. More precisely, we combine a two stage dynamic motion model with the dual particle filter, to capture the dynamic motion of the object and to quickly recover from failures in tracking. This approach is then tested on a Scitos G5 mobile robot through various experiments.

Abstract Motivation: Mass spectrometry-based protein profiling has become a key technology in biomedical research and biomarker discovery. Sample preparation strategies that reduce the complexity of tryptic digests by immunoaffinity substantially increase throughput and sensitivity in proteomic mass spectrometry. The scarce availability of peptide-specific capture antibodies limits these approaches. Recently antibodies directed against short terminal motifs were found to enrich subsets of peptides with identical terminal sequences. This approach holds the promise of a significant gain in efficiency. TXP (Triple X Proteomics) and context-independent motif specific/global proteome survey binders are variants of this concept. Principally the binding motifs of such antibodies have to be elucidated after generating these antibodies. This entails a substantial effort in the lab, as it requires synthetic peptide libraries and numerous mass spectrometry experiments. Results: We present an algorithm for predicting the antibody-binding motif in a mass spectrum obtained from a tryptic digest of a common cell line after immunoprecipitation. The epitope prediction, based on peptide mass fingerprinting, reveals the most enriched terminal epitopes. The tool provides a P-value for each potential epitope, estimated by sampling random spectra from a peptide database. The second algorithm combines the predicted sequences to more complex binding motifs. A comparison with library screenings shows that the predictions made by the novel methods are reliable and reproducible indicators of the binding properties of an antibody. Availability: Mass spectrum data, predictions, sampling tables, consensus peptide databases and the applied protocols are available as Supplementary Material. TXP-Terminus Enrichment Analysis (TEA) and MATERICS (Mass-spectrometric Analysis of Terminal Epitope Enrichment in Complex Samples) are available as web services at http://webservices.nmi.de/materics. Contact: [email protected] Supplementary information: Supplementary data are available at Bioinformatics online.

Recently, researchers showed growing interest in utilizing UHF Radio-Frequency Identification (RFID) technology for localizing tagged items with mobile robots in industrial scenarios. In this paper we present a novel three-dimensional (3D) probability sensor model of RFID antennas in the context of mapping passive RFID tags with mobile robots. The proposed 3D sensor model characterizes both detection rates and received signal strength (RSS). Compared to 2D-sensor model based approaches, the 3D model gains a higher mapping accuracy for 2D position estimation. Specially, with this sensor model, we are able to localize the tags in 3D by integrating the measurements from a pair of RFID antennas mounted at different heights of the robot. Furthermore, by integrating negative information (i.e., non-detections), the 3D mapping accuracy can be improved. Additionally, we utilize KLD-sampling to reduce the number of particles for our specific application, so that our algorithm can be performed online. Indoor experiments with a Scitos G5 robot demonstrate the effectiveness of our approach. We also provide the datasets of this work for download.

We present a computationally inexpensive RGBD-SLAM solution taylored to the application on autonomous MAVs, which enables our MAV to fly in an unknown environment and create a map of its surroundings completely autonomously, with all computations running on its onboard computer. We achieve this by implementing efficient methods for both tracking its current location with respect to a heavily processed previously seen RGBD image (keyframe) and efficient relative registration of a set of keyframes using bundle adjustment with depth constraints as a front-end for pose graph optimization. We prove the accuracy and efficiency of our system based on a public benchmark dataset and demonstrate that the proposed method enables our quadrotor to fly autonomously.

SUMMARY: Microarrays are commonly used to detect changes in gene expression between different biological samples. For this purpose, many analysis tools have been developed that offer visualization, statistical analysis and more sophisticated analysis methods. Most of these tools are designed specifically for messenger RNA microarrays. However, today, more and more different microarray platforms are available. Changes in DNA methylation, microRNA expression or even protein phosphorylation states can be detected with specialized arrays. For these microarray technologies, the number of available tools is small compared with mRNA analysis tools. Especially, a joint analysis of different microarray platforms that have been used on the same set of biological samples is hardly supported by most microarray analysis tools. Here, we present InCroMAP, a tool for the analysis and visualization of high-level microarray data from individual or multiple different platforms. Currently, InCroMAP supports mRNA, microRNA, DNA methylation and protein modification datasets. Several methods are offered that allow for an integrated analysis of data from those platforms. The available features of InCroMAP range from visualization of DNA methylation data over annotation of microRNA targets and integrated gene set enrichment analysis to a joint visualization of data from all platforms in the context of metabolic or signalling pathways. AVAILABILITY: InCroMAP is freely available as Java™ application at www.cogsys.cs.uni-tuebingen.de/software/InCroMAP, including a comprehensive user's guide and example files.

In this paper, we present the application of Evolutionary Algorithms (EAs) and linear programming for minimizing thermal impacts in the ground by operating a low-enthalpy geothermal plant with a field of multiple borehole heat exchangers (BHEs). The new methodology is demonstrated on two synthetic case studies with 36 BHEs that are grounded in reality and operated to produce given seasonal heating energy demand. We compare the performance of six different Evolutionary Algorithms (EAs) (two Differential Evolution variants, Particle Swarm Optimization, two Evolution Strategy based Algorithms, real valued Genetic Algorithm) and Monte-Carlo random search to find the optimal BHE positions. Additionally, linear programming is applied to adjust the energy extraction (loads) for the individual BHEs in the field. Both optimization steps are applied separately and in combination, and the achieved system improvements are compared to the conditions for the non-optimized case. The EAs were able to find constellations that cause less pronounced temperature changes in the subsurface (18% - 25%) than those associated with non-optimized BHE fields. Further, we could show that exclusive optimization of BHE energy extraction rates delivers slightly better results than the optimization of BHE positions. Combining both optimization approaches is the best choice and, ideally, adjusts the geothermal plant.

In this paper we investigate the effectiveness of SURF features for visual terrain classification for outdoor flying robots. A quadrocopter fitted with a single camera is flown over different terrains to take images of the ground below. Each image is divided into a grid and SURF features are calculated at grid intersections. A classifier is then used to learn to differentiate between different terrain types. Classification results of the SURF descriptor are compared with results from other texture descriptors like Local Binary Patterns and Local Ternary Patterns. Six different terrain types are considered in this approach. Random forests are used for classification on each descriptor. It is shown that SURF features perform better than other descriptors at higher resolutions.

We present a method of utilizing depth information as provided by RGBD sensors for robust real-time visual simultaneous localisation and mapping (SLAM) by augmenting monocular visual SLAM to take into account depth data. This is implemented based on the feely available software “Parallel Tracking and Mapping” by Georg Klein. Our modifications allow PTAM to be used as a 6D visual SLAM system even without any additional information about odometry or from an inertial measurement unit.

We present a novel approach incorporating a combination of Radio-Frequency Identification (RFID) and odometry information into the motion control of a mobile robot for the purpose of path following in unknown environments. Our method utilizes RFID measurements as landmarks and makes the mobile robot autonomously follow a path that was previously recorded in a manual training phase. The approach needs no prior information about RFID sensor models, the distribution and positioning of the tags nor does it require a map of the environment. Particularly, it is adaptive to different reader power levels and various tag densities, which have a major impact on RFID performance. Extensive experiments with a SCITOS G5 robot in different environments like a library, a supermarket and hallways confirm the effectiveness of our algorithm.

