CoRL 2024 Papers

Payload transport over flat terrain via multi-wheel robot carriers is well-understood, highly effective, and configurable. In this paper, our goal is to provide similar effectiveness and configurability for transport over rough terrain that is more suitable for legs rather than wheels. For this purpose, we consider multi-biped robot carriers, where wheels are replaced by multiple bipedal robots attached to the carrier. Our main contribution is to design a decentralized controller for such systems that can be effectively applied to varying numbers and configurations of rigidly attached bipedal robots without retraining. We present a reinforcement learning approach for training the controller in simulation that supports transfer to the real world. Our experiments in simulation provide quantitative metrics showing the effectiveness of the approach over a wide variety of simulated transport scenarios. In addition, we demonstrate the controller in the real-world for systems composed of two and three Cassie robots. To our knowledge, this is the first example of a scalable multi-biped payload transport system.

Recent works have proposed a number of general-purpose robotic foundation models that can control a variety of robotic platforms to perform a range of different tasks, including in the domains of navigation and manipulation. However, such models are typically trained via imitation learning, which precludes the ability to improve autonomously through experience that the robot gathers on the job. In this work, our aim is to train general-purpose robotic foundation models in the domain of robotic navigation specifically with the aim of enabling autonomous self-improvement. We show that a combination of pretraining with offline reinforcement learning and a complete system for continual autonomous operation leads to a robotic learning framework that not only starts off with broad and diverse capabilities, but can further improve and adapt those capabilities in the course of carrying out navigational tasks in a given deployment location. To our knowledge, our model LiReN is the first navigation robot foundation model that is capable of fine-tuning with autonomous online data in open-world settings.

Building effective imitation learning methods that enable robots to learn from limited data and still generalize across diverse real-world environments is a long-standing problem in robot learning. We propose EquiBot, a robust, data-efficient, and generalizable approach for robot manipulation task learning. Our approach combines SIM(3)-equivariant neural network architectures with diffusion models. This ensures that our learned policies are invariant to changes in scale, rotation, and translation, enhancing their applicability to unseen environments while retaining the benefits of diffusion-based policy learning such as multi-modality and robustness. We show on a suite of 6 simulation tasks that our proposed method reduces the data requirements and improves generalization to novel scenarios. In the real world, with 10 variations of 6 mobile manipulation tasks, we show that our method can easily generalize to novel objects and scenes after learning from just 5 minutes of human demonstrations in each task.

Simulators are powerful tools for autonomous robot learning as they offer scalable data generation, flexible design, and optimization of trajectories. However, transferring behavior learned from simulation data into the real world proves to be difficult, usually mitigated with compute-heavy domain randomization methods or further model fine-tuning. We present a method to improve generalization and robustness to distribution shifts in sim-to-real visual quadrotor navigation tasks. To this end, we first build a simulator by integrating Gaussian Splatting with quadrotor flight dynamics, and then, train robust navigation policies using Liquid neural networks. In this way, we obtain a full-stack imitation learning protocol that combines advances in 3D Gaussian splatting radiance field rendering, crafty programming of expert demonstration training data, and the task understanding capabilities of Liquid networks. Through a series of quantitative flight tests, we demonstrate the robust transfer of navigation skills learned in a single simulation scene directly to the real world. We further show the ability to maintain performance beyond the training environment under drastic distribution and physical environment changes. Our learned Liquid policies, trained on single target maneuvers curated from a photorealistic simulated indoor flight only, generalize to multi-step hikes onboard a real hardware platform outdoors.

Stable locomotion in precipitous environments is an essential task for quadruped robots, requiring the ability to resist various external disturbances. Recent neural policies enhance robustness against disturbances by learning to resist external forces sampled from a fixed distribution in the simulated environment. However, the force generation process doesn’t consider the robot’s current state, making it difficult to identify the most effective direction and magnitude that can push the robot to the most unstable but recoverable state. Thus, challenging cases in the buffer are insufficient to optimize robustness. In this paper, we propose to model the robust locomotion learning process as an adversarial interaction between the locomotion policy and a learnable disturbance that is conditioned on the robot state to generate appropriate external forces. To make the joint optimization stable, our novel $H_{\infty}$ constraint mandates the bound of the ratio between the cost and the intensity of the external forces. We verify the robustness of our approach in both simulated environments and real-world deployment, on quadrupedal locomotion tasks and a more challenging task where the quadruped performs locomotion merely on hind legs. Training and deployment code will be made public.

