We present a flexible method for fusing information from optical and range sensors based on an accelerated high-dimensional filtering approach. Our system takes as input a sequence of monocular camera images as well as a stream of sparse range measurements as obtained from a laser or other sensor system. In contrast with existing approaches, we do not assume that the depth and color data streams have the same data rates or that the observed scene is fully static. Our method produces a dense, high-resolution depth map of the scene, automatically generating confidence values for every interpolated depth point. We describe how to integrate priors on object motion and appearance and how to achieve an efficient implementation using parallel processing hardware such as GPUs.

Markerless tracking of human pose is a hard yet relevant problem. In this paper, we derive an efficient filtering algorithm for tracking human pose using a stream of monocular depth images. The key idea is to combine an accurate generative model - which is achievable in this setting using programmable graphics hardware - with a discriminative model that provides data-driven evidence about body part locations. In each filter iteration, we apply a form of local model-based search that exploits the nature of the kinematic chain. As fast movements and occlusion can disrupt the local search, we utilize a set of discriminatively trained patch classifiers to detect body parts. We describe a novel algorithm for propagating this noisy evidence about body part locations up the kinematic chain using the unscented transform. The resulting distribution of body configurations allows us to reinitialize the model-based search. We provide extensive experimental results on 28 real-world sequences using automatic ground-truth annotations from a commercial motion capture system.

We consider the task of accurately controlling a complex system, such as autonomously sliding a car sideways into a parking spot. Although certain regions of this domain are extremely hard to model (i.e., the dynamics of the car while skidding), we observe that in practice such systems are often remarkably deterministic over short periods of time, even in difficult-to-model regions. Motivated by this intuition, we develop a probabilistic method for combining closed-loop control in the well-modeled regions and open-loop control in the difficult-to-model regions. In particular, we show that by combining 1) an inaccurate model of the system and 2) a demonstration of the desired behavior, our approach can accurately and robustly control highly challenging systems, without the need to explicitly model the dynamics in the most complex regions and without the need to hand-tune the switching control law. We apply our approach to the task of autonomous sideways sliding into a parking spot, and show that we can repeatedly and accurately control the system, placing the car within about 2 feet of the desired location; to the best of our knowledge, this represents the state of the art in terms of accurately controlling a vehicle in such a maneuver.

We deal with the problem of detecting and identifying body parts in depth images at video frame rates. Our solution involves a novel interest point detector for mesh and range data that is particularly well suited for analyzing human shape. The interest points, which are based on identifying geodesic extrema on the surface mesh, coincide with salient points of the body, which can be classified as, e.g., hand, foot or head using local shape descriptors. Our approach also provides a natural way of estimating a 3D orientation vector for a given interest point. This can be used to normalize the local shape descriptors to simplify the classification problem as well as to directly estimate the orientation of body parts in space. Experiments involving ground truth labels acquired via an active motion capture system show that our interest points in conjunction with a boosted patch classifier are significantly better in detecting body parts in depth images than state-of-the-art sliding-window based detectors.

The received signal strength (RSS) of radiofrequency signals emitted from beacons placed at known locations in an environment, can be used by a local positioning system (LPS) to estimate the location of a person or a mobile object. In indoor environments, interference, multipath propagation of RF signals, and the presence of obstacles and people, lead to a complex spatial distribution of the RSS, which is inaccurately described by simple parametric models. In this work, we present a Bayesian method for an indoor RFID location system which uses an observation model based in Gaussian processes (GPs) nonparametric regression to represent the environment-specific RSS distributions for the individual RFID tags. The experimental results in an indoor environment demonstrate the effectiveness of GPs in order to increase positioning accuracy.