Hardware platforms with limited processing power are often incapable of running dense stereo analysis algorithms at acceptable speed. Sparse algorithms provide an alternative but generally lack in accuracy. To overcome this predicament, we present an efficient sparse stereo analysis algorithm that applies a dense consistency check, leading to accurate matching results. We further improve matching accuracy by introducing a new feature detector based on FAST, which exhibits a less clustered feature distribution. The new feature detector leads to a superior performance of our stereo analysis algorithm. Performance evaluation shows that the proposed stereo matching system achieves processing rates above 200 frames per second on a commodity dual core CPU, and faster than video frame-rate processing on a low-performance embedded platform. The stereo matching results prove to be superior to those obtained with ordinary sparse matching algorithms.

We present a follow-the-leader scenario with a system of two small low-cost quadrocopters of different types and configurations. The leader is a Parrot AR.Drone which is controlled by an iPad App utilizing the visual odometry provided by the quadrocopter and pilots it autonomously. The follower is an Asctec Hummingbird which is controlled by an onboard 8-bit microcontroller. Neither communication nor external sensors are required. A custom-built pan/tilt unit and the camera of a Nintendo Wii remote tracks a pattern of infrared lights and allows for online pose estimation. A base station allows for monitoring the behavior but is not required for autonomous flights. Our efficient solution of the perspective-3-point problem allows for estimating the pose of the camera relative to the pattern in six degrees of freedom at a high frequency on the microcontroller. The presented experiments include a scenario in which the follower follows the leader with a constant distance of two meters flying different shapes in narrow, GPS-denied indoor environment.

MOTIVATION: The biological pathway exchange language (BioPAX) and the systems biology markup language (SBML) belong to the most popular modeling and data exchange languages in systems biology. The focus of SBML is quantitative modeling and dynamic simulation of models, whereas the BioPAX specification concentrates mainly on visualization and qualitative analysis of pathway maps. BioPAX describes reactions and relations. In contrast, SBML core exclusively describes quantitative processes such as reactions. With the SBML qualitative models extension (qual), it has recently also become possible to describe relations in SBML. Before the development of SBML qual, relations could not be properly translated into SBML. Until now, there exists no BioPAX to SBML converter that is fully capable of translating both reactions and relations. RESULTS: The entire nature pathway interaction database has been converted from BioPAX (Level 2 and Level 3) into SBML (Level 3 Version 1) including both reactions and relations by using the new qual extension package. Additionally, we present the new webtool BioPAX2SBML for further BioPAX to SBML conversions. Compared with previous conversion tools, BioPAX2SBML is more comprehensive, more robust and more exact. AVAILABILITY: BioPAX2SBML is freely available at http://webservices.cs.uni-tuebingen.de/ and the complete collection of the PID models is available at http://www.cogsys.cs.uni-tuebingen.de/downloads/Qualitative-Models/.

MOTIVATION: Traditionally, microarrays were almost exclusively used for the genome-wide analysis of differential gene expression. But nowadays, their scope of application has been extended to various genomic features, such as microRNAs (miRNAs), proteins and DNA methylation (DNAm). Most available methods for the visualization of these datasets are focused on individual platforms and are not capable of integratively visualizing multiple microarray datasets from cross-platform studies. Above all, there is a demand for methods that can visualize genomic features that are not directly linked to protein-coding genes, such as regulatory RNAs (e.g. miRNAs) and epigenetic alterations (e.g. DNAm), in a pathway-centred manner. RESULTS: We present a novel pathway-based visualization method that is especially suitable for the visualization of high-throughput datasets from multiple different microarray platforms that were used for the analysis of diverse genomic features in the same set of biological samples. The proposed methodology includes concepts for linking DNAm and miRNA expression datasets to canonical signalling and metabolic pathways. We further point out strategies for displaying data from multiple proteins and protein modifications corresponding to the same gene. Ultimately, we show how data from four distinct platform types (messenger RNA, miRNA, protein and DNAm arrays) can be integratively visualized in the context of canonical pathways. AVAILABILITY: The described method is implemented as part of the InCroMAP application that is freely available at www.cogsys.cs.uni-tuebingen.de/software/InCroMAP. CONTACT: [email protected] or [email protected].

Higher organisms are able to respond to continuously changing external conditions by transducing cellular signals into specific regulatory programs, which control gene expression states of thousands of different genes. One of the central problems in understanding gene regulation is to decipher how combinations of transcription factors control sets of co expressed genes under specific experimental conditions. Existing methods in this field mainly focus on sequence aspects and pattern recognition, e.g., by detecting cis-regulatory modules (CRMs) based on gene expression profiling data. We propose a novel approach by combining experimental data with a priori knowledge of respective experimental conditions. These various sources of evidence are likewise considered using multi-objective evolutionary optimization. In this work, we present three objective functions that are especially designed for stimulus-response experiments and can be used to integrate a priori knowledge into the detection of gene regulatory modules. This method was tested and evaluated on whole-genome microarray measurements of drug-response in human hepatocytes.

Mapping is regarded as one of the most fundamental tasks for mobile robots. In this work, we present an approach that enables multiple resource-limited mobile robots to cooperatively build an image-based map of the environment and to afterwards localize in it. To achieve this, we deploy a hierarchical team of mobile robots. A parent robot possesses state-of-the-art sensors, computation power and acts as a leader. It teleoperates small child robots within its line-of-sight. In contrast to other approaches and due to the cooperation among the robots, we can relax the requirement that every robot must be able to self-localize to take part in multi-robot mapping. Additionally, our algorithm ensures the mapping of the entire area in an efficient way, i.e., it fulfills the requirements of area coverage. To test our approach, extensive experiments have been performed both in simulation and real-world. In the latter case, a team of four heterogeneous mobile robots was deployed. Besides the successful cooperation in the robot team, localization results are presented to validate the applicability of the proposed mapping procedure.

SUMMARY: The specifications of the Systems Biology Markup Language (SBML) define standards for storing and exchanging computer models of biological processes in text files. In order to perform model simulations, graphical visualizations and other software manipulations, an in-memory representation of SBML is required. We developed JSBML for this purpose. In contrast to prior implementations of SBML APIs, JSBML has been designed from the ground up for the Java programming language, and can therefore be used on all platforms supported by a Java Runtime Environment. This offers important benefits for Java users, including the ability to distribute software as Java Web Start applications. JSBML supports all SBML Levels and Versions through Level 3 Version 1, and we have strived to maintain the highest possible degree of compatibility with the popular library libSBML. JSBML also supports modules that can facilitate the development of plugins for end user applications, as well as ease migration from a libSBML-based backend. AVAILABILITY: Source code, binaries and documentation for JSBML can be freely obtained under the terms of the LGPL 2.1 from the website http://sbml.org/Software/JSBML.

