CoRL 2022 Papers

Proactive robot assistance enables a robot to anticipate and provide for a user's needs without being explicitly asked. We formulate proactive assistance as the problem of the robot anticipating temporal patterns of object movements associated with everyday user routines, and proactively assisting the user by placing objects to adapt the environment to their needs. We introduce a generative graph neural network to learn a unified spatio-temporal predictive model of object dynamics from temporal sequences of object arrangements. We additionally contribute the Household Object Movements from Everyday Routines (HOMER) dataset, which tracks household objects associated with human activities of daily living across 50+ days for five simulated households. Our model outperforms the leading baseline in predicting object movement, correctly predicting locations for 11.1% more objects and wrongly predicting locations for 11.5% fewer objects used by the human user.

Autonomous vehicle (AV) stacks are typically built in a modular fashion, with explicit components performing detection, tracking, prediction, planning, control, etc. While modularity improves reusability, interpretability, and generalizability, it also suffers from compounding errors, information bottlenecks, and integration challenges. To overcome these challenges, a prominent approach is to convert the AV stack into an end-to-end neural network and train it with data. While such approaches have achieved impressive results, they typically lack interpretability and reusability, and they eschew principled analytical components, such as planning and control, in favor of deep neural networks. To enable the joint optimization of AV stacks while retaining modularity, we present DiffStack, a differentiable and modular stack for prediction, planning, and control. Crucially, our model-based planning and control algorithms leverage recent advancements in differentiable optimization to produce gradients, enabling optimization of upstream components, such as prediction, via backpropagation through planning and control. Our results on the nuScenes dataset indicate that end-to-end training with DiffStack yields substantial improvements in open-loop and closed-loop planning metrics by, e.g., learning to make fewer prediction errors that would affect planning. Beyond these immediate benefits, DiffStack opens up new opportunities for fully data-driven yet modular and interpretable AV architectures.

Deep motion forecasting models have achieved great success when trained on a massive amount of data. Yet, they often perform poorly when training data is limited. To address this challenge, we propose a transfer learning approach for efficiently adapting pre-trained forecasting models to new domains, such as unseen agent types and scene contexts. Unlike the conventional fine-tuning approach that updates the whole encoder, our main idea is to reduce the amount of tunable parameters that can precisely account for the target domain-specific motion style. To this end, we introduce two components that exploit our prior knowledge of motion style shifts: (i) a low-rank motion style adapter that projects and adjusts the style features at a low-dimensional bottleneck; and (ii) a modular adapter strategy that disentangles the features of scene context and motion history to facilitate a fine-grained choice of adaptation layers. Through extensive experimentation, we show that our proposed adapter design, coined MoSA, outperforms prior methods on several forecasting benchmarks.

We propose a sim-to-real framework for dexterous manipulation which can generalize to new objects of the same category in the real world. The key of our framework is to train the manipulation policy with point cloud inputs and dexterous hands. We propose two new techniques to enable joint learning on multiple objects and sim-to-real generalization: (i) using imagined hand point clouds as augmented inputs; and (ii) designing novel contact-based rewards. We empirically evaluate our method using an Allegro Hand to grasp novel objects in both simulation and real world. To the best of our knowledge, this is the first policy learning-based framework that achieves such generalization results with dexterous hands. Our project page is available at https://yzqin.github.io/dexpoint.

We study how visual representations pre-trained on diverse human video data can enable data-efficient learning of downstream robotic manipulation tasks. Concretely, we pre-train a visual representation using the Ego4D human video dataset using a combination of time-contrastive learning, video-language alignment, and an L1 penalty to encourage sparse and compact representations. The resulting representation, R3M, can be used as a frozen perception module for downstream policy learning. Across a suite of 12 simulated robot manipulation tasks, we find that R3M improves task success by over 20% compared to training from scratch and by over 10% compared to state-of-the-art visual representations like CLIP and MoCo. Furthermore, R3M enables a Franka Emika Panda arm to learn a range of manipulation tasks in a real, cluttered apartment given just 20 demonstrations.

