CoRL 2023 Papers

Recent advances in multi-agent reinforcement learning (MARL) are enabling impressive coordination in heterogeneous multi-robot teams. However, existing approaches often overlook the challenge of generalizing learned policies to teams of new compositions, sizes, and robots. While such generalization might not be important in teams of virtual agents that can retrain policies on-demand, it is pivotal in multi-robot systems that are deployed in the real-world and must readily adapt to inevitable changes. As such, multi-robot policies must remain robust to team changes -- an ability we call adaptive teaming. In this work, we investigate if awareness and communication of robot capabilities can provide such generalization by conducting detailed experiments involving an established multi-robot test bed. We demonstrate that shared decentralized policies, that enable robots to be both aware of and communicate their capabilities, can achieve adaptive teaming by implicitly capturing the fundamental relationship between collective capabilities and effective coordination. Videos of trained policies can be viewed at https://sites.google.com/view/cap-comm .

We present CAJun, a novel hierarchical learning and control framework that enables legged robots to jump continuously with adaptive jumping distances. CAJun consists of a high-level centroidal policy and a low-level leg controller. In particular, we use reinforcement learning (RL) to train the centroidal policy, which specifies the gait timing, base velocity, and swing foot position for the leg controller. The leg controller optimizes motor commands for the swing and stance legs according to the gait timing to track the swing foot target and base velocity commands.% using optimal control. Additionally, we reformulate the stance leg optimizer in the leg controller to speed up policy training by an order of magnitude. Our system combines the versatility of learning with the robustness of optimal control. % By combining RL with optimal control methods, our system achieves the versatility of learning while enjoys the robustness from control methods.% making it easily transferable to real robots. We show that after 20 minutes of training on a single GPU, CAJun can achieve continuous, long jumps with adaptive distances on a Go1 robot with small sim-to-real gaps. Moreover, the robot can jump across gaps with a maximum width of 70cm, which is over 40% wider than existing methods.

Trajectory prediction modules are key enablers for safe and efficient planning of autonomous vehicles (AVs), particularly in highly interactive traffic scenarios. Recently, learning-based trajectory predictors have experienced considerable success in providing state-of-the-art performance due to their ability to learn multimodal behaviors of other agents from data. In this paper, we present an algorithm called multi-predictor fusion (MPF) that augments the performance of learning-based predictors by imbuing them with motion planners that are tasked with satisfying logic-based rules. MPF probabilistically combines learning- and rule-based predictors by mixing trajectories from both standalone predictors in accordance with a belief distribution that reflects the online performance of each predictor. In our results, we show that MPF outperforms the two standalone predictors on various metrics and delivers the most consistent performance.

We present a learning-based dynamics model for granular material manipulation. Drawing inspiration from computer graphics' Eulerian approach, our method adopts a fully convolutional neural network that operates on a density field-based representation of object piles, allowing it to exploit the spatial locality of inter-object interactions through the convolution operations. This approach greatly improves the learning and computation efficiency compared to existing latent or particle-based methods and sidesteps the need for state estimation, making it directly applicable to real-world settings. Furthermore, our differentiable action rendering module makes the model fully differentiable and can be directly integrated with a gradient-based algorithm for curvilinear trajectory optimization. We evaluate our model with a wide array of piles manipulation tasks both in simulation and real-world experiments and demonstrate that it significantly exceeds existing methods in both accuracy and computation efficiency. More details can be found at https://sites.google.com/view/nfd-corl23/

Reinforcement Learning (RL) is notoriously data-inefficient, which makes training on a real robot difficult. While model-based RL algorithms (world models) improve data-efficiency to some extent, they still require hours or days of interaction to learn skills. Recently, offline RL has been proposed as a framework for training RL policies on pre-existing datasets without any online interaction. However, constraining an algorithm to a fixed dataset induces a state-action distribution shift between training and inference, and limits its applicability to new tasks. In this work, we seek to get the best of both worlds: we consider the problem of pretraining a world model with offline data collected on a real robot, and then finetuning the model on online data collected by planning with the learned model. To mitigate extrapolation errors during online interaction, we propose to regularize the planner at test-time by balancing estimated returns and (epistemic) model uncertainty. We evaluate our method on a variety of visuo-motor control tasks in simulation and on a real robot, and find that our method enables few-shot finetuning to seen and unseen tasks even when offline data is limited. Videos are available at https://yunhaifeng.com/FOWM

