CoRL 2024 论文列表

Grasping in cluttered scenes remains highly challenging for dexterous hands due to the scarcity of data. To address this problem, we present a large-scale synthetic dataset, encompassing 1319 objects, 8270 scenes, and 426 million grasps. Beyond benchmarking, we also explore data-efficient learning strategies from grasping data. We reveal that the combination of a conditional generative model that focuses on local geometry and a grasp dataset that emphasizes complex scene variations is key to achieving effective generalization. Our proposed generative method outperforms all baselines in simulation experiments. Furthermore, it demonstrates zero-shot sim-to-real transfer through test-time depth restoration, attaining 91% real-world success rate, showcasing the robust potential of utilizing fully synthetic training data.

Tensegrity robots are composed of rigid struts and flexible cables. They constitute an emerging class of hybrid rigid-soft robotic systems and are promising systems for a wide array of applications, ranging from locomotion to assembly. They are difficult to control and model accurately, however, due to their compliance and high number of degrees of freedom. To address this issue, prior work has introduced a differentiable physics engine designed for tensegrity robots based on first principles. In contrast, this work proposes the use of graph neural networks to model contact dynamics over a graph representation of tensegrity robots, which leverages their natural graph-like cable connec- tivity between end caps of rigid rods. This learned simulator can accurately model 3-bar and 6-bar tensegrity robot dynamics in simulation-to-simulation experiments where MuJoCo is used as the ground truth. It can also achieve higher accuracy than the previous differentiable engine for a real 3-bar tensegrity robot, for which the robot state is only partially observable. When compared against direct applications of recent mesh-based graph neural network simulators, the proposed approach is computationally more efficient, both for training and inference, while achieving higher accuracy. Code and data are available at https://github.com/nchen9191/tensegrity_gnn_simulator_public

Humans can imagine goal states during planning and perform actions to match those goals. In this work, we propose IMAGINATION POLICY, a novel multi-task key-frame policy network for solving high-precision pick and place tasks. Instead of learning actions directly, IMAGINATION POLICY generates point clouds to imagine desired states which are then translated to actions using rigid action estimation. This transforms action inference into a local generative task. We leverage pick and place symmetries underlying the tasks in the generation process and achieve extremely high sample efficiency and generalizability to unseen configurations. Finally, we demonstrate state-of-the-art performance across various tasks on the RLbench benchmark compared with several strong baselines and validate our approach on a real robot.

To train generalist robot policies, machine learning methods often require a substantial amount of expert human teleoperation data. An ideal robot for humans collecting data is one that closely mimics them: bimanual arms and dexterous hands. However, creating such a bimanual teleoperation system with over 50 DoF is a significant challenge. To address this, we introduce Bidex, an extremely dexterous, low-cost, low-latency and portable bimanual dexterous teleoperation system which relies on motion capture gloves and teacher arms. We compare Bidex to a Vision Pro teleoperation system and a SteamVR system and find Bidex to produce better quality data for more complex tasks at a faster rate. Additionally, we show Bidex operating a mobile bimanual robot for in the wild tasks. Please refer to https://bidex-teleop.github.io for video results and instructions to recreate Bidex. The robot hands (5k USD) and teleoperation system (7k USD) is readily reproducible and can be used on many robot arms including two xArms ($16k USD).

Endowing robot hands with human-level dexterity is a long-lasting research objective. Bi-manual robot piano playing constitutes a task that combines challenges from dynamic tasks, such as generating fast while precise motions, with slower but contact-rich manipulation problems. Although reinforcement learning based approaches have shown promising results in single-task performance, these methods struggle in a multi-song setting. Our work aims to close this gap and, thereby, enable imitation learning approaches for robot piano playing at scale. To this end, we introduce the Robot Piano 1 Million (RP1M) dataset, containing bi-manual robot piano playing motion data of more than one million trajectories. We formulate finger placements as an optimal transport problem, thus, enabling automatic annotation of vast amounts of unlabeled songs. Benchmarking existing imitation learning approaches shows that such approaches reach state-of-the-art robot piano playing performance by leveraging RP1M.

