We present Neural Memory Object (NeMO), a novel object-centric representation that can be used to detect, segment and estimate the 6DoF pose of objects unseen during training using RGB images. Our method consists of an encoder that requires only a few RGB template views depicting an object to generate a sparse object-like point cloud using a learned UDF containing semantic and geometric information. Next, a decoder takes the object encoding together with a query image to generate a variety of dense predictions. Through extensive experiments, we show that our method can be used for few-shot object perception without requiring any camera-specific parameters or retraining on target data. Our proposed concept of outsourcing object information in a NeMO and using a single network for multiple perception tasks enhances interaction with novel objects, improving scalability and efficiency by enabling quick object onboarding without retraining or extensive pre-processing. We report competitive and state-of-the-art results on various datasets and perception tasks of the BOP benchmark, demonstrating the versatility of our approach. https://github.com/DLR-RM/nemo

The interaction of robots with their environment requires robust object-centric perception capabilities, typically achieved using learning-based methods trained on synthetic data. However, real-world deployment demands evaluating these capabilities in relevant environments, often involving extensive manual annotation for a quantitative analysis. Additionally, standardized evaluations for robotic tasks, such as grasping, need reproducible object scene configurations and performance benchmarks. We propose a solution to both problems by temporarily employing 3D-printed components, socalled fixtures, which can be designed for any rigid object. Once the scene is set up and object poses are extracted, the fixtures are removed, leaving the natural scene without any artificial distractions. The presented approach is seemingly applicable for pre-determined configurations of multiple objects, which enables precise re-building of scenes with consistent object-toobject relations. Our suggested annotation procedure achieves strong pose accuracy solely on RGB images without any manual involvement. We evaluate and show the usability of the proposed fixtures for automated real-world data annotation to fine-tune a detector and for benchmarking object pose estimation algorithms for robotic grasping. Code and fixture meshes for 3D printing are available at https://github.com/DLRRM/fixture_generation.

We present PK-ROKED, a learning-based pipeline for probabilistic robot pose estimation relative to a camera, addressing inaccuracies in forward kinematics, particularly in systems with elastic and lightweight modules. Our approach integrates a probabilistic 2D keypoint detection mechanism that leverages prior knowledge derived from the robot's imprecise kinematics. We further improve the detection accuracy and geometric understanding by incorporating segmentation of the robot arm. The method computes reliable uncertainty estimates, enabling a robust 2D-6D fusion for precise robot arm pose estimation from a single detected keypoint. PK-ROKED requires only synthetic training data, effectively exploits imperfect kinematics as valuable prior knowledge, and introduces a novel fusion framework for enhanced robot pose estimation. We validate our method on the Panda-Orb dataset, demonstrating competitive performance against state-of-the-art approaches. Additionally, we evaluate on two other robotic systems in real-world scenarios and show its practicality by using the predictions to initialize a tracking algorithm. Code and pre-trained models are available.

We propose a novel pipeline for unknown object grasping in shared robotic autonomy scenarios. State-of-the-art methods for fully autonomous scenarios are typically learning-based approaches optimised for a specific end-effector, that generate grasp poses directly from sensor input. In the domain of assistive robotics, we seek instead to utilise the user’s cognitive abilities for enhanced satisfaction, grasping performance, and alignment with their high level task-specific goals. Given a pair of stereo images, we perform unknown object instance segmentation and generate a 3D reconstruction of the object of interest. In shared control, the user then guides the robot end-effector across a virtual hemisphere centered around the object to their desired approach direction. A physics-based grasp planner finds the most stable local grasp on the reconstruction, and finally the user is guided by shared control to this grasp. In experiments on the DLR EDAN platform, we report a grasp success rate of 87% for 10 unknown objects, and demonstrate the method’s capability to grasp objects in structured clutter and from shelves.

Automatic Robotic Assembly Sequence Planning (RASP) can significantly improve productivity and resilience in modern manufacturing along with the growing need for greater product customization. One of the main challenges in realizing such automation resides in efficiently finding solutions from a growing number of potential sequences for increasingly complex assemblies. Besides, costly feasibility checks are always required for the robotic system. To address this, we propose a holistic graphical approach including a graph representation called Assembly Graph for product assemblies and a policy architecture, Graph Assembly Processing Network, dubbed GRACE for assembly sequence generation. With GRACE, we are able to extract meaningful information from the graph input and predict assembly sequences in a step-by-step manner. In experiments, we show that our approach can predict feasible assembly sequences across product variants of aluminum profiles based on data collected in simulation of a dual-armed robotic system. We further demonstrate that our method is capable of detecting infeasible assemblies, substantially alleviating the undesirable impacts from false predictions, and hence facilitating real-world deployment soon. Code and training data are available at https://github.com/DLR-RM/GRACE.