MOTIVATION: Statins are the most widely used cholesterol-lowering drugs. The primary target of statins is HMG-CoA reductase, a key enzyme in cholesterol synthesis. However, statins elicit pleitropic responses including beneficial as well as adverse effects in the liver or other organs. Today, the regulatory mechanisms that cause these pleiotropic effects are not sufficiently understood. RESULTS: In this work, genome-wide RNA expression changes in primary human hepatocytes of six individuals were measured at up to six time points upon atorvastatin treatment. A computational analysis workflow was applied to reconstruct regulatory mechanisms based on these drug-response data and available knowledge about transcription factor (TF) binding specificities and protein-drug interactions. Several previously unknown TFs were predicted to be involved in atorvastatin-responsive gene expression. The novel relationships of nuclear receptors NR2C2 and PPARA on CYP3A4 were successfully validated in wet-lab experiments. AVAILABILITY: Microarray data are available at the Gene Expression Omnibus (GEO) database at www.ncbi.nlm.nih.gov/geo/, under accession number GSE29868. CONTACT: [email protected]; [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

SUMMARY: The KEGG PATHWAY database provides a widely used service for metabolic and nonmetabolic pathways. It contains manually drawn pathway maps with information about the genes, reactions and relations contained therein. To store these pathways, KEGG uses KGML, a proprietary XML-format. Parsers and translators are needed to process the pathway maps for usage in other applications and algorithms. We have developed KEGGtranslator, an easy-to-use stand-alone application that can visualize and convert KGML formatted XML-files into multiple output formats. Unlike other translators, KEGGtranslator supports a plethora of output formats, is able to augment the information in translated documents (e.g. MIRIAM annotations) beyond the scope of the KGML document, and amends missing components to fragmentary reactions within the pathway to allow simulations on those. AVAILABILITY: KEGGtranslator is freely available as a Java(™) Web Start application and for download at http://www.cogsys.cs.uni-tuebingen.de/software/KEGGtranslator/. KGML files can be downloaded from within the application. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Geothermal energy use from shallow groundwater systems is attractive for the supply of heat and hot water to buildings. It offers economic and environmental advantages over traditional fossil-fuel based technologies, in particular when large scale systems are well adapted to the always unique hydrogeological conditions. Computer based numerical simulations are used to examine the performance of multiple borehole heat exchangers installed in the ground. This paper demonstrates how evolutionary algorithms can be utilized to configure the elements of a geothermal system in an ideal way, and thus substantially enhance the energy extraction rate in comparison to standardized approaches. Differential evolution (DE), evolution strategies (ES) and particle swarm optimizers (PSO) are combined with a local search approach and compared with respect to their efficiency in the optimization of synthetic, real case oriented and static systems. First results are promising, especially for the PSO and the DE with the local search approach.

The scaling properties of multimodal optimization methods have seldom been studied, and existing studies often concentrated on the idea that all local optima of a multimodal function can be found and their number can be estimated a priori. We argue that this approach is impractical for complex, high-dimensional target functions, and we formulate alternative criteria for scalable multimodal optimization methods. We suggest that a scalable niching method should return the more local optima the longer it is run, without relying on a fixed number of expected optima. This can be fulfilled by sequential and semi-sequential niching methods, several of which are presented and analyzed in that respect. Results show that, while sequential local search is very successful on simpler functions, a clustering-based particle swarm approach is most successful on multi-funnel functions, offering scalability even under deceptive multimodality, and denoting it a starting point towards effective scalable niching.

The determination of crop production functions which describe the relationship between irrigation water and crop yield under the assumption of optimal irrigation scheduling is a major building block for a more efficient and sustainable water management. In this paper we introduce a methodology to determine the entire crop production function for a given scenario within a single run of a multi-objective evolutionary algorithm. Further we compare the performance of four major algorithms (NSGA-II, NSDE, DEMO, and MO-CMA-ES), and a single-objective approach based on differential evolution on three different scenarios and two different population initialization methods on this problem. We show that the combination of a problem specific initialization with MO-CMA-ES is able to determine crop production functions which are extremely close to actual ones.

In the domain of agricultural robotics, one major application is crop scouting, e.g., for the task of weed control. For this task a key enabler is a robust detection and classification of the plant and species. Automatically distinguishing between plant species is a challenging task, because some species look very similar. It is also difficult to translate the symbolic high level description of the appearances and the differences between the plants used by humans, into a formal, computer understandable form. Also it is not possible to reliably detect structures, like leaves and branches in 3D data provided by our sensor. One approach to solve this problem is to learn how to classify the species by using a set of example plants and machine learning methods. In this paper we are introducing a method for distinguishing plant species using a 3D LIDAR sensor and supervised learning. For that we have developed a set of size and rotation invariant features and evaluated experimentally which are the most descriptive ones. Besides these features we have also compared different learning methods using the toolbox Weka. It turned out that the best methods for our application are simple logistic regression functions, support vector machines and neural networks. In our experiments we used six different plant species, typically available at common nurseries, and about 20 examples of each species. In the laboratory we were able to identify over 98% of these plants correctly.

We present a novel approach which enables a mobile robot to estimate its trajectory in an unknown environment with long-range passive radio-frequency identification (RFID). The estimation is based only on odometry and RFID measurements. The technique requires no prior observation model and makes no assumptions on the RFID setup. In particular, it is adaptive to the power level, the way the RFID antennas are mounted on the robot, and environmental characteristics, which have major impact on long-range RFID measurements. Tag positions need not be known in advance, and only the arbitrary, given infrastructure of RFID tags in the environment is utilized. By a series of experiments with a mobile robot, we show that trajectory estimation is achieved accurately and robustly.

A safe traversal of a mobile robot in an unknown environment requires the determination of local ground surface properties. As a first step, a broad structure of the underlying environment can be established by clustering terrain sections which exhibit similar features. In this work, we focus on an unsupervised learning approach to segment different terrain types according to the clustering of acquired vibration signals. Therefore, we present a Markov random field-based clustering approach taking the inherent temporal dependencies between consecutive measurements into account. The applied generative model assumes that the class labels of neighboring vibration segments are generated by prior distributions with similar parameters. A temporally constrained expectation maximization algorithm enables the efficient estimation of its parameters considering a predefined set of neighboring vibration segments. Since the size of the neighbor set proves to be data-dependent, we derive a general means of estimating this set size from the observed data. We show that the Markov random field clustering approach generates valid models for a variety of driving speeds even in situations of frequent terrain changes.

BACKGROUND: Bioinformatics data analysis often deals with additive mixtures of signals for which only class labels are known. Then, the overall goal is to estimate class related signals for data mining purposes. A convenient application is metabolic monitoring of patients using infrared spectroscopy. Within an infrared spectrum each single compound contributes quantitatively to the measurement. RESULTS: In this work, we propose a novel factorization technique for additive signal factorization that allows learning from classified samples. We define a composed loss function for this task and analytically derive a closed form equation such that training a model reduces to searching for an optimal threshold vector. Our experiments, carried out on synthetic and clinical data, show a sensitivity of up to 0.958 and specificity of up to 0.841 for a 15-class problem of disease classification. Using class and regression information in parallel, our algorithm outperforms linear SVM for training cases having many classes and few data. CONCLUSIONS: The presented factorization method provides a simple and generative model and, therefore, represents a first step towards predictive factorization methods.

