CoRL 2024 论文列表

Humanoid robots, with their human-like embodiment, have the potential to integrate seamlessly into human environments. Critical to their coexistence and cooperation with humans is the ability to understand natural language communications and exhibit human-like behaviors. This work focuses on generating diverse whole-body motions for humanoid robots from language descriptions. We leverage human motion priors from extensive human motion datasets to initialize humanoid motions and employ the commonsense reasoning capabilities of Vision Language Models (VLMs) to edit and refine these motions. Our approach demonstrates the capability to produce natural, expressive, and text-aligned humanoid motions, validated through both simulated and real-world experiments. More videos can be found on our website https://ut-austin-rpl.github.io/Harmon/.

We study the problem of teaching humanoid robots manipulation skills by imitating from single video demonstrations. We introduce OKAMI, a method that generates a manipulation plan from a single RGB-D video and derives a policy for execution. At the heart of our approach is object-aware retargeting, which enables the humanoid robot to mimic the human motions in an RGB-D video while adjusting to different object locations during deployment. OKAMI uses open-world vision models to identify task-relevant objects and retarget the body motions and hand poses separately. Our experiments show that OKAMI achieves strong generalizations across varying visual and spatial conditions, outperforming the state-of-the-art baseline on open-world imitation from observation. Furthermore, OKAMI rollout trajectories are leveraged to train closed-loop visuomotor policies, which achieve an average success rate of $79.2\%$ without the need for labor-intensive teleoperation. More videos can be found on our website https://ut-austin-rpl.github.io/OKAMI/.

We introduce the novel task of interactive scene exploration, wherein robots autonomously explore environments and produce an action-conditioned scene graph (ACSG) that captures the structure of the underlying environment. The ACSG accounts for both low-level information (geometry and semantics) and high-level information (action-conditioned relationships between different entities) in the scene. To this end, we present the Robotic Exploration (RoboEXP) system, which incorporates the Large Multimodal Model (LMM) and an explicit memory design to enhance our system's capabilities. The robot reasons about what and how to explore an object, accumulating new information through the interaction process and incrementally constructing the ACSG. Leveraging the constructed ACSG, we illustrate the effectiveness and efficiency of our RoboEXP system in facilitating a wide range of real-world manipulation tasks involving rigid, articulated objects, nested objects, and deformable objects. Project Page: https://jianghanxiao.github.io/roboexp-web/

Fetching, which includes approaching, grasping, and retrieving, is a critical challenge for robot manipulation tasks. Existing methods primarily focus on table-top scenarios, which do not adequately capture the complexities of environments where both grasping and planning are essential. To address this gap, we propose a new benchmark FetchBench, featuring diverse procedural scenes that integrate both grasping and motion planning challenges. Additionally, FetchBench includes a data generation pipeline that collects successful fetch trajectories for use in imitation learning methods. We implement multiple baselines from the traditional sense-plan-act pipeline to end-to-end behavior models. Our empirical analysis reveals that these methods achieve a maximum success rate of only 20%, indicating substantial room for improvement. Additionally, we identify key bottlenecks within the sense-plan-act pipeline and make recommendations based on the systematic analysis.

Estimating visual-tactile models of deformable objects is challenging because vision suffers from occlusion, while touch data is sparse and noisy. We propose a novel data-efficient method for dense heterogeneous model estimation by leveraging experience from diverse training objects. The method is based on Bayesian Meta-Learning (BML), which can mitigate overfitting high-capacity visual-tactile models by meta-learning an informed prior and naturally achieves few-shot online estimation via posterior estimation. However, BML requires a shared parametric model across tasks but visual-tactile models for diverse objects have different parameter spaces. To address this issue, we introduce Structured Bayesian Meta-Learning (SBML) that incorporates heterogeneous physics models, enabling learning from training objects with varying appearances and geometries. SBML performs zero-shot vision-only prediction of deformable model parameters and few-shot adaptation after a handful of touches. Experiments show that in two classes of heterogeneous objects, namely plants and shoes, SBML outperforms existing approaches in force and torque prediction accuracy in zero- and few-shot settings.

