CoRL 2024 论文列表

The increasing complexity of tasks in robotics demands efficient strategies for multitask and continual learning. Traditional models typically rely on a universal policy for all tasks, facing challenges such as high computational costs and catastrophic forgetting when learning new tasks. To address these issues, we introduce a sparse, reusable, and flexible policy, Sparse Diffusion Policy (SDP). By adopting Mixture of Experts (MoE) within a transformer-based diffusion policy, SDP selectively activates experts and skills, enabling task-specific learning without retraining the entire model. It not only reduces the burden of active parameters but also facilitates the seamless integration and reuse of experts across various tasks. Extensive experiments on diverse tasks in both simulators and the real world show that SDP 1) excels in multitask scenarios with negligible increases in active parameters, 2) prevents forgetting in continual learning new tasks, and 3) enables efficient task transfer, offering a promising solution for advanced robotic applications. More demos and codes can be found on our https://anonymous.4open.science/w/sparse_diffusion_policy-24E7/.

We present our method, LeLaN, which uses action-free egocentric data to learn robust language-conditioned object navigation. By leveraging the knowledge of large vision and language models and grounding this knowledge using pre-trained segmentation and depth estimation models, we can label in-the-wild data from a variety of indoor and outdoor environments with diverse instructions that capture a range of objects with varied granularity and noise in their descriptions. Leveraging this method to label over 50 hours of data collected in indoor and outdoor environments, including robot observations, YouTube video tours, and human-collected walking data allows us to train a policy that can outperform state-of-the-art methods on the zero-shot object navigation task in both success rate and precision.

Robot behavior policies trained via imitation learning are prone to failure under conditions that deviate from their training data. Thus, algorithms that monitor learned policies at test time and provide early warnings of failure are necessary to facilitate scalable deployment. We propose Sentinel, a runtime monitoring framework that splits the detection of failures into two complementary categories: 1) Erratic failures, which we detect using statistical measures of temporal action consistency, and 2) task progression failures, where we use Vision Language Models (VLMs) to detect when the policy confidently and consistently takes actions that do not solve the task. Our approach has two key strengths. First, because learned policies exhibit diverse failure modes, combining complementary detectors leads to significantly higher accuracy at failure detection. Second, using a statistical temporal action consistency measure ensures that we quickly detect when multimodal, generative policies exhibit erratic behavior at negligible computational cost. In contrast, we only use VLMs to detect modes that are less time-sensitive. We demonstrate our approach in the context of diffusion policies trained on robotic mobile manipulation domains in both simulation and the real world. By unifying temporal consistency detection and VLM runtime monitoring, Sentinel detects 18% more failures than using either of the two detectors alone and significantly outperforms baselines, thus highlighting the importance of assigning specialized detectors to complementary categories of failure. Qualitative results are made available at sites.google.com/stanford.edu/sentinel.

We introduce Vocal Sandbox, a framework for enabling seamless human-robot collaboration in situated environments. Systems in our framework are characterized by their ability to *adapt and continually learn* at multiple levels of abstraction from diverse teaching modalities such as spoken dialogue, object keypoints, and kinesthetic demonstrations. To enable such adaptation, we design lightweight and interpretable learning algorithms that allow users to build an understanding and co-adapt to a robot's capabilities in real-time, as they teach new behaviors. For example, after demonstrating a new low-level skill for "tracking around" an object, users are provided with trajectory visualizations of the robot's intended motion when asked to track a new object. Similarly, users teach high-level planning behaviors through spoken dialogue, using pretrained language models to synthesize behaviors such as "packing an object away" as compositions of low-level skills -- concepts that can be reused and built upon. We evaluate Vocal Sandbox in two settings: collaborative gift bag assembly and LEGO stop-motion animation. In the first setting, we run systematic ablations and user studies with 8 non-expert participants, highlighting the impact of multi-level teaching. Across 23 hours of total robot interaction time, users teach 17 new high-level behaviors with an average of 16 novel low-level skills, requiring 22.1% less active supervision compared to baselines. Qualitatively, users strongly prefer Vocal Sandbox systems due to their ease of use (+31.2%), helpfulness (+13.0%), and overall performance (+18.2%). Finally, we pair an experienced system-user with a robot to film a stop-motion animation; over two hours of continuous collaboration, the user teaches progressively more complex motion skills to produce a 52 second (232 frame) movie. Videos & Supplementary Material: https://vocal-sandbox.github.io

Vision-based robot policy learning, which maps visual inputs to actions, necessitates a holistic understanding of diverse visual tasks beyond single-task needs like classification or segmentation. Inspired by this, we introduce Theia, a vision foundation model for robot learning that distills multiple off-the-shelf vision foundation models trained on varied vision tasks. Theia's rich visual representations encode diverse visual knowledge, enhancing downstream robot learning. Extensive experiments demonstrate that Theia outperforms its teacher models and prior robot learning models using less training data and smaller model sizes. Additionally, we quantify the quality of pre-trained visual representations and hypothesize that higher entropy in feature norm distributions leads to improved robot learning performance. Code, models, and demo are available at https://theia.theaiinstitute.com.