In this paper, the path following problem of an omnidirectional mobile robot has been studied. Given the error dynamic model derived from the robot state vector and the path state vector, model predictive control (MPC) is employed to design the control law, which can deal explicitly with the rate of progression of a virtual vehicle to be followed along the path. The distinct advantage over other control strategies is that input and system constraints are able to be handled straightforwardly in the optimization problem so that the robot can travel safely with a high velocity. Unlike nonholonomic mobile robots, omnidirectional mobile robots, which we focus on in this paper, have simultaneously and independently controlled rotational and translational motion capabilities. Then, our purposed MPC controller was validated by experiments with a real omnidirectional mobile robot.

Outdoor robots are faced with a variety of terrain types each possessing different characteristics. To ensure a safe traversal a robot has to infer the current ground surface from sensor readings. Recent techniques generate a model which predicts the terrain class from single vibration signals disregarding the temporal coherence between consecutive measurements. In this paper, we present a novel approach in which the final classification relies on the analysis of not only one, but several recent observations. Therefore, the probabilistic framework of the Bayes filter is adopted to the problem of terrain classification. We propose an adaptive approach which automatically adjusts its parameters according to the history of observations. To demonstrate the performance of our method we further describe and compare another technique based on temporal coherence. The evaluation using data collected from our RWI ATRV-Jr robot shows that our approach is both reactive and stable enough to detect fast terrain transitions and selective misclassifications.

One key principle of natural immune systems is the extracellular presentation of peptides bound to MHC-II complexes on the cell surface to represent the internal state. The prediction of those peptides that are presented became a current research topic in machine learning, as they may be used as potential vaccines for immunization. In addition the biological immune system (IS) is a learning system in its own right. In this work, we design an artificial immune system (AIS) that is based on observations of the natural immune system to predict MHC-II binding peptides. Our strategy simulates the mutable receptors of T lymphocytes as well as their selection during life time. We model the receptor specificity and binding mode as well as the lymphocyte's influence during an inflammatory response. Finally, our implementation uses the pathogen specificity of T cells to model the prediction problem.

In this paper we present a novel approach to estimating the trajectory of a robot by means of inexpensive passive RFID tags and odometry in unknown environments. We show how trajectory estimation, a prerequisite of mapping RFID transponder positions without a reference positioning system, can be achieved using a particle filter. The presented technique is based on a non-parametric model of spatial relationships between RFID measurements. It overcomes the noisy nature of RFID measurements and the absence of distance and bearing information. The accuracy of our method is investigated in a series of experiments with a mobile robot.

SUMMARY: The XML-based Systems Biology Markup Language (SBML) has emerged as a standard for storage, communication and interchange of models in systems biology. As a machine-readable format XML is difficult for humans to read and understand. Many tools are available that visualize the reaction pathways stored in SBML files, but many components, e.g. unit declarations, complex kinetic equations or links to MIRIAM resources, are often not made visible in these diagrams. For a broader understanding of the models, support in scientific writing and error detection, a human-readable report of the complete model is needed. We present SBML2L(A)T(E)X, a Java-based stand-alone program to fill this gap. A convenient web service allows users to directly convert SBML to various formats, including DVI, L(A)T(E)X and PDF, and provides many settings for customization. AVAILABILITY: Source code, documentation and a web service are freely available at (http://www.ra.cs.uni-tuebingen.de/software/SBML2LaTeX).

Reconstructing and modeling regulatory networks is an active area of research in bioinformatics and systems biology. Hence, various computational methods have been published, often successfully modeling one aspect of regulatory control. Gene regulation, however, is a process that depends on many different components such as transcription factors (TFs), cis-regulatory motifs and their temporal and spatial coordination. Accordingly, a promising new direction for computational analysis is the incorporation of multiple data types to discover, for instance, cluster membership, the spatial organization of cis-regulatory motifs and TFs that bind to these motifs. Here, we present such a data-driven framework, comprising four stages, to infer gene regulatory networks (GRNs) by modeling: 1. motif presence in the promoter, 2. spatial motif arrangement in co-regulated genes, 3. TFs that bind the respective motifs, and 4. dynamic properties of the GRN. A novel method is presented in stage 2, where we optimize for the spatial motif properties: orientation, occurrence of multiple motifs, relative distance between two motifs and distance to the Transcription Start Site (TSS). To find optimal distance based properties in efficient time we describe a dynamic programming approach. To combine multiple motif properties that are shared by genes with similar expression profiles a Hill-climber is employed. Subsequently, in stage 3 and 4, we infer GRNs by assigning TFs to the derived motifs and model time-dependent regulatory relationships between them with the Inferelator approach. None of the stages require the user to manually adjust any parameter, and thus derived properties can be analyzed without the bias introduced by parametrization. We applied this approach to S. cerevisiae data and obtained insight into individual and general properties of the spatial assembly of regulatory elements and inferred the corresponding GRN.

This paper addresses the problem of object classification in a biosonar based mobile robot in a natural environment using a boosting method. We present an algorithm based on gradient boosting for biosanar-based robots that recognize different objects such as different trees via reflected sonar echoes. Gradient boosting is a machine learning approach, that builds one strong classifier from many base learners. We present two kinds of base learners for the gradient boosting: ordinary least squares (OLS) and kernel-based base learners. Compared with our previous works, in which we presented a time resolved spectrum kernel to extract the similarities between echoes, we get more efficient and accurate results with the newly proposed boosting method. We compare the methods in terms of sensitivity, specificity, accuracy and Matthew's correlation coefficient and also the runtime of training and testing.

This paper focuses on the dribbling control problem of an omnidirectional mobile robot and a rolling ball in the RoboCup Middle Size domain. Because the ball easily slides away from the robot when the ball moves along a curve, dribbling control is more challenging than the normal mobile robot motion control problem. Based on an introduced reference point with respect to the robot body and a sophisticated planning method of robot pose, the nonlinear predictive control is used to steer the robot to follow the planned poses so as to prevent the ball from leaving the robot. Real-world experiments showed that nonlinear predictive control is capable of solving the pose following problem in a real-time application.

In this paper, we implement a new method for classification of biological signals in general, and use it in the animal behavior classification as an example. The forced swimming test of rats or mice is a frequently used behavioral test to evaluate the efficacy of drugs in rats or mice. Frequently used features for that evaluation are obtained through observing three states: immobility, struggling/climbing and swimming in activity profiles. We consider that those activity profiles (signals) inherently contain undesired and interference noise that should be removed before feature extraction and classification. We use a Finite Impulse Response (FIR) filter to filter out that additive noise from the activity profile. The parameters of the FIR filter are obtained via maximizing the accuracy of a classifier that tries to make a discrimination between two classes of the activity profiles (e.g. drug vs. control). We use the kernel Fisher discriminant criterion as a criterion for the discrimination, the DIviding RECTangles (DIRECT) search method for solving the optimization problem and Support Vector Machines (SVMs) for the classification task. We show that Autoregressive (AR) coefficients are suitable features for the extraction of the dynamic behavior of rats and also the classification of activity profiles. Our proposed behavior classification method provides a reliable discrimination of different classes of antidepressant drugs (imipramine and desipramine) administered to rats versus a vehicle-treated group.