We present a novel technique to estimate the 6D pose of objects from single images where the 3D geometry of the object is only given approximately and not as a precise 3D model. To achieve this, we employ a dense 2D-to-3D correspondence predictor that regresses 3D model coordinates for every pixel. In addition to the 3D coordinates, our model also estimates the pixel-wise coordinate error to discard correspondences that are likely wrong. This allows us to generate multiple 6D pose hypotheses of the object, which we then refine iteratively using a highly efficient region-based approach. We also introduce a novel pixel-wise posterior formulation by which we can estimate the probability for each hypothesis and select the most likely one. As we show in experiments, our approach is capable of dealing with extreme visual conditions including overexposure, high contrast, or low signal-to-noise ratio. This makes it a powerful technique for the particularly challenging task of estimating the pose of tumbling satellites for in-orbit robotic applications. Our method achieves state-of-the-art performance on the SPEED+ dataset and has won the SPEC2021 post-mortem competition. Code, trained models, and the used satellite model will be made publicly available.

While learning from synthetic training data has recently gained an increased attention, in real-world robotic applications, there are still performance deficiencies due to the so-called Sim-to-Real gap. In practice, this gap is hard to resolve with only synthetic data. Therefore, we focus on an efficient acquisition of real data within a Sim-to-Real learning pipeline. Concretely, we employ deep Bayesian active learning to minimize manual annotation efforts and devise an autonomous learning paradigm to select the data that is considered useful for the human expert to annotate. To achieve this, a Bayesian Neural Network (BNN) object detector providing reliable un-certainty estimates is adapted to infer the informativeness of the unlabeled data. Furthermore, to cope with misalignments of the label distribution in uncertainty-based sampling, we develop an effective randomized sampling strategy that performs favorably compared to other complex alternatives. In our experiments on object classification and detection, we show benefits of our approach and provide evidence that labeling efforts can be reduced significantly. Finally, we demonstrate the practical effectiveness of this idea in a grasping task on an assistive robot.

We present a novel framework for self-supervised grasped object segmentation with a robotic manipulator. Our method successively learns an agnostic foreground segmentation followed by a distinction between manipulator and object solely by observing the motion between consecutive RGB frames. In contrast to previous approaches, we propose a single, end-to-end trainable architecture which jointly incorporates motion cues and semantic knowledge. Furthermore, while the motion of the manipulator and the object are substantial cues for our algorithm, we present means to robustly deal with distraction objects moving in the background, as well as with completely static scenes. Our method neither depends on any visual registration of a kinematic robot or 3D object models, nor on precise hand-eye calibration or any additional sensor data. By extensive experimental evaluation we demonstrate the superiority of our framework and provide detailed insights on its capability of dealing with the aforementioned extreme cases of motion. We also show that training a semantic segmentation network with the automatically labeled data achieves results on par with manually annotated training data. Code and pretrained model are available at https://github.com/DLR-RM/DistinctNet.

Although instance-aware perception is a key prerequisite for many autonomous robotic applications, most of the methods only partially solve the problem by focusing solely on known object categories. However, for robots interacting in dynamic and cluttered environments, this is not realistic and severely limits the range of potential applications. Therefore, we propose a novel object instance segmentation approach that does not require any semantic or geometric information of the objects beforehand. In contrast to existing works, we do not explicitly use depth data as input, but rely on the insight that slight viewpoint changes, which for example are provided by stereo image pairs, are often sufficient to determine object boundaries and thus to segment objects. Focusing on the versatility of stereo sensors, we employ a transformer-based architecture that maps directly from the pair of input images to the object instances. This has the major advantage that instead of a noisy, and potentially incomplete depth map as an input, on which the segmentation is computed, we use the original image pair to infer the object instances and a dense depth map. In experiments in several different application domains, we show that our Instance Stereo Transformer (INSTR) algorithm outperforms current state-of-the-art methods that are based on depth maps. Training code and pretrained models are available at https://github.com/DLR-RM/instr.