We present a fully autonomous real-world RL framework for mobile manipulation that can learn policies without extensive instrumentation or human supervision. This is enabled by 1) task-relevant autonomy, which guides exploration towards object interactions and prevents stagnation near goal states, 2) efficient policy learning by leveraging basic task knowledge in behavior priors, and 3) formulating generic rewards that combine human-interpretable semantic information with low-level, fine-grained observations. We demonstrate that our approach allows Spot robots to continually improve their performance on a set of four challenging mobile manipulation tasks, obtaining an average success rate of 80% across tasks, a 3-4 times improvement over existing approaches. Videos can be found at https://continual-mobile-manip.github.io/.

Manipulating objects with two multi-fingered hands has been a long-standing challenge in robotics, due to the contact-rich nature of many manipulation tasks and the complexity inherent in coordinating a high-dimensional bimanual system. In this work, we share novel insights into physical modeling, real-time perception, and reward design that enable policies trained in simulation using deep reinforcement learning (RL) to be effectively and efficiently transferred to the real world. Specifically, we consider the problem of twisting lids of various bottle-like objects with two hands, demonstrating policies with generalization capabilities across a diverse set of unseen objects as well as dynamic and dexterous behaviors. To the best of our knowledge, this is the first sim-to-real RL system that enables such capabilities on bimanual multi-fingered hands.

Perceiving the environment via cameras is crucial for Reinforcement Learning (RL) in robotics. While images are a convenient form of representation, they often complicate extracting important geometric details, especially with varying geometries or deformable objects. In contrast, point clouds naturally represent this geometry and easily integrate color and positional data from multiple camera views. However, while point-cloud processing with deep learning has seen many recent successes, RL on point clouds is under-researched, with only the simplest encoder architecture considered in the literature. We introduce PointPatchRL (PPRL), a method for RL on point clouds that builds on the common paradigm of dividing point clouds into overlapping patches, tokenizing them, and processing the tokens with transformers. PPRL provides significant improvements compared with other point-cloud processing architectures previously used for RL. We then complement PPRL with masked reconstruction for representation learning and show that our method outperforms strong model-free and model-based baselines on image observations in complex manipulation tasks containing deformable objects and variations in target object geometry.

We introduce UMI-on-Legs, a new framework that combines real-world and simulation data for quadruped manipulation systems. We scale task-centric data collection in the real world using a handheld gripper (UMI), providing a cheap way to demonstrate task-relevant manipulation skills without a robot. Simultaneously, we scale robot-centric data in simulation by training a whole-body controller. The interface between these two policies are end-effector trajectories in the task-frame, which are inferred by the manipulation policy and passed to the whole-body controller for tracking. We evaluate UMI-on-Legs on prehensile, non-prehensile, and dynamic manipulation tasks, and report over 70% success rate for all tasks. Lastly, we also demonstrate the zero-shot cross-embodiment deployment of a pre-trained manipulation policy checkpoint from a prior work, originally intended for a fixed-base robot arm, on our quadruped system. We believe this framework provides a scalable path towards learning expressive manipulation skills on dynamic robot embodiments.

Generating safety-critical scenarios, which are essential yet difficult to collect at scale, offers an effective method to evaluate the robustness of autonomous vehicles (AVs). Existing methods focus on optimizing adversariality while preserving the naturalness of scenarios, aiming to achieve a balance through data-driven approaches. However, without an appropriate upper bound for adversariality, the scenarios might exhibit excessive adversariality, potentially leading to unavoidable collisions. In this paper, we introduce FREA, a novel safety-critical scenarios generation method that incorporates the Largest Feasible Region (LFR) of AV as guidance to ensure the reasonableness of the adversarial scenarios. Concretely, FREA initially pre-calculates the LFR of AV from offline datasets. Subsequently, it learns a reasonable adversarial policy that controls critical background vehicles (CBVs) in the scene to generate adversarial yet AV-feasible scenarios by maximizing a novel feasibility-dependent objective function. Extensive experiments illustrate that FREA can effectively generate safety-critical scenarios, yielding considerable near-miss events while ensuring AV's feasibility. Generalization analysis also confirms the robustness of FREA in AV testing across various surrogate AV methods and traffic environments.