This paper presents a solution to the problem of steering a group of real omnidirectional mobile robots along a given path, while maintaining a desired formation pattern. This problem can be divided into a leader agent subproblem and a follower agent subproblem such that a leader agent follows a given path and each follower agent tracks a trajectory, estimated by using the leaderpsilas information. In this paper, we exploit nonlinear model predictive control (NMPC) as a local control law for real-world experiments due to its advantages of taking the robot constraints and future information into account. To solve the path following problem for the leader agent, we propose to integrate the rate of progression of a virtual vehicle to be followed along that path into the local cost function of NMPC. After the open-loop optimization problem is solved, the optimal rate of progression at each time step in the future is obtained. This information and the leaderpsilas current state are broadcasted to all follower agents. With respect to a desired formation configuration and a reference path, each follower agent can estimate its own reference trajectory by using the leaderpsilas information and its time stamp. NMPC is also employed as a local control law to steer the follower agent to track that reference trajectory. Our approach was validated by experiments using three omnidirectional mobile robots.

In this paper we show that, despite some disadvantageous properties of radio frequency identification (RFID), it is possible to localize a mobile robot quite accurately in environments which are densely tagged. We therefore employ a recently presented probabilistic fingerprinting technique called RFID snapshots. This method interprets short series of RFID measurements as feature vectors and is able to position a mobile robot after a training phase. It requires no explicit sensor model and is capable of exploiting given tag infrastructures, e.g., provided by supermarket shelves containing labeled products.

For safe navigation in outdoor environments, a mobile robot should be able to estimate the type of the current and forthcoming terrain. Based on this estimation, the robot can decide if the terrain is safe and may be traversed securely, or if the terrain is potentially dangerous and must be avoided or traversed carefully. This paper presents a terrain classification approach which fuses terrain predictions based on image data with predictions made by a vibration-based method. Using color images, the robot classifies terrain in front of it. When the robot later traverses the classified area, it uses vibration data to verify its former prediction. Our experiments on 14 different terrain types show that by fusing both predictions, the classification rates are significantly larger than predictions based on data from a single sensor alone.

This paper considers the problem of steering and coordinating a group of omnidirectional mobile robots along given paths. Two subproblems, i.e., a path following subproblem and a motion coordination subproblem are solved by using distributed nonlinear model predictive control (NMPC) whose cost function is coupled with neighbors. The distinct features of NMPC are that constraints can be explicitly accommodated, as well as nonlinear and time-varying systems can be easily handled. An idea of morphing is employed to generate intermediate configurations between a source and a target configuration with a desired transition rate. Each mobile robot adjusts its own speed along its reference path to achieve motion coordination. Experiments with three omnidirectional mobile robots are presented to illustrate the validity of our proposed method. Unlike nonholonomic mobile robots, omnidirectional mobile robots have simultaneously and independently controlled rotational and translational motion capabilities.

An important problem in systems biology is parameter estimation for biochemical system models. Our work concentrates on the metabolic subnetwork of the valine and leucine biosynthesis in Corynebacterium glutamicum, an anaerobic actinobacterium of high biotechnological importance. Using data of an in vivo experiment measuring 13 metabolites during a glucose stimulus-response experiment we investigate the performance of various evolutionary algorithms on the parameter inference problem in biochemical modeling. Due to the inconclusive information on the reversibility of the reactions in the pathway, we develop both a reversible and an irreversible differential equation model based on the recent convenience kinetics approach. As the reversible model allows better approximation on the whole, we use it to analyze the impact of different settings on four especially promising EAs. We show that Particle Swarm Optimization as well as Differential Evolution are useful methods for parameter estimation on convenience kinetics models outperforming Genetic Algorithm and Evolution Strategy approaches and nearly reaching the quality of independent spline approximations on the raw data.

Parameter estimation for biochemical model systems has become an important problem in systems biology. Here we focus on the metabolic subnetwork of the valine and leucine biosynthesis in C. glutamicum. Due to the lack of indisputable information regarding reversibility of the reactions in the pathway we derived two alternative ordinary differential equation models based on the formalisms of the generalized mass-action rate law. We introduced two alternative modeling approaches for feedback inhibition and evaluated the applicability of six optimization procedures (multi start Hill Climber, binary and real valued Genetic Algorithm, standard and covariance matrix adaption Evolution Strategy as well as Simulated Annealing) to the problem of parameter fitting. The model considering irreversible reactions performed worse and was therefore rejected from further analysis. We benchmarked the impact of different mutation and crossover operators as well as the influence of the population size on the remaining system and the two best optimization procedures namely binary Genetic Algorithm and the Evolution Strategy. The GA performed best on average and found the best total result based on the relative squared error.

This paper focuses on the dribbling control problem of an omnidirectional mobile robot. Because the movement of the dribbled object must be considered, dribbling control is more challenging than normal mobile robot motion control. A new feedback control algorithm, which steers a reference point to follow the desired movement and keeps the ball near to this point simultaneously, is proposed. To dribble a rolling ball along a given path, the robot should provide the ball with appropriate force by consecutive pushing operations when they travel in an environment with obstacles. Based on the analysis of the forces acting on the ball with respect to the mobile robot coordinate system, a constraint for robot movement in the dribbling process is also introduced. The simulation and real-world experiments address the performance of this control algorithm.

Vision-based robot localization in outdoor environments is difficult because of changing illumination conditions. Another problem is the rough and cluttered environment which makes it hard to use visual features that are not rotation invariant. A popular method that is rotation invariant and relatively robust to changing illumination is the Scale Invariant Feature Transform (SIFT). However, due to the computationally intensive feature extraction and image matching, localization using SIFT is slow. On the other hand, techniques which use global image features are in general less robust and exact than SIFT, but are often much faster due to fast image matching. In this paper, we present a hybrid localization approach that switches between local and global image features. For most images, the hybrid approach uses fast global features. Only in difficult situations, e.g. containing strong illumination changes, the hybrid approach switches to local features. To decide which features to use for an image, we analyze the particle cloud of the particle filter that we use for position estimation. Experiments on outdoor images taken under varying illumination conditions show that the position estimates of the hybrid approach are about as exact as the estimates of SIFT alone. However, the average localization time using the hybrid approach is more than 3.5 times faster than using SIFT.