Problems which require both long-horizon planning and continuous control capabilities pose significant challenges to existing reinforcement learning agents. In this paper we introduce a novel hierarchical reinforcement learning agent which links temporally extended skills for continuous control with a forward model in a symbolic discrete abstraction of the environment's state for planning. We term our agent SEADS for Symbolic Effect-Aware Diverse Skills. We formulate an objective and corresponding algorithm which leads to unsupervised learning of a diverse set of skills through intrinsic motivation given a known state abstraction. The skills are jointly learned with the symbolic forward model which captures the effect of skill execution in the state abstraction. After training, we can leverage the skills as symbolic actions using the forward model for long-horizon planning and subsequently execute the plan using the learned continuous-action control skills. The proposed algorithm learns skills and forward models that can be used to solve complex tasks which require both continuous control and long-horizon planning capabilities with high success rate. It compares favorably with other flat and hierarchical reinforcement learning baseline agents and is successfully demonstrated with a real robot.

Humans use all of their senses to accomplish different tasks in everyday activities. In contrast, existing work on robotic manipulation mostly relies on one, or occasionally two modalities, such as vision and touch. In this work, we systematically study how visual, auditory, and tactile perception can jointly help robots to solve complex manipulation tasks. We build a robot system that can see with a camera, hear with a contact microphone, and feel with a vision-based tactile sensor, with all three sensory modalities fused with a self-attention model. Results on two challenging tasks, dense packing and pouring, demonstrate the necessity and power of multisensory perception for robotic manipulation: vision displays the global status of the robot but can often suffer from occlusion, audio provides immediate feedback of key moments that are even invisible, and touch offers precise local geometry for decision making. Leveraging all three modalities, our robotic system significantly outperforms prior methods.

Physical interactions can often help reveal information that is not readily apparent. For example, we may tug at a table leg to evaluate whether it is built well, or turn a water bottle upside down to check that it is watertight. We propose to train robots to acquire such interactive behaviors automatically, for the purpose of evaluating the result of an attempted robotic skill execution. These evaluations in turn serve as "interactive reward functions" (IRFs) for training reinforcement learning policies to perform the target skill, such as screwing the table leg tightly. In addition, even after task policies are fully trained, IRFs can serve as verification mechanisms that improve online task execution. For any given task, our IRFs can be conveniently trained using only examples of successful outcomes, and no further specification is needed to train the task policy thereafter. In our evaluations on door locking and weighted block stacking in simulation, and screw tightening on a real robot, IRFs enable large performance improvements, even outperforming baselines with access to demonstrations or carefully engineered rewards.

Cloth in the real world is often crumpled, self-occluded, or folded in on itself such that key regions, such as corners, are not directly graspable, making manipulation difficult. We propose a system that leverages visual and tactile perception to unfold the cloth via grasping and sliding on edges. Doing so, the robot is able to grasp two adjacent corners, enabling subsequent manipulation tasks like folding or hanging. We develop tactile perception networks that classify whether an edge is grasped and estimate the pose of the edge. We use the edge classification network to supervise a visuotactile edge grasp affordance network that can grasp edges with a 90% success rate. Once an edge is grasped, we demonstrate that the robot can slide along the cloth to the adjacent corner using tactile pose estimation/control in real time.

Recent years have seen a rise in techniques for modeling discontinuous dynamics, such as rigid contact or switching motion modes, using deep learning. A common claim is that deep networks are incapable of accurately modeling rigid-body dynamics without explicit modules for handling contacts, due to the continuous nature of how deep networks are parameterized. Here we investigate this claim with experiments on established real and simulated datasets and show that general-purpose graph network simulators, with no contact-specific assumptions, can learn and predict contact discontinuities. Furthermore, contact dynamics learned by graph network simulators capture real-world cube tossing trajectories more accurately than highly engineered robotics simulators, even when provided with only 8 -- 16 trajectories. Overall, this suggests that rigid-body dynamics do not pose a fundamental challenge for deep networks with the appropriate general architecture and parameterization. Instead, our work opens new directions for considering when deep learning-based models might be preferable to traditional simulation environments for accurately modeling real-world contact dynamics.