In recent years, there has been an increasing interest within the robotics community in investigating whether Radio Frequency Identification (RFID) technology can be utilized to solve localization and mapping problems in the context of mobile robots. We present a novel sensor model which can be utilized for localizing RFID tags and for tracking a mobile agent moving through an RFID-equipped environment. The proposed probabilistic sensor model characterizes the received signal strength indication (RSSI) information as well as the tag detection events to achieve a higher modeling accuracy compared to state-of-the-art models which deal with one of these aspects only. We furthermore propose a method that is able to bootstrap such a sensor model in a fully unsupervised fashion. Real-world experiments demonstrate the effectiveness of our approach also in comparison to existing techniques.

To act intelligently in dynamic environments, a system must understand the current situation it is involved in at any given time. This requires dealing with temporal context, handling multiple and ambiguous interpretations, and accounting for various sources of uncertainty. In this paper we propose a probabilistic approach to modeling and recognizing situations. We define a situation as a distribution over sequences of states that have some meaningful interpretation. Each situation is characterized by an individual hidden Markov model that describes the corresponding distribution. In particular, we consider typical traffic scenarios and describe how our framework can be used to model and track different situations while they are evolving. The approach was evaluated experimentally in vehicular traffic scenarios using real and simulated data. The results show that our system is able to recognize and track multiple situation instances in parallel and make sensible decisions between competing hypotheses. Additionally, we show that our models can be used for predicting the position of the tracked vehicles.

One of the key tasks during the realization of probabilistic approaches to localization is the design of a proper sensor model, that calculates the likelihood of a measurement given the current pose of the vehicle and the map of the environment. In the past, range sensors have become popular for mobile robot localization since they directly measure distance. However, in situations in which the robot operates close to edges of obstacles or in highly cluttered environments, small changes in the pose of the robot can lead to large variations in the acquired range scans. If the sensor model used does not appropriately characterize the resulting fluctuations, the performance of probabilistic approaches may substantially degrade. A common solution is to artificially smooth the likelihood function or to only integrate a small fraction of the measurements. In this paper we present a more fundamental and robust approach which uses mixtures of Gaussians to model the likelihood function for single range measurements. In practical experiments we compare our approach to previous methods and demonstrate that it yields a substantially increase in robustness.

Mobile robots rely on the ability to sense the geometry of their local environment in order to avoid obstacles or to explore the surroundings. For this task, dedicated proximity sensors such as laser range finders or sonars are typically employed. Cameras are a cheap and lightweight alternative to such sensors, but do not directly offer proximity information. In this paper, we present a novel approach to learning the relationship between range measurements and visual features extracted from a single monocular camera image. As the learning engine, we apply Gaussian processes, a non-parametric learning technique that not only yields the most likely range prediction corresponding to a certain visual input but also the predictive uncertainty. This information, in turn, can be utilized in an extended grid-based mapping scheme to more accurately update the map. In practical experiments carried out in different environments with a mobile robot equipped with an omnidirectional camera system, we demonstrate that our system is able to produce proximity estimates with an accuracy comparable to that of dedicated sensors such as sonars or infrared range finders.

In this paper, we present an approach allowing a robot to learn a generative model of its own physical body from scratch using self-perception with a single monocular camera. Our approach yields a compact Bayesian network for the robot's kinematic structure including the forward and inverse models relating action signals and body pose. We propose to simultaneously learn local action models for all pairs of perceivable body parts from data generated through random "motor babbling." From this repertoire of local models, we construct a Bayesian network for the full system using the pose prediction accuracy on a separate cross validation data set as the criterion for model selection. The resulting model can be used to predict the body pose when no perception is available and allows for gradient-based posture control. In experiments with real and simulated manipulator arms, we show that our system is able to quickly learn compact and accurate models and to robustly deal with noisy observations.

The problem of estimating the positions of landmarks using a mobile robot equipped with a camera has intensively been studied in the past. In this paper, we consider a variant of this problem in which the robot should estimate the locations of observed landmarks based on a sparse set of geo-referenced images for which no heading information is available. Sources for such kind of data are image portals such as Flickr or Google Image Search. We formulate the problem of estimating the landmark locations as an optimization problem and show that it is possible to accurately localize the landmarks in real world settings.