For 6-DoF grasp detection, simulated data is expandable to train more powerful model, but it faces the challenge of the large gap between simulation and real world. Previous works bridge this gap with a sim-to-real way. However, this way explicitly or implicitly forces the simulated data to adapt to the noisy real data when training grasp detectors, where the positional drift and structural distortion within the camera noise will harm the grasp learning. In this work, we propose a Real-to-Sim framework for 6-DoF Grasp detection, named R2SGrasp, with the key insight of bridging this gap in a real-to-sim way, which directly bypasses the camera noise in grasp detector training through an inference-time real-to-sim adaption. To achieve this real-to-sim adaptation, our R2SGrasp designs the Real-to-Sim Data Repairer (R2SRepairer) to mitigate the camera noise of real depth maps in data-level, and the Real-to-Sim Feature Enhancer (R2SEnhancer) to enhance real features with precise simulated geometric primitives in feature-level. To endow our framework with the generalization ability, we construct a large-scale simulated dataset cost-efficiently to train our grasp detector, which includes 64,000 RGB-D images with 14.4 million grasp annotations. Sufficient experiments show that R2SGrasp is powerful and our real-to-sim perspective is effective. The real-world experiments further show great generalization ability of R2SGrasp. Project page is available on https://isee-laboratory.github.io/R2SGrasp.

Demonstrations are an effective alternative to task specification for learning agents in settings where designing a reward function is difficult. However, demonstrating expert behavior in the action space of the agent becomes unwieldy when robots have complex, unintuitive morphologies. We consider the practical setting where an agent has a dataset of prior interactions with the environment and is provided with observation-only expert demonstrations. Typical learning from observations approaches have required either learning an inverse dynamics model or a discriminator as intermediate steps of training. Errors in these intermediate one-step models compound during downstream policy learning or deployment. We overcome these limitations by directly learning a multi-step utility function that quantifies how each action impacts the agent's divergence from the expert's visitation distribution. Using the principle of duality, we derive DILO (Dual Imitation Learning from Observations), an algorithm that can leverage arbitrary suboptimal data to learn imitating policies without requiring expert actions. DILO reduces the learning from observations problem to that of simply learning an actor and a critic, bearing similar complexity to vanilla offline RL. This allows DILO to gracefully scale to high dimensional observations, and demonstrate improved performance across the board.

Most reinforcement learning (RL) methods focus on learning optimal policies over low-level action spaces. While these methods can perform well in their training environments, they lack the flexibility to transfer to new tasks. Instead, RL agents that can act over useful, temporally extended skills rather than low-level actions can learn new tasks more easily. Prior work in skill-based RL either requires expert supervision to define useful skills, which is hard to scale, or learns a skill-space from offline data with heuristics that limit the adaptability of the skills, making them difficult to transfer during downstream RL. Our approach, EXTRACT, instead utilizes pre-trained vision language models to extract a discrete set of semantically meaningful skills from offline data, each of which is parameterized by continuous arguments, without human supervision. This skill parameterization allows robots to learn new tasks by only needing to learn when to select a specific skill and how to modify its arguments for the specific task. We demonstrate through experiments in sparse-reward, image-based, robot manipulation environments that EXTRACT can more quickly learn new tasks than prior works, with major gains in sample efficiency and performance over prior skill-based RL.

Natural language and images are commonly used as goal representations in goal-conditioned imitation learning. However, language can be ambiguous and images can be over-specified. In this work, we study hand-drawn sketches as a modality for goal specification. Sketches can be easy to provide on the fly like language, but like images they can also help a downstream policy to be spatially-aware. By virtue of being minimal, sketches can further help disambiguate task-relevant from irrelevant objects. We present RT-Sketch, a goal-conditioned policy for manipulation that takes a hand-drawn sketch of the desired scene as input, and outputs actions. We train RT-Sketch on a dataset of trajectories paired with synthetically generated goal sketches. We evaluate this approach on six manipulation skills involving tabletop object rearrangements on an articulated countertop. Experimentally we find that RT-Sketch performs comparably to image or language-conditioned agents in straightforward settings, while achieving greater robustness when language goals are ambiguous or visual distractors are present. Additionally, we show that RT-Sketch handles sketches with varied levels of specificity, ranging from minimal line drawings to detailed, colored drawings. For supplementary material and videos, please visit http://rt-sketch.github.io.