Language-based robot control is a powerful and versatile method to control a robot manipulator where large language models (LLMs) are used to reason about the environment. However, the generated robot motions by these controllers often lack safety and performance, resulting in jerky movements. In this work, a novel modular framework for zero-shot motion planning for manipulation tasks is developed. The modular components do not require any motion-planning-specific training. An LLM is combined with a vision model to create Python code that interacts with a novel path planner, which creates a piecewise linear reference path with bounds around the path that ensure safety. An optimization-based planner, the BoundMPC framework, is utilized to execute optimal, safe, and collision-free trajectories along the reference path. The effectiveness of the approach is shown on various everyday manipulation tasks in simulation and experiment, shown in the video at www.acin.tuwien.ac.at/42d2.

The current state-of-the-art in quadruped locomotion is able to produce a variety of complex motions. These methods either rely on switching between a discrete set of skills or learn a distribution across gaits using complex black-box models. Alternatively, we present Gaitor, which learns a disentangled and 2D representation across locomotion gaits. This learnt representation forms a planning space for closed-loop control delivering continuous gait transitions and perceptive terrain traversal. Gaitor’s latent space is readily interpretable and we discover that during gait transitions, novel unseen gaits emerge. The latent space is disentangled with respect to footswing heights and lengths. This means that these gait characteristics can be varied independently in the 2D latent representation. Together with a simple terrain encoding and a learnt planner operating in the latent space, Gaitor can take motion commands including desired gait type and swing characteristics all while reacting to uneven terrain. We evaluate Gaitor in both simulation and the real world on the ANYmal C platform. To the best of our knowledge, this is the first work learning a unified and interpretable latent space for multiple gaits, resulting in continuous blending between different locomotion modes on a real quadruped robot. An overview of the methods and results in this paper is found at https://youtu.be/eVFQbRyilCA.

Contact-rich manipulation plays an important role in everyday life, but uncertain parameters pose significant challenges to model-based planning and control. To address this issue, domain adaptation and domain randomization have been proposed to learn robust policies. However, they either lose the generalization ability to diverse instances or perform conservatively due to neglecting instance-specific information. In this paper, we propose a bi-level approach to learn robust manipulation primitives, including parameter-augmented policy learning using multiple models with tensor approximation, and parameter-conditioned policy retrieval through domain contraction. This approach unifies domain randomization and domain adaptation, providing optimal behaviors while keeping generalization ability. We validate the proposed method on three contact-rich manipulation primitives: hitting, pushing, and reorientation. The experimental results showcase the superior performance of our approach in generating robust policies for instances with diverse physical parameters.

Data collection in imitation learning often requires significant, laborious human supervision, such as numerous demonstrations, and/or frequent environment resets for methods that incorporate reinforcement learning. In this work, we propose an alternative approach, MILES: a fully autonomous, self-supervised data collection paradigm, and we show that this enables efficient policy learning from just a single demonstration and a single environment reset. MILES autonomously learns a policy for returning to and then following the single demonstration, whilst being self-guided during data collection, eliminating the need for additional human interventions. We evaluated MILES across several realworld tasks, including tasks that require precise contact-rich manipulation such as locking a lock with a key. We found that, under the constraints of a single demonstration and no repeated environment resetting, MILES significantly outperforms state-of-the-art alternatives like imitation learning methods that leverage reinforcement learning. Videos of our experiments and code can be found on our webpage: www.robot-learning.uk/miles.

Control tuning and adaptation present a significant challenge to the usage of robots in diverse environments. It is often nontrivial to find a single set of control parameters by hand that work well across the broad array of environments and conditions that a robot might encounter. Automated adaptation approaches must utilize prior knowledge about the system while adapting to significant domain shifts to find new control parameters quickly. In this work, we present a general framework for online controller adaptation that deals with these challenges. We combine meta-learning with Bayesian recursive estimation to learn prior predictive models of system performance that quickly adapt to online data, even when there is significant domain shift. These predictive models can be used as cost functions within efficient sampling-based optimization routines to find new control parameters online that maximize system performance. Our framework is powerful and flexible enough to adapt controllers for four diverse systems: a simulated race car, a simulated quadrupedal robot, and a simulated and physical quadrotor.