BACKGROUND: Cells dynamically adapt their gene expression patterns in response to various stimuli. This response is orchestrated into a number of gene expression modules consisting of co-regulated genes. A growing pool of publicly available microarray datasets allows the identification of modules by monitoring expression changes over time. These time-series datasets can be searched for gene expression modules by one of the many clustering methods published to date. For an integrative analysis, several time-series datasets can be joined into a three-dimensional gene-condition-time dataset, to which standard clustering or biclustering methods are, however, not applicable. We thus devise a probabilistic clustering algorithm for gene-condition-time datasets. RESULTS: In this work, we present the EDISA (Extended Dimension Iterative Signature Algorithm), a novel probabilistic clustering approach for 3D gene-condition-time datasets. Based on mathematical definitions of gene expression modules, the EDISA samples initial modules from the dataset which are then refined by removing genes and conditions until they comply with the module definition. A subsequent extension step ensures gene and condition maximality. We applied the algorithm to a synthetic dataset and were able to successfully recover the implanted modules over a range of background noise intensities. Analysis of microarray datasets has lead us to define three biologically relevant module types: 1) We found modules with independent response profiles to be the most prevalent ones. These modules comprise genes which are co-regulated under several conditions, yet with a different response pattern under each condition. 2) Coherent modules with similar responses under all conditions occurred frequently, too, and were often contained within these modules. 3) A third module type, which covers a response specific to a single condition was also detected, but rarely. All of these modules are essentially different types of biclusters. CONCLUSION: We successfully applied the EDISA to different 3D datasets. While previous studies were mostly aimed at detecting coherent modules only, our results show that coherent responses are often part of a more general module type with independent response profiles under different conditions. Our approach thus allows for a more comprehensive view of the gene expression response. After subsequent analysis of the resulting modules, the EDISA helped to shed light on the global organization of transcriptional control. An implementation of the algorithm is available at http://www-ra.informatik.uni-tuebingen.de/software/IAGEN/.

A common practical problem in mobile robotics is the task to calibrate the robot's sensors. Although, the general mapping of the sensor data to robot-centered world coordinates is given by the hardware configuration, the parameters of this mapping vary even between robots with the same configuration. In the RoboCup domain, these parameters can change drastically after transport or physical contact during game play. It is therefore necessary to recalibrate the robots for their next assignment within a few minutes, not only in order to fulfill the future regulatory requirements of the RoboCup organization committee to keep the setup time as low as possible. As camera systems, especially omni-directional systems, are currently the most important sensors in RoboCup, a reliable and fast calibration method for the mapping of image to world coordinates is necessary. Since the RoboCup environment, i.e. the soccer field, has known dimensions and is also static, automatic calibration using the features and landmarks of the soccer field is possible if the robot is given an image from a known pose. In this paper, an efficient evolutionary approach to automatic camera calibration is presented, which is independent of the hardware configuration. It only requires a quality function for the parameter settings, which allows lazy evaluation. To meet the time constraints given for this real-world optimization problem, a novel mutation operator is introduced to enhance the performance of the evolutionary algorithm. It samples a number of alternative solutions using a high rate of lazy evaluation, before deciding on the true mutative change applied on the given individual. This new mutation operator proves to be fast and most reliable on the camera to world calibration problem.

During the last years, high throughput experiments have become very popular. During the analysis of such data the need for a functional grouping of genes arises. In this paper, we propose grouping genes according to their biological function by means of kernel functions, which are similarity measures having special mathematical properties and play a crucial role e.g. in SVM classification. Thereby our kernel functions rely on functional information on the genes provided by Gene Ontology annotation. We investigate and compare several provably symmetric, positive semidefinite kernel functions in combination with spectral clustering, dual k-means and average linkage and demonstrate that our approach leads to good clustering results.

Self-localization is a major research task in mobile robotics for several years. Efficient self-localization methods have been developed, among which probabilistic Monte-Carlo localization (MCL) is one of the most popular. It enables robots to localize themselves in real-time and to recover from localization errors. However, even those versions of MCL using an adaptive number of samples need at least a minimum in the order of 100 samples to compute an acceptable position estimation. This paper presents a novel approach to MCL based on images from an omnidirectional camera system. The approach uses an adaptive number of samples that drops down to a single sample if the pose estimation is sufficiently accurate. We show that the method enters this efficient tracking mode after a few cycles and remains there using only a single sample for more than 90% of the cycles. Nevertheless, it is still able to cope with the kidnapped robot problem

Many approaches for extracting landmarks and objects from a camera image based on their color coding were published in the RoboCup domain. They are quite sophisticated and tuned to the typical RoboCup scenario of constant bright lighting using a static subdivision of the color space into different color classes. However, such algorithms would soon be of limited use, as the future requirements of RoboCup include the possibility to play under changing and finally natural lighting. This paper presents an algorithm for automatic online color training, which is able to robustly adapt the mapping of colors to color classes onto different lighting situations online. Using the ACT algorithm a robot will be able to play a RoboCup match while the illumination of the field varies

In outdoor environments, there is a variety of different types of ground surfaces. If some of them are slippery or bumpy, for example, the ground surface itself is a possible hazard for an autonomous mobile vehicle traversing the surface. Therefore, it is beneficial if the vehicle is able to estimate, which terrain it is currently traversing. Using this estimation, the vehicle can adapt its driving style to the terrain. In this paper, we present a method for terrain classification based on vibration induced in the vehicle's body. An accelerometer mounted on the vehicle measures the vibration perpendicular to the ground surface. We experimentally compare representations of the data based on the fast Fourier transform (FFT) and on the power spectral density (PSD). Additionally, we suggest a simpler and more compact representation based on features calculated from the raw data vectors and a combination of this representation with the PSD. We train and classify the data with a support vector machine (SVM). Experiments on a large real-world dataset containing seven different terrain types evaluate our approach

MOTIVATION: JCell is a Java-based application for reconstructing gene regulatory networks from experimental data. The framework provides several algorithms to identify genetic and metabolic dependencies based on experimental data conjoint with mathematical models to describe and simulate regulatory systems. Owing to the modular structure, researchers can easily implement new methods. JCell is a pure Java application with additional scripting capabilities and thus widely usable, e.g. on parallel or cluster computers. AVAILABILITY: The software is freely available for download at http://www-ra.informatik.uni-tuebingen.de/software/JCell.

In this paper we address the problem of finding valid solutions for the problem of inferring gene regulatory networks. Different approaches to directly infer the dependencies of gene regulatory networks by identifying parameters of mathematical models can be found in literature. The problem of reconstructing regulatory systems from experimental data is often multimodal and thus appropriate optimization strategies become necessary. Thus, we propose to use a clustering based niching evolutionary algorithm to maintain diversity in the optimization population to prevent premature convergence and to raise the probability of finding the global optimum by identifying multiple alternative networks. With this set of alternatives, the identification of the true solution has then to be addressed in a second post-processing step

In this paper, we address the problem of finding gene regulatory networks from experimental DNA microarray data. The problem often is multi-modal and therefore appropriate optimization strategies become necessary. We propose to use a clustering based niching evolutionary algorithm to maintain diversity in the optimization population to prevent premature convergence and to raise the probability of finding the global optimum by identifying multiple alternative networks than standard algorithms. With this set of alternatives, the identification of the true solution has then to be addressed in a second post-processing step.