We introduce a novel approach to manipulate articulated objects with ambiguities, such as opening a door, in which multi-modality and occlusions create ambiguities about the opening side and direction. Multi-modality occurs when the method to open a fully closed door (push, pull, slide) is uncertain, or the side from which it should be opened is uncertain. Occlusions further obscure the door’s shape from certain angles, creating further ambiguities during the occlusion. To tackle these challenges, we propose a history-aware diffusion network that models the multi-modal distribution of the articulated object and uses history to disambiguate actions and make stable predictions under occlusions. Experiments and analysis demonstrate the state-of-art performance of our method and specifically improvements in ambiguity-caused failure modes. Our project website is available at https://flowbothd.github.io/.

Recent progress in deep reinforcement learning (RL) and computer vision enables artificial agents to solve complex tasks, including locomotion, manipulation, and video games from high-dimensional pixel observations. However, RL usually relies on domain-specific reward functions for sufficient learning signals, requiring expert knowledge. While vision-based agents could learn skills from only sparse rewards, exploration challenges arise. We present Latent Nearest-demonstration-guided Exploration (LaNE), a novel and efficient method to solve sparse-reward robot manipulation tasks from image observations and a few demonstrations. First, LaNE builds on the pre-trained DINOv2 feature extractor to learn an embedding space for forward prediction. Next, it rewards the agent for exploring near the demos, quantified by quadratic control costs in the embedding space. Finally, LaNE optimizes the policy for the augmented rewards with RL. Experiments demonstrate that our method achieves state-of-the-art sample efficiency in Robosuite simulation and enables under-an-hour RL training from scratch on a Franka Panda robot, using only a few demonstrations.

In complex real-world tasks such as robotic manipulation and autonomous driving, collecting expert demonstrations is often more straightforward than specifying precise learning objectives and task descriptions. Learning from expert data can be achieved through behavioral cloning or by learning a reward function, i.e., inverse reinforcement learning. The latter allows for training with additional data outside the training distribution, guided by the inferred reward function. We propose a novel approach to construct compact and interpretable reward models from automatically selected state features. These inferred rewards have an explicit form and enable the learning of policies that closely match expert behavior by training standard reinforcement learning algorithms from scratch. We validate our method's performance in various robotic environments with continuous and high-dimensional state spaces.

Large-scale endeavors like RT-1 and widespread community efforts such as Open-X-Embodiment have contributed to growing the scale of robot demonstration data. However, there is still an opportunity to improve the quality, quantity, and diversity of robot demonstration data. Although vision-language models have been shown to automatically generate demonstration data, their utility has been limited to environments with privileged state information, they require hand-designed skills, and are limited to interactions with few object instances. We propose Manipulate-Anything, a scalable automated generation method for real-world robotic manipulation. Unlike prior work, our method can operate in real-world environments without any privileged state information, hand-designed skills, and can manipulate any static object. We evaluate our method using two setups. First, Manipulate-Anything successfully generates trajectories for all 5 real-world and 12 simulation tasks, significantly outperforming existing methods like VoxPoser. Second, Manipulate-Anything's demonstrations can train more robust behavior cloning policies than training with human demonstrations, or from data generated by VoxPoser and Code-As-Policies. We believe Manipulate-Anything can be the scalable method for both generating data for robotics and solving novel tasks in a zero-shot setting. Anonymous project page: manipulate-anything.github.io.

Humans can learn to manipulate new objects by simply watching others; providing robots with the ability to learn from such demonstrations would enable a natural interface specifying new behaviors. This work develops Robot See Robot Do (RSRD), a method for imitating articulated object manipulation from a single monocular RGB human demonstration given a single static multi- view object scan. We first propose 4D Differentiable Part Models (4D-DPM), a method for recovering 3D part motion from a monocular video with differentiable rendering. This analysis-by-synthesis approach uses part-centric feature fields in an iterative optimization which enables the use of geometric regularizers to re- cover 3D motions from only a single video. Given this 4D reconstruction, the robot replicates object trajectories by planning bimanual arm motions that induce the demonstrated object part motion. By representing demonstrations as part- centric trajectories, RSRD focuses on replicating the demonstration’s intended behavior while considering the robot’s own morphological limits, rather than at- tempting to reproduce the hand’s motion. We evaluate 4D-DPM’s 3D tracking accuracy on ground truth annotated 3D part trajectories and RSRD’s physical ex- ecution performance on 9 objects across 10 trials each on a bimanual YuMi robot. Each phase of RSRD achieves an average of 87% success rate, for a total end- to-end success rate of 60% across 90 trials. Notably, this is accomplished using only feature fields distilled from large pretrained vision models — without any task-specific training, fine-tuning, dataset collection, or annotation. Project page: https://robot-see-robot-do.github.io