Abstract: Human actions manipulating articulated objects, such as opening and closing a drawer, can be categorized into multiple modalities we define as interaction modes. Traditional robot learning approaches lack discrete representations of these modes, which are crucial for empirical sampling and grounding. In this paper, we present ActAIM2, which learns a discrete representation of robot manipulation interaction modes in a purely unsupervised fashion, without the use of expert labels or simulator-based privileged information. Utilizing novel data collection methods involving simulator rollouts, ActAIM2 consists of an interaction mode selector and a low-level action predictor. The selector generates discrete representations of potential interaction modes with self-supervision, while the predictor outputs corresponding action trajectories. Our method is validated through its success rate in manipulating articulated objects and its robustness in sampling meaningful actions from the discrete representation. Extensive experiments demonstrate ActAIM2’s effectiveness in enhancing manipulability and generalizability over baselines and ablation studies. For videos and additional results, see our website: https://actaim2.github.io/.

Model-free reinforcement learning has emerged as a powerful method for developing robust robot control policies capable of navigating through complex and unstructured terrains. The effectiveness of these methods hinges on two essential elements: (1) the use of massively parallel physics simulations to expedite policy training, and (2) the deployment of an environment generator tasked with crafting terrains that are sufficiently challenging yet attainable, thereby facilitating continuous policy improvement. Existing methods of environment generation often rely on heuristics constrained by a set of parameters, limiting the diversity and realism. In this work, we introduce the Adaptive Diffusion Terrain Generator (ADTG), a novel method that leverages Denoising Diffusion Probabilistic Models (DDPMs) to dynamically expand an existing training environment by adding more diverse and complex terrains tailored to the current policy. Unlike conventional methods, ADTG adapts the terrain complexity and variety based on the evolving capabilities of the current policy. This is achieved through two primary mechanisms: First, by blending terrains from the initial dataset within their latent spaces using performance-informed weights, ADTG creates terrains that suitably challenge the policy. Secondly, by manipulating the initial noise in the diffusion process, ADTG seamlessly shifts between creating similar terrains for fine-tuning the current policy and entirely novel ones for expanding training diversity. Our experiments show that the policy trained by ADTG outperforms both procedural generated and natural environments, along with popular navigation methods.

In this work, we introduce general purpose touch representations for the increasingly accessible class of vision-based tactile sensors. Such sensors have led to many recent advances in robot manipulation as they markedly complement vision, yet solutions today often rely on task and sensor specific handcrafted perception models. Collecting real data at scale with task centric ground truth labels, like contact forces and slip, is a challenge further compounded by sensors of various form factor differing in aspects like lighting and gel markings. To tackle this, we turn to self-supervised learning (SSL) that has demonstrated remarkable performance in computer vision. We present Sparsh, a family of SSL models that can support various vision-based tactile sensors, alleviating the need for custom labels through pre-training on 460k+ tactile images with masking and self-distillation in pixel and latent spaces. We also build TacBench, to facilitate standardized benchmarking across sensors and models, comprising of six tasks ranging from comprehending tactile properties to enabling physical perception and manipulation planning. In evaluations, we find that SSL pre-training for touch representation outperforms task and sensor-specific end-to-end training by 95.1% on average over TacBench, and Sparsh (DINO) and Sparsh (IJEPA) are the most competitive, indicating the merits of learning in latent space for tactile images. Project page: https://sparsh-ssl.github.io

This paper presents a novel learning-based control framework that uses keyframing to incorporate high-level objectives in natural locomotion for legged robots. These high-level objectives are specified as a variable number of partial or complete pose targets that are spaced arbitrarily in time. Our proposed framework utilizes a multi-critic reinforcement learning algorithm to effectively handle the mixture of dense and sparse rewards. Additionally, it employs a transformer-based encoder to accommodate a variable number of input targets, each associated with specific time-to-arrivals. Throughout simulation and hardware experiments, we demonstrate that our framework can effectively satisfy the target keyframe sequence at the required times. In the experiments, the multi-critic method significantly reduces the effort of hyperparameter tuning compared to the standard single-critic alternative. Moreover, the proposed transformer-based architecture enables robots to anticipate future goals, which results in quantitative improvements in their ability to reach their targets.