We propose a new kernel function for attributed molecular graphs, which is based on the idea of computing an optimal assignment from the atoms of one molecule to those of another one, including information on neighborhood, membership to a certain structural element and other characteristics for each atom. As a byproduct this leads to a new class of kernel functions. We demonstrate how the necessary computations can be carried out efficiently. Compared to marginalized graph kernels our method in some cases leads to a significant reduction of the prediction error. Further improvement can be gained, if expert knowledge is combined with our method. We also investigate a reduced graph representation of molecules by collapsing certain structural elements, like e.g. rings, into a single node of the molecular graph.

This work addresses the capability to use concentration gridmaps to locate a static gas source. In previous works it was found that depending on the shape of the mapped gas distribution (corresponding to different airflow conditions) the gas source location can be sometimes approximated with high accuracy by the maximum in the concentration map while this is not possible in other cases. This paper introduces a method to distinguish both cases by analysing the shape of the obtained concentration map in terms of a model of the time-averaged gas distribution known from physics. The parameters of the model that approximates the concentration map most closely are determined by nonlinear least squares fitting using evolution strategies (ES). The best fit also provides a better estimate of the gas source position in situations where the concentration maximum estimate fails. Different methods to select the most truthful estimate are introduced in this work and a comparison regarding their accuracy is presented, based on a total of 34h of concentration mapping experiments.

During the last years kernel methods like the support vector machine (SVM) have gained a growing interest in machine learning. One of the strengths of this approach is the ability to deal easily with arbitrarily structured data by means of the kernel function. In this paper we propose a kernel for chemical compounds which is based on the idea of computing optimal assignments between atoms of two different molecules including information about their neighborhood. As a byproduct this leads to a new class of kernel functions. We demonstrate how the necessary computations can be carried out efficiently. We compare our method against the marginalized graph kernels by Kashima et al. and show its good performance on classifying toxicological and human intestinal absorption data.

Support vector machines (SVMs) have become one of the most popular methods in machine learning during the last years. A special strength is the use of a kernel function to introduce nonlinearity and to deal with arbitrarily structured data. Usually the kernel function depends on certain parameters, which, together with other parameters of the SVM, have to be tuned to achieve good results. However, finding good parameters can become a real computational burden as the number of parameters and the size of the dataset increases. In this paper we propose an algorithm to deal with the model selection problem, which is based on the idea of learning an online Gaussian process model of the error surface in parameter space and sampling systematically at points for which the so called expected improvement is highest. Our experiments show that on this way we can find good parameters very efficiently.

With the invention of high throughput methods, researchers are capable of producing large amounts of biological data. During the analysis of such data the need for a functional grouping of genes arises. In this paper, we propose a new method based on spectral clustering for the partitioning of genes according to their biological function. The functional information is based on Gene Ontology annotation, a mechanism to capture functional knowledge in a shareable and computer processable form. Our functional cluster method promises to automates, speed up and therefore improve biological data analysis.

With the invention of microarrays, researchers are capable of measuring thousands of gene expression levels in parallel at various time points of the biological process. To investigate general regulatory mechanisms, biologists cluster genes based on their expression patterns. In this paper, we propose a new memetic co-clustering algorithm for expression profiles, which incorporates a priori knowledge in the form of gene ontology information. Ontologies offer a mechanism to capture knowledge in a shareable form that is also processable by computers. The use of this additional annotation information promises to improve biological data analysis and simplifies the identification of processes that are relevant under the measured conditions.

In this paper we address the problem of finding gene regulatory networks from artificial data sets of DNA microarray experiments. Some researchers suggested evolutionary algorithms for this purpose. We suggest to use an enhancement for evolutionary algorithms to infer the parameters of the nonlinear system given by the observed data more reliably and precisely. At present, we use S-Systems because they are a general mathematical model for simulating the complex interactions of gene regulatory networks. Due to the limited number of available data, the inferring problem is highly under-determined and ambiguous. Further on, the problem often is highly multi-modal and therefore appropriate optimization strategies become necessary. We propose to use an island model to maintain diversity in the EA population to prevent premature convergence and to raise the probability of finding the global optimum.

In this paper, we address the problem of finding gene regulatory networks from experimental DNA microarray data. As underlying mathematical model we used S-Systems, a quantitative model, which recently has found increased attention in the literature. Due to the complexity of the inference problem some researchers suggested evolutionary algorithms for this purpose. We introduce enhancements to this optimization process to infer the parameters of sparsely connected non-linear systems given by the observed data more reliably and precisely. Due to the limited number of available data the inferring problem is under-determined and ambiguous. Further on, the problem often is multi-modal and therefore appropriate optimization strategies become necessary. In this paper, we propose a new method, which evolves the topology as well as the parameters of the mathematical model to find the correct network. This method is compared to standard algorithms found in the literature.

In this paper we investigate the impact of different crossover operators for a real-valued evolutionary algorithm on the constrained portfolio selection problem based on the Markowitz mean-variance model. We also introduce an extension of a real-valued genotype, which increases the performance of the evolutionary algorithm significantly, independent of the crossover operator used. This extension is based on the effect that most efficient portfolios only consist of a selection of few assets. Therefore, the portfolio selection problem is actually a combination of a knapsack and continuous parameter problem. We also introduce a repair mechanism and examine the impact of Lamarckism on the performance of the evolutionary algorithm.

Recently, P. Viola and M.J. Jones (2001) presented a method for real-time object detection in images using a boosted cascade of simple features. In This work we show how an evolutionary algorithm can be used within the Adaboost framework to find new features providing better classifiers. The evolutionary algorithm replaces the exhaustive search over all features so that even very large feature sets can be searched in reasonable time. Experiments on two different sets of images prove that by the use of evolutionary search we are able to find object detectors that are faster and have higher detection rates.

Evolutionary algorithms (EA) are excellent optimization tools for complex high-dimensional multimodal problems. However, they require a very large number of problem function evaluations. In many engineering and design optimization problems a single fitness evaluation is very expensive or time consuming. Therefore, standard evolutionary computation methods are not practical for such applications. Applying models as a surrogate of the true fitness function is a quite popular approach to handle this restriction. It is straightforward that the success of this approach depends highly on the quality of the approximation model. We propose a controlled model assisted evolution strategy (C-MAES), which uses a support vector regression (SVR) approximation by preselecting the most promising individuals. The model assistance on the evolutionary optimization process is dynamically controlled by a model quality based on the number of correctly preselected individuals. Numerical results from extensive simulations on high dimensional test functions including noisy functions and noisy functions with changing noise level are presented. The proposed C-MAES algorithm with controlled model assistance has a much better convergence rate and achieves better results than the model assisted algorithms without model control.

With the invention of high throughput methods, researchers are capable of producing large amounts of biological data. During the analysis of such data the need of a functional grouping of genes arises. We propose a new clustering algorithm for the partition of genes or gene products according to their known biological function based on Gene Ontology terms. Ontologies offer a mechanism to capture knowledge in a shareable form that is also processable by computers. Our functional cluster algorithm promises to automatize, speed up and therefore improve biological data analysis.