Diligently gathered human demonstrations serve as the unsung heroes empowering the progression of robot learning. Today, demonstrations are collected by training people to use specialized controllers, which (tele-)operate robots to manipulate a small number of objects. By contrast, we introduce AR2-D2: a system for collecting demonstrations which (1) does not require people with specialized training, (2) does not require any real robots during data collection, and therefore, (3) enables manipulation of diverse objects with a real robot. AR2-D2 is a framework in the form of an iOS app that people can use to record a video of themselves manipulating any object while simultaneously capturing essential data modalities for training a real robot. We show that data collected via our system enables the training of behavior cloning agents in manipulating real objects. Our experiments further show that training with our AR data is as effective as training with real-world robot demonstrations. Moreover, our user study indicates that users find AR2-D2 intuitive to use and require no training in contrast to four other frequently employed methods for collecting robot demonstrations.

Grasping occluded objects in cluttered environments is an essential component in complex robotic manipulation tasks. In this paper, we introduce an AffordanCE-driven Next-Best-View planning policy (ACE-NBV) that tries to find a feasible grasp for target object via continuously observing scenes from new viewpoints. This policy is motivated by the observation that the grasp affordances of an occluded object can be better-measured under the view when the view-direction are the same as the grasp view. Specifically, our method leverages the paradigm of novel view imagery to predict the grasps affordances under previously unobserved view, and select next observation view based on the highest imagined grasp quality of the target object. The experimental results in simulation and on a real robot demonstrate the effectiveness of the proposed affordance-driven next-best-view planning policy. Project page: https://sszxc.net/ace-nbv/.

Motion planning for robot manipulation systems operating in complex environments remains a challenging problem. It requires the evaluation of both the collision distance and its derivative. Owing to its computational complexity, recent studies have attempted to utilize data-driven approaches to learn the collision distance. However, their performance degrades significantly for complicated high-dof systems, such as multi-arm robots. Additionally, the model must be retrained every time the environment undergoes even slight changes. In this paper, we propose PairwiseNet, a model that estimates the minimum distance between two geometric shapes and overcomes many of the limitations of current models. By dividing the problem of global collision distance learning into smaller pairwise sub-problems, PairwiseNet can be used to efficiently calculate the global collision distance. PairwiseNet can be deployed without further modifications or training for any system comprised of the same shape elements (as those in the training dataset). Experiments with multi-arm manipulation systems of various dof indicate that our model achieves significant performance improvements concerning several performance metrics, especially the false positive rate with the collision-free guaranteed threshold. Results further demonstrate that our single trained PairwiseNet model is applicable to all multi-arm systems used in the evaluation. The code is available at https://github.com/kjh6526/PairwiseNet.

Gradient-based methods enable efficient search capabilities in high dimensions. However, in order to apply them effectively in offline optimization paradigms such as offline Reinforcement Learning (RL) or Imitation Learning (IL), we require a more careful consideration of how uncertainty estimation interplays with first-order methods that attempt to minimize them. We study smoothed distance to data as an uncertainty metric, and claim that it has two beneficial properties: (i) it allows gradient-based methods that attempt to minimize uncertainty to drive iterates to data as smoothing is annealed, and (ii) it facilitates analysis of model bias with Lipschitz constants. As distance to data can be expensive to compute online, we consider settings where we need amortize this computation. Instead of learning the distance however, we propose to learn its gradients directly as an oracle for first-order optimizers. We show these gradients can be efficiently learned with score-matching techniques by leveraging the equivalence between distance to data and data likelihood. Using this insight, we propose Score-Guided Planning (SGP), a planning algorithm for offline RL that utilizes score-matching to enable first-order planning in high-dimensional problems, where zeroth-order methods were unable to scale, and ensembles were unable to overcome local minima. Website: https://sites.google.com/view/score-guided-planning/home

Long-horizon tasks, usually characterized by complex subtask dependencies, present a significant challenge in manipulation planning. Skill chaining is a practical approach to solving unseen tasks by combining learned skill priors. However, such methods are myopic if sequenced greedily and face scalability issues with search-based planning strategy. To address these challenges, we introduce Generative Skill Chaining (GSC), a probabilistic framework that learns skill-centric diffusion models and composes their learned distributions to generate long-horizon plans during inference. GSC samples from all skill models in parallel to efficiently solve unseen tasks while enforcing geometric constraints. We evaluate the method on various long-horizon tasks and demonstrate its capability in reasoning about action dependencies, constraint handling, and generalization, along with its ability to replan in the face of perturbations. We show results in simulation and on real robot to validate the efficiency and scalability of GSC, highlighting its potential for advancing long-horizon task planning. More details are available at: https://generative-skill-chaining.github.io/