Large-scale visuomotor policy learning is a promising approach toward developing generalizable manipulation systems. Yet, policies that can be deployed on diverse embodiments, environments, and observational modalities remain elusive. In this work, we investigate how knowledge from large-scale visual data of the world may be used to address one axis of variation for generalizable manipulation: observational viewpoint. Specifically, we study single-image novel view synthesis models, which learn 3D-aware scene-level priors by rendering images of the same scene from alternate camera viewpoints given a single input image. For practical application to diverse robotic data, these models must operate *zero-shot*, performing view synthesis on unseen tasks and environments. We empirically analyze view synthesis models within a simple data-augmentation scheme that we call View Synthesis Augmentation (VISTA) to understand their capabilities for learning viewpoint-invariant policies from single-viewpoint demonstration data. Upon evaluating the robustness of policies trained with our method to out-of-distribution camera viewpoints, we find that they outperform baselines in both simulated and real-world manipulation tasks.

In this work, we introduce PianoMime, a framework for training a piano-playing agent using internet demonstrations. The internet is a promising source of large-scale demonstrations for training our robot agents. In particular, for the case of piano-playing, Youtube is full of videos of professional pianists playing a wide myriad of songs. In our work, we leverage these demonstrations to learn a generalist piano-playing agent capable of playing any arbitrary song. Our framework is divided into three parts: a data preparation phase to extract the informative features from the Youtube videos, a policy learning phase to train song-specific expert policies from the demonstrations and a policy distillation phase to distil the policies into a single generalist agent. We explore different policy designs to represent the agent and evaluate the influence of the amount of training data on the generalization capability of the agent to novel songs not available in the dataset. We show that we are able to learn a policy with up to 57% F1 score on unseen songs.

Applying reinforcement learning (RL) to learn effective policies on physical robots without supervision remains challenging when it comes to tasks where safe exploration is critical. Constrained model-based RL (CMBRL) presents a promising approach to this problem. These methods are designed to learn constraint-adhering policies through constrained optimization approaches. Yet, such policies often fail to meet stringent safety requirements during learning and exploration. Our solution ``CASE'' aims to reduce the instances where constraints are breached during the learning phase. Specifically, CASE integrates techniques for optimizing constrained policies and employs planning-based safety filters as backup policies, effectively lowering constraint violations during learning and making it a more reliable option than other recent constrained model-based policy optimization methods.

Traditional autonomous driving methods adopt modular design, decomposing tasks into sub-tasks, including perception, prediction, planning, and control. In contrast, end-to-end autonomous driving directly outputs actions from raw sensor data, avoiding error accumulation. However, training an end-to-end model requires a comprehensive dataset. Without adequate data, the end-to-end model exhibits poor generalization capabilities. Recently, large language models (LLMs) have been applied to enhance the generalization property of end-to-end driving models. Most studies explore LLMs in an open-loop manner, where the output actions are compared to those of experts without direct activation in the real world. Other studies in closed-loop settings examine their results in simulated environments. In comparison, this paper proposes an efficient architecture that integrates multimodal LLMs into end-to-end real-world driving models in a closed-loop setting. The LLM periodically takes raw sensor data to generate high-level driving instructions. In our architecture, LLMs can effectively guide the end-to-end model, even at a slower rate than the raw sensor data, because updates aren't needed every time frame. This architecture relaxes the trade-off between the latency and inference quality of the LLM. It also allows us to choose a wide variety of LLMs to improve high-level driving instructions and minimize fine-tuning costs. Consequently, our architecture reduces the data collection requirements because the LLMs do not directly output actions, and we only need to train a simple imitation learning model to output actions. In our experiments, the training data for the end-to-end model in a real-world environment consists of only simple obstacle configurations with one traffic cone, while the test environment is more complex and contains different types of obstacles. Experiments show that the proposed architecture enhances the generalization capabilities of the end-to-end model even without fine-tuning the LLM.

Adjusting robot behavior to human preferences can require intensive human feedback, preventing quick adaptation to new users and changing circumstances. Moreover, current approaches typically treat user preferences as a reward, which requires a manual balance between task success and user satisfaction. To integrate new user preferences in a zero-shot manner, our proposed Text2Interaction framework invokes large language models to generate a task plan, motion preferences as Python code, and parameters of a safety controller. By maximizing the combined probability of task completion and user satisfaction instead of a weighted sum of rewards, we can reliably find plans that fulfill both requirements. We find that 83% of users working with Text2Interaction agree that it integrates their preferences into the plan of the robot, and 94% prefer Text2Interaction over the baseline. Our ablation study shows that Text2Interaction aligns better with unseen preferences than other baselines while maintaining a high success rate. Real-world demonstrations and code are made available at [sites.google.com/view/text2interaction](sites.google.com/view/text2interaction).