Suitable datasets are an integral part of robotics research, especially for training neural networks in robot perception. However, in many domains, suitable real-world data are scarce and cannot be easily obtained. This problem is especially prevalent for unstructured outdoor environments, in particular, planetary ones. Recent advances in photorealistic simulations help researchers to simulate close-to-real data in many domains. Yet, there exists no high-quality synthetic data for planetary exploration tasks. Also, existing simulators lack the fidelity required for generating planetary data, which is inherently less structured than human environments. Synthetic planetary data requires careful modeling and annotation of many different terrain aspect and details, such as textures and distributions of rocks, to become a valuable test-bed for robotics. To fill this gap, we present a novel simulator specifically designed for the needs of planetary robotics visual tasks, but also applicable for other outdoor environments. Our simulator is capable of generating large varieties of (planetary) outdoor scenes with rich generation of meta data, such as multilevel semantic and instance annotations. To demonstrate the wide applicability of this new simulator, we evaluate its performance on typical robotics applications, i.e. semantic segmentation, instance segmentation, and visual SLAM. Our simulator is accessible under https://github.com/DLR-RM/oaisys.

We introduce a scalable approach for object pose estimation trained on simulated RGB views of multiple 3D models together. We learn an encoding of object views that does not only describe an implicit orientation of all objects seen during training, but can also relate views of untrained objects. Our single-encoder-multi-decoder network is trained using a technique we denote ”multi-path learning”: While the encoder is shared by all objects, each decoder only reconstructs views of a single object. Consequently, views of different instances do not have to be separated in the latent space and can share common features. The resulting encoder generalizes well from synthetic to real data and across various instances, categories, model types and datasets. We systematically investigate the learned encodings, their generalization, and iterative refinement strategies on the ModelNet40 and T-LESS dataset. Despite training jointly on multiple objects, our 6D Object Detection pipeline achieves state-of-the-art results on T-LESS at much lower runtimes than competing approaches.

One of the main challenges in sampling-based motion planners is to find an efficient sampling strategy. While methods such as Rapidly-exploring Random Tree (RRT) have shown to be more reliable in complex environments than optimization-based methods, they often require longer planning times, which reduces their usability for real-time applications. Recently, biased sampling methods have shown to remedy this issue. For example Gaussian Mixture Models (GMMs) have been used to sample more efficiently in feasible regions of the configuration space. Once the GMM is learned, however, this approach does not adapt its biases to individual planning scene during inference. Hence, we propose in this work a more efficient sampling strategy to further bias the GMM based on visual input upon query. We employ an autoencoder trained entirely in simulation to extract features from depth images and use the latent representation to adjust the weights of each mixture components in the GMM. We show empirically that this improves the sampling efficiency of an RRT motion planner in both real and simulated scenes.

We present an approach to automatically assign semantic labels to rooms reconstructed from 3D RGB maps of apartments. Evidence for the room types is generated using state-of-the-art deep-learning techniques for scene classification and object detection based on automatically generated virtual RGB views, as well as from a geometric analysis of the map's 3D structure. The evidence is merged in a conditional random field, using statistics mined from different datasets of indoor environments. We evaluate our approach qualitatively and quantitatively and compare it to related methods.

We present a novel method for classifying 3D objects that is particularly tailored for the requirements in robotic applications. The major challenges here are the comparably small amount of available training data and the fact that often data is perceived in streams and not in fixed-size pools. Traditional state-of-the-art learning methods, however, require a large amount of training data, and their online learning capabilities are usually limited. Therefore, we propose a modality-specific selection of convolutional neural networks (CNN), pre-trained or fine-tuned, in combination with a classifier that is designed particularly for online learning from data streams, namely the Mondrian Forest (MF). We show that this \ncombination of trained features obtained from a CNN can be improved further if a feature selection algorithm is applied. In our experiments, we use the resulting features both with a MF and a linear Support Vector Machine (SVM). With SVM we beat the state of the art on an RGB-D dataset, while with MF a strong result for active learning is achieved.

In this work, a new ensemble method for the task of category recognition in different environments is presented. The focus is on service robotic perception in an open environment, where the robot’s task is to recognize previously unseen objects of predefined categories, based on training on a public dataset. We propose an ensemble learning approach to be able to flexibly combine complementary sources of information (different state-of-the-art descriptors computed on color and depth images), based on a Markov Random Field (MRF). By exploiting its specific characteristics, the MRF ensemble method can also be executed as a Dynamic Classifier Selection (DCS) system. In the experiments, the committee- and topology-dependent performance boost of our ensemble is shown. Despite reduced computational costs and using less information, our strategy performs on the same level as common ensemble approaches. Finally, the impact of large differences between datasets is analyzed.