We approach the fundamental problem of obstacle avoidance for robotic systems via the lens of online learning. In contrast to prior work that either assumes worst-case realizations of uncertainty in the environment or a stationary stochastic model of uncertainty, we propose a method that is efficient to implement and provably grants instance-optimality with respect to perturbations of trajectories generated from an open-loop planner (in the sense of minimizing worst-case regret). The resulting policy adapts online to realizations of uncertainty and provably compares well with the best obstacle avoidance policy in hindsight from a rich class of policies. The method is validated in simulation on a dynamical system environment and compared to baseline open-loop planning and robust Hamilton-Jacobi reachability techniques. Further, it is implemented on a hardware example where a quadruped robot traverses a dense obstacle field and encounters input disturbances due to time delays, model uncertainty, and dynamics nonlinearities.

Reusing large datasets is crucial to scale vision-based robotic manipulators to everyday scenarios due to the high cost of collecting robotic datasets. However, robotic platforms possess varying control schemes, camera viewpoints, kinematic configurations, and end-effector morphologies, posing significant challenges when transferring manipulation skills from one platform to another. To tackle this problem, we propose a set of key design decisions to train a single policy for deployment on multiple robotic platforms. Our framework first aligns the observation and action spaces of our policy across embodiments via utilizing wrist cameras and a unified, but modular codebase. To bridge the remaining domain shift, we align our policy's internal representations across embodiments via contrastive learning. We evaluate our method on a dataset collected over 60 hours spanning 6 tasks and 3 robots with varying joint configurations and sizes: the WidowX 250S, Franka Emika Panda, and Sawyer. Our results demonstrate significant improvements in success rate and sample efficiency for our policy when using new task data collected on a different robot, validating our proposed design decisions. More details and videos can be found on our project website: https://sites.google.com/view/cradle-multirobot

Replicating human-like dexterity in robot hands represents one of the largest open problems in robotics. Reinforcement learning is a promising approach that has achieved impressive progress in the last few years; however, the class of problems it has typically addressed corresponds to a rather narrow definition of dexterity as compared to human capabilities. To address this gap, we investigate piano-playing, a skill that challenges even the human limits of dexterity, as a means to test high-dimensional control, and which requires high spatial and temporal precision, and complex finger coordination and planning. We introduce RoboPianist, a system that enables simulated anthropomorphic hands to learn an extensive repertoire of 150 piano pieces where traditional model-based optimization struggles. We additionally introduce an open-sourced environment, benchmark of tasks, interpretable evaluation metrics, and open challenges for future study. Our website featuring videos, code, and datasets is available at https://kzakka.com/robopianist/

Monocular 3D object detection has seen significant advancements with the incorporation of depth information. However, there remains a considerable performance gap compared to LiDAR-based methods, largely due to inaccurate depth estimation. We argue that this issue stems from the commonly used pixel-wise depth map loss, which inherently creates the imbalance of loss weighting between near and distant objects. To address these challenges, we propose MonoHBD (Monocular Hierarchical Balanced Depth), a comprehensive solution with the hierarchical mechanism. We introduce the Hierarchical Depth Map (HDM) structure that incorporates depth bins and depth offsets to enhance the localization accuracy for objects. Leveraging RoIAlign, our Balanced Depth Extractor (BDE) module captures both scene-level depth relationships and object-specific depth characteristics while considering the geometry properties through the inclusion of camera calibration parameters. Furthermore, we propose a novel depth map loss that regularizes object-level depth features to mitigate imbalanced loss propagation. Our model reaches state-of-the-art results on the KITTI 3D object detection benchmark while supporting real-time detection. Excessive ablation studies are also conducted to prove the efficacy of our proposed modules.

Heteroscedastic Gaussian processes (HGPs) are kernel-based, non-parametric models that can be used to infer nonlinear functions with time-varying noise. In robotics, they can be employed for learning from demonstration as motion primitives, i.e. as a model of the trajectories to be executed by the robot. HGPs provide variance estimates around the reference signal modeling the trajectory, capturing both the predictive uncertainty and the motion variability. However, similarly to standard Gaussian processes they suffer from a cubic complexity in the number of training points, due to the inversion of the kernel matrix. The uncertainty can be leveraged for more complex learning tasks, such as inferring the variable impedance profile required from a robotic manipulator. However, suitable approximations are needed to make HGPs scalable, at the price of potentially worsening the posterior mean and variance profiles. Motivated by these observations, we study the combination of HGPs and random features, which are a popular, data-independent approximation strategy of kernel functions. In a theoretical analysis, we provide novel guarantees on the approximation error of the HGP posterior due to random features. Moreover, we validate this scalable motion primitive on real robot data, related to the problem of variable impedance learning. In this way, we show that random features offer a viable and theoretically sound alternative for speeding up the trajectory processing, without sacrificing accuracy.