Large Vision-Language-Action (VLA) models, leveraging powerful pre-trained Vision-Language Models (VLMs) backends, have shown promise in robotic control due to their impressive generalization ability. However, the success comes at a cost. Their reliance on VLM backends with billions of parameters leads to high computational costs and inference latency, limiting the testing scenarios to mainly quasi-static tasks and hindering performance in dynamic tasks requiring rapid interactions. To address these limitations, this paper proposes \textbf{HiRT}, a \textbf{Hi}erarchical \textbf{R}obot \textbf{T}ransformer framework that enables flexible frequency and performance trade-off. HiRT keeps VLMs running at low frequencies to capture temporarily invariant features while enabling real-time interaction through a high-frequency vision-based policy guided by the slowly updated features. Experiment results in both simulation and real-world settings demonstrate significant improvements over baseline methods. Empirically, we achieve a 58\% reduction in inference time delay while maintaining comparable success rates. Additionally, on novel dynamic manipulation benchmarks which are challenging for previous VLA models, HiRT improves the success rate from 48% to 75%.

Audio signals provide rich information for the robot interaction and object properties through contact. These information can surprisingly ease the learning of contact-rich robot manipulation skills, especially when the visual information alone is ambiguous or incomplete. However, the usage of audio data in robot manipulation has been constrained to teleoperated demonstrations collected by either attaching a microphone to the robot or object, which significantly limits its usage in robot learning pipelines. In this work, we introduce ManiWAV: an 'ear-in-hand' data collection device to collect in-the-wild human demonstrations with synchronous audio and visual feedback, and a corresponding policy interface to learn robot manipulation policy directly from the demonstrations. We demonstrate the capabilities of our system through four contact-rich manipulation tasks that require either passively sensing the contact events and modes, or actively sensing the object surface materials and states. In addition, we show that our system can generalize to unseen in-the-wild environments, by learning from diverse in-the-wild human demonstrations. All data, code, and policy will be public.

Imitation learning from human motion capture (MoCap) data provides a promising way to train humanoid robots. However, due to differences in morphology, such as varying degrees of joint freedom and force limits, exact replication of human behaviors may not be feasible for humanoid robots. Consequently, incorporating physically infeasible MoCap data in training datasets can adversely affect the performance of the robot policy. To address this issue, we propose a bi-level optimization-based imitation learning framework that alternates between optimizing both the robot policy and the target MoCap data. Specifically, we first develop a generative latent dynamics model using a novel self-consistent auto-encoder, which learns sparse and structured motion representations while capturing desired motion patterns in the dataset. The dynamics model is then utilized to generate reference motions while the latent representation regularizes the bi-level motion imitation process. Simulations conducted with a realistic model of a humanoid robot demonstrate that our method enhances the robot policy by modifying reference motions to be physically consistent.

Recent work has shown diffusion models are an effective approach to learning the multimodal distributions arising from demonstration data in behavior cloning. However, a drawback of this approach is the need to learn a denoising function, which is significantly more complex than learning an explicit policy. In this work, we propose Equivariant Diffusion Policy, a novel diffusion policy learning method that leverages domain symmetries to obtain better sample efficiency and generalization in the denoising function. We theoretically analyze the $\mathrm{SO}(2)$ symmetry of full 6-DoF control and characterize when a diffusion model is $\mathrm{SO}(2)$-equivariant. We furthermore evaluate the method empirically on a set of 12 simulation tasks in MimicGen, and show that it obtains a success rate that is, on average, 21.9\% higher than the baseline Diffusion Policy. We also evaluate the method on a real-world system to show that effective policies can be learned with relatively few training samples, whereas the baseline Diffusion Policy cannot.

Learning from expert demonstrations is a popular approach to train robotic manipulation policies from limited data. However, imitation learning algorithms require a number of design choices ranging from the input modality, training objective, and 6-DoF end-effector pose representation. Diffusion-based methods have gained popularity as they allow to predict long horizon trajectories and handle multimodal action distributions. Recently, Conditional Flow Matching (CFM) (or Rectified Flow) has been proposed as a more flexible generalization of diffusion models. In this paper we investigate the application of CFM in the context of robotic policy learning, and specifically study the interplay with the other design choices required to build an imitation learning algorithm. We show that CFM gives the best performance when combined with point cloud input observations. Additionally, we study the feasibility of a CFM formulation on the SO(3) manifold and evaluate its suitability with a simplified example. We perform extensive experiments on RLBench which demonstrate that our proposed PointFlowMatch approach achieves a state-of-the-art average success rate of 67.8% over eight tasks, double the performance of the next best method.