As robots become increasingly capable of manipulation and long-term autonomy, long-horizon task and motion planning problems are becoming increasingly important. A key challenge in such problems is that early actions in the plan may make future actions infeasible. When reaching a dead-end in the search, most existing planners use backtracking, which exhaustively reevaluates motion-level actions, often resulting in inefficient planning, especially when the search depth is large. In this paper, we propose to learn backjumping heuristics which identify the culprit action directly using supervised learning models to guide the task-level search. Based on evaluations of two different tasks, we find that our method significantly improves planning efficiency compared to backtracking and also generalizes to problems with novel numbers of objects.

Neural fields can be trained from scratch to represent the shape and appearance of 3D scenes efficiently. It has also been shown that they can densely map correlated properties such as semantics, via sparse interactions from a human labeller. In this work, we show that a robot can densely annotate a scene with arbitrary discrete or continuous physical properties via its own fully-autonomous experimental interactions, as it simultaneously scans and maps it with an RGB-D camera. A variety of scene interactions are possible, including poking with force sensing to determine rigidity, measuring local material type with single-pixel spectroscopy or predicting force distributions by pushing. Sparse experimental interactions are guided by entropy to enable high efficiency, with tabletop scene properties densely mapped from scratch in a few minutes from a few tens of interactions.

Large-scale deployment of autonomous vehicles has been continually delayed due to safety concerns. On the one hand, comprehensive scene understanding is indispensable, a lack of which would result in vulnerability to rare but complex traffic situations, such as the sudden emergence of unknown objects. However, reasoning from a global context requires access to sensors of multiple types and adequate fusion of multi-modal sensor signals, which is difficult to achieve. On the other hand, the lack of interpretability in learning models also hampers the safety with unverifiable failure causes. In this paper, we propose a safety-enhanced autonomous driving framework, named Interpretable Sensor Fusion Transformer (InterFuser), to fully process and fuse information from multi-modal multi-view sensors for achieving comprehensive scene understanding and adversarial event detection. Besides, intermediate interpretable features are generated from our framework, which provide more semantics and are exploited to better constrain actions to be within the safe sets. We conducted extensive experiments on CARLA benchmarks, where our model outperforms prior methods, ranking the first on the public CARLA Leaderboard.

Robustly classifying ground infrastructure such as roads and street crossings is an essential task for mobile robots operating alongside pedestrians. While many semantic segmentation datasets are available for autonomous vehicles, models trained on such datasets exhibit a large domain gap when deployed on robots operating in pedestrian spaces. Manually annotating images recorded from pedestrian viewpoints is both expensive and time-consuming. To overcome this challenge, we propose \textit{TrackletMapper}, a framework for annotating ground surface types such as sidewalks, roads, and street crossings from object tracklets without requiring human-annotated data. To this end, we project the robot ego-trajectory and the paths of other traffic participants into the ego-view camera images, creating sparse semantic annotations for multiple types of ground surfaces from which a ground segmentation model can be trained. We further show that the model can be self-distilled for additional performance benefits by aggregating a ground surface map and projecting it into the camera images, creating a denser set of training annotations compared to the sparse tracklet annotations. We qualitatively and quantitatively attest our findings on a novel large-scale dataset for mobile robots operating in pedestrian areas. Code and dataset will be made available upon acceptance of the manuscript.

We present a method to learn compositional multi-object dynamics models from image observations based on implicit object encoders, Neural Radiance Fields (NeRFs), and graph neural networks. NeRFs have become a popular choice for representing scenes due to their strong 3D prior. However, most NeRF approaches are trained on a single scene, representing the whole scene with a global model, making generalization to novel scenes, containing different numbers of objects, challenging. Instead, we present a compositional, object-centric auto-encoder framework that maps multiple views of the scene to a set of latent vectors representing each object separately. The latent vectors parameterize individual NeRFs from which the scene can be reconstructed. Based on those latent vectors, we train a graph neural network dynamics model in the latent space to achieve compositionality for dynamics prediction. A key feature of our approach is that the latent vectors are forced to encode 3D information through the NeRF decoder, which enables us to incorporate structural priors in learning the dynamics models, making long-term predictions more stable compared to several baselines. Simulated and real world experiments show that our method can model and learn the dynamics of compositional scenes including rigid and deformable objects. Video: https://dannydriess.github.io/compnerfdyn/