Imitation learning from human demonstrations can teach robots complex manipulation skills, but is time-consuming and labor intensive. In contrast, Task and Motion Planning (TAMP) systems are automated and excel at solving long-horizon tasks, but they are difficult to apply to contact-rich tasks. In this paper, we present Human-in-the-Loop Task and Motion Planning (HITL-TAMP), a novel system that leverages the benefits of both approaches. The system employs a TAMP-gated control mechanism, which selectively gives and takes control to and from a human teleoperator. This enables the human teleoperator to manage a fleet of robots, maximizing data collection efficiency. The collected human data is then combined with an imitation learning framework to train a TAMP-gated policy, leading to superior performance compared to training on full task demonstrations. We compared HITL-TAMP to a conventional teleoperation system --- users gathered more than 3x the number of demos given the same time budget. Furthermore, proficient agents (75\%+ success) could be trained from just 10 minutes of non-expert teleoperation data. Finally, we collected 2.1K demos with HITL-TAMP across 12 contact-rich, long-horizon tasks and show that the system often produces near-perfect agents. Videos and additional results at https://hitltamp.github.io .

Gestures serve as a fundamental and significant mode of non-verbal communication among humans. Deictic gestures (such as pointing towards an object), in particular, offer valuable means of efficiently expressing intent in situations where language is inaccessible, restricted, or highly specialized. As a result, it is essential for robots to comprehend gestures in order to infer human intentions and establish more effective coordination with them. Prior work often rely on a rigid hand-coded library of gestures along with their meanings. However, interpretation of gestures is often context-dependent, requiring more flexibility and common-sense reasoning. In this work, we propose a framework, GIRAF, for more flexibly interpreting gesture and language instructions by leveraging the power of large language models. Our framework is able to accurately infer human intent and contextualize the meaning of their gestures for more effective human-robot collaboration. We instantiate the framework for three table-top manipulation tasks and demonstrate that it is both effective and preferred by users. We further demonstrate GIRAF’s ability on reasoning about diverse types of gestures by curating a GestureInstruct dataset consisting of 36 different task scenarios. GIRAF achieved 81% success rate on finding the correct plan for tasks in GestureInstruct. Videos and datasets can be found on our project website: https://tinyurl.com/giraf23

Touch-based control is a promising approach to dexterous manipulation. However, existing tactile control methods often overlook tactile geometric aliasing which can compromise control performance and reliability. This type of aliasing occurs when different contact locations yield similar tactile signatures. To address this, we propose TactileVAD, a generative decoder-only linear latent dynamics formulation compatible with standard control methods that is capable of resolving geometric aliasing. We evaluate TactileVAD on two mechanically-distinct tactile sensors, SoftBubbles (pointcloud data) and Gelslim 3.0 (RGB data), showcasing its effectiveness in handling different sensing modalities. Additionally, we introduce the tactile cartpole, a novel benchmarking setup to evaluate the ability of a control method to respond to disturbances based on tactile input. Evaluations comparing TactileVAD to baselines suggest that our method is better able to achieve goal tactile configurations and hand poses.

We present a system that enables an autonomous small-scale RC car to drive aggressively from visual observations using reinforcement learning (RL). Our system, FastRLAP, trains autonomously in the real world, without human interventions, and without requiring any simulation or expert demonstrations. Our system integrates a number of important components to make this possible: we initialize the representations for the RL policy and value function from a large prior dataset of other robots navigating in other environments (at low speed), which provides a navigation-relevant representation. From here, a sample-efficient online RL method uses a single low-speed user-provided demonstration to determine the desired driving course, extracts a set of navigational checkpoints, and autonomously practices driving through these checkpoints, resetting automatically on collision or failure. Perhaps surprisingly, we find that with appropriate initialization and choice of algorithm, our system can learn to drive over a variety of racing courses with less than 20 minutes of online training. The resulting policies exhibit emergent aggressive driving skills, such as timing braking and acceleration around turns and avoiding areas which impede the robot's motion, approaching the performance of a human driver using a similar first-person interface over the course of training.