Whereas probabilistic approaches are a powerful tool for mobile robot localization, they heavily rely on the proper definition of the so-called observation model which defines the likelihood of an observation given the position and orientation of the robot and the map of the environment. Most of the sensor models for range sensors proposed in the past either consider the individual beam measurements independently or apply uni-modal models to represent the likelihood function. In this paper, we present an approach that learns place-dependent sensor models for entire range scans using Gaussian mixture models. To deal with the high dimensionality of the measurement space, we utilize principle component analysis for dimensionality reduction. In practical experiments carried out with data obtained from a real robot, we demonstrate that our model substantially outperforms existing and popular sensor models.

Legged robots require accurate models of their environment in order to plan and execute paths. We present a probabilistic technique based on Gaussian processes that allows terrain models to be learned and updated efficiently using sparse approximation techniques. The major benefit of our terrain model is its ability to predict elevations at unseen locations more reliably than alternative approaches, while it also yields estimates of the uncertainty in the prediction. In particular, our nonstationary Gaussian process model adapts its covariance to the situation at hand, allowing more accurate inference of terrain height at points that have not been observed directly. We show how a conventional motion planner can use the learned terrain model to plan a path to a goal location, using a terrain-specific cost model to accept or reject candidate footholds. In experiments with a real quadruped robot equipped with a laser range finder, we demonstrate the usefulness of our approach and discuss its benefits compared to simpler terrain models such as elevations grids.

This paper presents a novel Gaussian process (GP) approach to regression with input-dependent noise rates. We follow Goldberg et al.'s approach and model the noise variance using a second GP in addition to the GP governing the noise-free output value. In contrast to Goldberg et al., however, we do not use a Markov chain Monte Carlo method to approximate the posterior noise variance but a most likely noise approach. The resulting model is easy to implement and can directly be used in combination with various existing extensions of the standard GPs such as sparse approximations. Extensive experiments on both synthetic and real-world data, including a challenging perception problem in robotics, show the effectiveness of most likely heteroscedastic GP regression.

Range sensors are popular for localization since they directly measure the geometry of the local environment. Another distinct benefit is their typically high accuracy and spatial resolution. It is a well-known problem, however, that the high precision of these sensors leads to practical problems in probabilistic localization approaches such as Monte Carlo localization (MCL), because the likelihood function becomes extremely peaked if no means of regularization are applied. In practice, one therefore artificially smoothes the likelihood function or only integrates a small fraction of the measurements. In this paper we present a more fundamental and robust approach, that provides a smooth likelihood model for entire range scans. Additionally, it is location-dependent. In practical experiments we compare our approach to previous methods and demonstrate that it leads to a more robust localization.

In this paper, we present an approach that applies the reinforcement learning principle to the problem of learning height control policies for aerial blimps. In contrast to previous approaches, our method does not require sophisticated hand- tuned models, but rather learns the policy online, which makes the system easily adaptable to changing conditions. The blimp we apply our approach to is a small-scale vehicle equipped with an ultrasound sensor that measures its elevation relative to the ground. The major problem in the context of learning control policies lies in the high-dimensional state-action space that needs to be explored in order to identify the values of all state-action pairs. In this paper, we propose a solution to learning continuous control policies based on the Gaussian process model. In practical experiments carried out on a real robot we demonstrate that the system is able to learn a policy online within a few minutes only.

Navigation is one of the fundamental tasks for a mobile robot. The majority of path planning approaches has been designed to entirely solve the given problem from scratch given the current and goal configurations of the robot. Although these approaches yield highly efficient plans, the computed policies typically do not transfer to other, similar tasks. We propose to learn relational decision trees as abstract navigation strategies from example paths. Relational abstraction has several interesting and important properties. First, it allows a mobile robot to generalize navigation plans from specific examples provided by users or exploration. Second, the navigation policy learned in one environment can be transferred to unknown environments. In several experiments with real robots in a real environment and in simulated runs, we demonstrate the usefulness of our approach