As a sub-task of the general gas source localisation problem, gas source declaration is the process of determining the certainty that a source is in the immediate vicinity. Due to the turbulent character of gas transport in a natural indoor environment, it is not sufficient to search for instantaneous concentration maxima, in order to solve this task. Therefore, this paper introduces a method to classify whether an object is a gas source or not from a series of concentration measurements, recorded while the robot performs a rotation manoeuvre in front of a possible source. For three different gas source positions, a total of 288 declaration experiments were carried out at different robot-to-source distances. Based on these readings, two machine learning techniques (ANN, SVM) were evaluated in terms of their classification performance. With learning parameters that were optimised by grid search, a maximal hit rate of approximately 87.5% could be obtained using a support vector machine.

In This work we present a novel feature selection algorithm for SVMs which works by decreasing the regularized risk in an iterative manner by using a combination of a backward elimination procedure together with an exchange algorithm. It is applicable to linear as well as to nonlinear problems. We test this new algorithm on toy and real life data sets and show its good performance in comparison to state-of-the-art feature selection methods.

As a sub-task of the general gas source localisation problem, gas source declaration is the process of determining the certainty that a source is in the immediate vicinity. Due to the turbulent character of gas transport in a natural indoor environment, it is not sufficient to search for instantaneous concentration maxima, in order to solve this task. Therefore, this paper introduces a method to classify whether an object is a gas source from a series of concentration measurements, recorded while the robot performs a rotation manoeuvre in front of a possible source. For three different gas source positions, a total of 1056 declaration experiments were carried out at different robot-to-source distances. Based on these readings, support vector machines (SVM) with optimised learning parameters were trained and the cross-validation classification performance was evaluated. The results demonstrate the feasibility of the approach to detect proximity to a gas source using only gas sensors. The paper also presents an analysis of the classification rate depending on the desired declaration accuracy, and a comparison with the classification rate that can be achieved by selecting an optimal threshold value regarding the mean sensor signal.

We show here that it is possible to track a face or an object in real-time using usual PC hardware, even if the object quickly changes its rotation and scale. An algorithm based on the discriminator technique (used earlier in analogue signal processing) was developed. Tracking control was realized by a scale factor and roll angle discriminator control loops and convolution based 2D cross-correlation. Experiments revealed that this method enables real-time scale and rotation invariant tracking control of an object or face template in a wide scale range and provides robustness against high frequency camera vibrations, from which cameras on mobile robots suffer.

The aim of this article is to present the potential of kernel principal component analysis (kernel PCA) in the field of vision based robot localization. Using kernel PCA we can extract features from the visual scene of a mobile robot. The analysis is applied only to local features so as to guarantee better computational performance as well as translation invariance. Compared with the classical principal component analysis (PCA), kernel PCA results show superiority in localization and robustness in presence of noisy scenes. The key success of the kernel PCA is the use of fractional power polynomial kernels.

With the invention of microarray technology, researchers are capable of measuring the expression levels of ten thousands of genes in parallel at various time points of the biological process. During the investigation of gene regulatory networks and general cellular mechanisms, biologists are attempting to group genes based on the time-depending pattern of the obtained expression levels. In this paper, we propose a new memetic algorithm - a genetic algorithm combined with local search-based on a tree representation of the data - a minimum spanning tree minus; for clustering gene expression data. The combination of both concepts is shown to find near-optimal solutions quickly. Due to the minimum spanning tree representation of the data, our algorithm is capable of finding clusters of different shapes. We show that our approach is superior in solution quality compared to classical clustering methods.

In many engineering optimization problems, the number of fitness function evaluations is limited by time and cost. These problems pose a special challenge to the field of evolutionary computation, since existing evolutionary methods require a very large number of problem function evaluations. One popular way to address this challenge is the application of approximation models as a surrogate of the real fitness function. We propose a model assisted evolution strategy, which uses a Gaussian process approximation model to preselect the most promising solutions. To refine the preselection process we determine the likelihood of each individual to improve the overall best found solution. Due to this, the new algorithm has a much better convergence behavior and achieves better results than standard evolutionary optimization approaches with less fitness evaluations. Numerical results from extensive simulations on several high dimensional test functions including multimodal functions are presented.

The purpose of this article is to provide rescue robots with an adaptive behaviour during searching for victims in disasters such as fire, earthquake, flood, wars etc. This experimental research work took place in previously unknown dynamic indoor terrains. The main phases of this framework are; 1) modelling of robot behaviours/dynamics in collapsed environments, 2) designing an adaptive controller, which regulates robot longitudinal velocity and heading (collision avoidance) based on the obstacles distribution histogram, 3) prediction of robot behaviours in another unknown terrain. Two approaches have been used to design the adaptive controller: the first one is the stochastic control theory, based on Kalman filter algorithms. The second approach relies on fuzzy inference systems (FIS). Throughout this work, robot dynamics have been modelled using the auto regressive exogenous (ARX) scheme, while ARX model parameters have been identified using recursive least squares (RLS). This contribution presents a description and some discussion of the discrete Kalman filter, modelling techniques, and some discussion of robot behaviour analysis. Furthermore, the design of adaptive controllers using FIS-based techniques versus stochastic control systems bas been demonstrated.

This paper describes the assembly of a mobile odour sensing system and investigates its practical operation in an indoor environment without a constant airflow. Lacking a constant airflow leads to a problem which cannot be neglected in real world applications. The response of the metal oxide gas sensors used is dominated by air turbulence rather than concentration differences. We show that this problem can be overcome by driving the robot with a constant speed, thus adding an extra constant airflow relative to the gas sensors location. If the robot's speed is not too low the system described proved to be well suited to detect even weak odour sources. Since driving with constant speed is an indispensable condition to perform the basic tasks of a mobile odour sensing system, a new localization strategy is proposed, which takes this into account.

We present a system which is able to visually detect human faces, to track them by controlling a robot-head and to pursue a detected person by means of driving movements. The detection is based on a multimodal approach combining color, motion, and contour information. By using a stereo algorithm the position of the person in the scene is determined. Both the path of the person going ahead and a local environment map built by means of range sensor data are used to perform the navigation task. Stationary and dynamic obstacles are avoided during the process of pursuit.

We here present a parallel implementation of artificial neural networks on the connection machine CM-5 and compare it with other parallel implementations on SIMD and MIMD architectures. This parallel implementation was developed with the goal of efficiently training large neural networks with huge training pattern sets for applications in molecular biology, in particular the prediction of coding regions in DNA sequences. The implementation uses training pattern parallelism and makes use of the parallel I/O facilities of the CM-5 and its efficient reduction operations available within the control network to achieve a high scalability. The parallel simulator obtains a maximum speed of 149.25 MCUPS for training feedforward networks with backpropagation on a 512 processor CM-5 system without using the CM-5 vector facility. The implementation poses no restriction on the type of network topology and works with different batch training algorithms like BP. Quickprop and Rprop.

No abstract available.