As the deployment of autonomous agents in real-world scenarios grows, so does the interest in their application to competitive environments with other robots. Self-play in Reinforcement Learning (RL) enables agents to develop competitive strategies. However, the complexity arising from multi-agent interactions and the tendency for RL agents to disrupt competitors' training introduce instability and a risk of overfitting. While traditional methods depend on costly Nash equilibrium approximations or random exploration for training scenario optimization, this can be inefficient in large search spaces often prevalent in multi-agent problems. However, related works in single-agent setups show that genetic algorithms perform better in large scenario spaces. Therefore, we propose using genetic algorithms to adaptively adjust environment parameters and opponent policies in a multi-agent context to find and synthesize coherent scenarios efficiently. We also introduce GenOpt Agent—a genetically optimized, open-loop agent executing scheduled actions. The open-loop aspect of GenOpt prevents RL agents from winning through adversarial perturbations, thereby fostering generalizable strategies. Also, GenOpt is genetically optimized without expert supervision, negating the need for expensive expert supervision to have meaningful opponents at the start of training. Our empirical studies indicate that this method surpasses several established baselines in two-player competitive settings with continuous action spaces, validating its effectiveness and stability in training.

Legged robots are able to overcome challenging terrains through diverse gaits formed by contact sequences. However, environments characterized by discrete footholds present significant challenges. In this paper, we tackle the problem of free gait motion planning for hexapod robots walking in randomly generated plum blossom pile environments. Specifically, we first address the complexity of multi-leg coordination in discrete environments by treating each leg of the hexapod robot as an individual agent. Then, we propose the Hybrid action space Multi-Agent Soft Actor Critic (Hybrid-MASAC) algorithm capable of handling both discrete and continuous actions. Finally, we present an integrated free gait motion planning method based on Hybrid-MASAC, streamlining gait, Center of Mass (COM), and foothold sequences planning into a single model. Comparative and ablation experiments in both of the simulated and real plum blossom pile environments demonstrate the feasibility and efficiency of our method.

Learning from human feedback has gained traction in fields like robotics and natural language processing in recent years. While prior works mostly rely on human feedback in the form of comparisons, language is a preferable modality that provides more informative insights into user preferences. In this work, we aim to incorporate comparative language feedback to iteratively improve robot trajectories and to learn reward functions that encode human preferences. To achieve this goal, we learn a shared latent space that integrates trajectory data and language feedback, and subsequently leverage the learned latent space to improve trajectories and learn human preferences. To the best of our knowledge, we are the first to incorporate comparative language feedback into reward learning. Our simulation experiments demonstrate the effectiveness of the learned latent space and the success of our learning algorithms. We also conduct human subject studies that show our reward learning algorithm achieves a 23.9% higher subjective score on average and is 11.3% more time-efficient compared to preference-based reward learning, underscoring the superior performance of our method. Our website is at https://liralab.usc.edu/comparative-language-feedback/.

This work explores conditions under which multi-finger grasping algorithms can attain robust sim-to-real transfer. While numerous large datasets facilitate learning *generative* models for multi-finger grasping at scale, reliable real-world dexterous grasping remains challenging, with most methods degrading when deployed on hardware. An alternate strategy is to use *discriminative* grasp evaluation models for grasp selection and refinement, conditioned on real-world sensor measurements. This paradigm has produced state-of-the-art results for vision-based parallel-jaw grasping, but remains unproven in the multi-finger setting. In this work, we find that existing datasets and methods have been insufficient for training discriminitive models for multi-finger grasping. To train grasp evaluators at scale, datasets must provide on the order of millions of grasps, including both positive *and negative examples*, with corresponding visual data resembling measurements at inference time. To that end, we release a new, open-source dataset of 3.5M grasps on 4.3K objects annotated with RGB images, point clouds, and trained NeRFs. Leveraging this dataset, we train vision-based grasp evaluators that outperform both analytic and generative modeling-based baselines on extensive simulated and real-world trials across a diverse range of objects. We show via numerous ablations that the key factor for performance is indeed the evaluator, and that its quality degrades as the dataset shrinks, demonstrating the importance of our new dataset. Project website at: https://sites.google.com/view/get-a-grip-dataset.