The task of "relative placement" is to predict the placement of one object in relation to another, e.g. placing a mug on a mug rack. Recent methods for relative placement have made tremendous progress towards data-efficient learning for robot manipulation; using explicit object-centric geometric reasoning, these approaches enable generalization to unseen task variations from a small number of demonstrations. State-of-the-art works in this area, however, have yet to represent deformable transformations, despite the ubiquity of non-rigid bodies in real world settings. As a first step towards bridging this gap, we propose "cross-displacement" - an extension of the principles of relative placement to geometric relationships between deformable objects - and present a novel vision-based method to learn cross-displacement for a non-rigid task through dense diffusion. To this end, we demonstrate our method's ability to generalize to unseen object instances, out-of-distribution scene configurations, and multimodal goals on a highly deformable cloth-hanging task beyond the scope of prior works.

Large language models (LLMs) can provide rich physical descriptions of most worldly objects, allowing robots to achieve more informed and capable grasping. We leverage LLMs' common sense physical reasoning and code-writing abilities to infer an object's physical characteristics-mass $m$, friction coefficient $\mu$, and spring constant $k$-from a semantic description, and then translate those characteristics into an executable adaptive grasp policy. Using a two-finger gripper with a built-in depth camera that can control its torque by limiting motor current, we demonstrate that LLM-parameterized but first-principles grasp policies outperform both traditional adaptive grasp policies and direct LLM-as-code policies on a custom benchmark of 12 delicate and deformable items including food, produce, toys, and other everyday items, spanning two orders of magnitude in mass and required pick-up force. We then improve property estimation and grasp performance on variable size objects with model finetuning on property-based comparisons and eliciting such comparisons via chain-of-thought prompting. We also demonstrate how compliance feedback from DeliGrasp policies can aid in downstream tasks such as measuring produce ripeness. Our code and videos are available at: https://deligrasp.github.io

Solving complex manipulation tasks in household and factory settings remains challenging due to long-horizon reasoning, fine-grained interactions, and broad object and scene diversity. Learning skills from demonstrations can be an effective strategy, but such methods often have limited generalizability beyond training data and struggle to solve long-horizon tasks. To overcome this, we propose to synergistically combine two paradigms: Neural Object Descriptors (NODs) that produce generalizable object-centric features and Task and Motion Planning (TAMP) frameworks that chain short-horizon skills to solve multi-step tasks. We introduce NOD-TAMP, a TAMP-based framework that extracts short manipulation trajectories from a handful of human demonstrations, adapts these trajectories using NOD features, and composes them to solve broad long-horizon, contact-rich tasks. NOD-TAMP solves existing manipulation benchmarks with a handful of demonstrations and significantly outperforms prior NOD-based approaches on new tabletop manipulation tasks that require diverse generalization. Finally, we deploy NOD-TAMP on a number of real-world tasks, including tool-use and high-precision insertion. For more details, please visit https://nodtamp.github.io/.

Autonomous self-improving robots that interact and improve with experience are key to the real-world deployment of robotic systems. In this paper, we propose an online learning method, SELFI, that leverages online robot experience to rapidly fine-tune pre-trained control policies efficiently. SELFI applies online model-free reinforcement learning on top of offline model-based learning to bring out the best parts of both learning paradigms. Specifically, SELFI stabilizes the online learning process by incorporating the same model-based learning objective from offline pre-training into the Q-values learned with online model-free reinforcement learning. We evaluate SELFI in multiple real-world environments and report improvements in terms of collision avoidance, as well as more socially compliant behavior, measured by a human user study. SELFI enables us to quickly learn useful robotic behaviors with less human interventions such as pre-emptive behavior for the pedestrians, collision avoidance for small and transparent objects, and avoiding travel on uneven floor surfaces. We provide supplementary videos to demonstrate the performance of our fine-tuned policy.

Object-level mapping builds a 3D map of objects in a scene with detailed shapes and poses from multi-view sensor observations. Conventional methods struggle to build complete shapes and estimate accurate poses due to partial occlusions and sensor noise. They require dense observations to cover all objects, which is challenging to achieve in robotics trajectories. Recent work introduces generative shape priors for object-level mapping from sparse views, but is limited to single-category objects. In this work, we propose a General Object-level Mapping system, GOM, which leverages a 3D diffusion model as shape prior with multi-category support and outputs Neural Radiance Fields (NeRFs) for both texture and geometry for all objects in a scene. GOM includes an effective formulation to guide a pre-trained diffusion model with extra nonlinear constraints from sensor measurements without finetuning. We also develop a probabilistic optimization formulation to fuse multi-view sensor observations and diffusion priors for joint 3D object pose and shape estimation. Our GOM system demonstrates superior multi-category mapping performance from sparse views, and achieves more accurate mapping results compared to state-of-the-art methods on the real-world benchmarks. We will release our code and model upon publication.