To build general robotic agents that can operate in many environments, it is often imperative for the robot to collect experience in the real world. However, this is often not feasible due to safety, time and hardware restrictions. We thus propose leveraging the next best thing as real world experience: internet videos of humans using their hands. Visual priors, such as visual features, are often learned from videos, but we believe that more information from videos can be utilized as a stronger prior. We build a learning algorithm, Videodex, that leverages visual, action and physical priors from human video datasets to guide robot behavior. These action and physical priors in the neural network dictate the typical human behavior for a particular robot task. We test our approach on a robot arm and dexterous hand based system and show strong results on many different manipulation tasks, outperforming various state-of-the-art methods. For videos and supplemental material visit our website at https://video-dex.github.io.

A robotic feeding system must be able to acquire a variety of foods. Prior bite acquisition works consider single-arm spoon scooping or fork skewering, which do not generalize to foods with complex geometries and deformabilities. For example, when acquiring a group of peas, skewering could smoosh the peas while scooping without a barrier could result in chasing the peas on the plate. In order to acquire foods with such diverse properties, we propose stabilizing food items during scooping using a second arm, for example, by pushing peas against the spoon with a flat surface to prevent dispersion. The addition of this second stabilizing arm can lead to a new set of challenges. Critically, these strategies should stabilize the food scene without interfering with the acquisition motion, which is especially difficult for easily breakable high-risk food items, such as tofu. These high-risk foods can break between the pusher and spoon during scooping, which can lead to food waste falling onto the plate or out of the workspace. We propose a general bimanual scooping primitive and an adaptive stabilization strategy that enables successful acquisition of a diverse set of food geometries and physical properties. Our approach, CARBS: Coordinated Acquisition with Reactive Bimanual Scooping, learns to stabilize without impeding task progress by identifying high-risk foods and robustly scooping them using closed-loop visual feedback. We find that CARBS is able to generalize across food shape, size, and deformability and is additionally able to manipulate multiple food items simultaneously. CARBS achieves 87.0% success on scooping rigid foods, which is 25.8% more successful than a single-arm baseline, and reduces food breakage by 16.2% compared to an analytical baseline. Videos can be found on our website at https://sites.google.com/view/bimanualscoop-corl22/home.

Reinforcement learning in partially observable domains is challenging due to the lack of observable state information. Thankfully, learning offline in a simulator with such state information is often possible. In particular, we propose a method for partially observable reinforcement learning that uses a fully observable policy (which we call a \emph{state expert}) during training to improve performance. Based on Soft Actor-Critic (SAC), our agent balances performing actions similar to the state expert and getting high returns under partial observability. Our approach can leverage the fully-observable policy for exploration and parts of the domain that are fully observable while still being able to learn under partial observability. On six robotics domains, our method outperforms pure imitation, pure reinforcement learning, the sequential or parallel combination of both types, and a recent state-of-the-art method in the same setting. A successful policy transfer to a physical robot in a manipulation task from pixels shows our approach's practicality in learning interesting policies under partial observability.

This paper presents a novel framework for planning in unknown and occluded urban spaces. We specifically focus on turns and intersections where occlusions significantly impact navigability. Our approach uses an inpainting model to fill in a sparse, occluded, semantic lidar point cloud and plans dynamically feasible paths for a vehicle to traverse through the open and inpainted spaces. We demonstrate our approach using a car’s lidar data with real-time occlusions, and show that by inpainting occluded areas, we can plan longer paths, with more turn options compared to without inpainting; in addition, our approach more closely follows paths derived from a planner with no occlusions (called the ground truth) compared to other state of the art approaches.