The tie-knotting task is highly challenging due to the tie's high deformation and long-horizon manipulation actions. This work presents TieBot, a Real-to-Sim-to-Real learning from visual demonstration system for the robots to learn to knot a tie. We introduce the Hierarchical Feature Matching approach to estimate a sequence of tie's meshes from the demonstration video. With these estimated meshes used as subgoals, we first learn a teacher policy using privileged information. Then, we learn a student policy with point cloud observation by imitating teacher policy. Lastly, our pipeline applies learned policy to real-world execution. We demonstrate the effectiveness of TieBot in simulation and the real world. In the real-world experiment, a dual-arm robot successfully knots a tie, achieving 50% success rate among 10 trials. Videos can be found on https://tiebots.github.io/.

Aerial manipulation combines the versatility and speed of flying platforms with the functional capabilities of mobile manipulation, which presents significant challenges due to the need for precise localization and control. Traditionally, researchers have relied on off-board perception systems, which are limited to expensive and impractical specially equipped indoor environments. In this work, we introduce a novel platform for autonomous aerial manipulation that exclusively utilizes onboard perception systems. Our platform can perform aerial manipulation in various indoor and outdoor environments without depending on external perception systems. Our experimental results demonstrate the platform's ability to autonomously grasp various objects in diverse settings. This advancement significantly improves the scalability and practicality of aerial manipulation applications by eliminating the need for costly tracking solutions. To accelerate future research, we open source our modern ROS 2 software stack and custom hardware design, making our contributions accessible to the broader research community.

Towel manipulation is a crucial step towards more general cloth manipulation. However, folding a towel from an arbitrarily crumpled state and recovering from a failed folding step remain critical challenges in robotics. We propose joint-probability action inference JA-TN, as a way to improve TransporterNet's operational efficiency; to our knowledge, this is the first single data-driven policy to achieve various types of folding from most crumpled states. We present three benchmark domains with a set of shaping tasks and the corresponding oracle policies to facilitate the further development of the field. We also present a simulation-to-reality transfer procedure for vision-based deep learning controllers by processing and augmenting RGB and/or depth images. We also demonstrate JA-TN's ability to integrate with a real camera and a UR3e robot arm, showcasing the method's applicability to real-world tasks.

In-hand manipulation of pen-like objects is a most basic and important skill in our daily lives, as many tools such as hammers and screwdrivers are similarly shaped. However, current learning-based methods struggle with this task due to a lack of high-quality demonstrations and the significant gap between simulation and the real world. In this work, we push the boundaries of learning-based in-hand manipulation systems by demonstrating the capability to spin pen-like objects. We use reinforcement learning to train a policy and generate a high-fidelity trajectory dataset in simulation. This serves two purposes: 1) pre-training a sensorimotor policy in simulation; 2) conducting open-loop trajectory replay in the real world. We then fine-tune the sensorimotor policy using these real-world trajectories to adapt to the real world. With less than 50 trajectories, our policy learns to rotate more than ten pen-like objects with different physical properties for multiple revolutions. We present a comprehensive analysis of our design choices and share the lessons learned during development. Videos are shown on https://corl-2024-dexpen.github.io/.

The ability to reflect on and correct failures is crucial for robotic systems to interact stably with real-life objects. Observing the generalization and reasoning capabilities of Multimodal Large Language Models (MLLMs), previous approaches have aimed to utilize these models to enhance robotic systems accordingly. However, these methods typically focus on high-level planning corrections using an additional MLLM, with limited utilization of failed samples to correct low-level contact poses which is particularly prone to occur during articulated object manipulation. To address this gap, we propose an Autonomous Interactive Correction (AIC) MLLM, which makes use of previous low-level interaction experiences to correct SE(3) pose predictions for articulated object. Specifically, AIC MLLM is initially fine-tuned to acquire both pose prediction and feedback prompt comprehension abilities. We design two types of prompt instructions for interactions with objects: 1) visual masks to highlight unmovable parts for position correction, and 2) textual descriptions to indicate potential directions for rotation correction. During inference, a Feedback Information Extraction module is introduced to recognize the failure cause, allowing AIC MLLM to adaptively correct the pose prediction using the corresponding prompts. To further enhance manipulation stability, we devise a Test Time Adaptation strategy that enables AIC MLLM to better adapt to the current scene configuration. Finally, extensive experiments are conducted in both simulated and real-world environments to evaluate the proposed method. The results demonstrate that our AIC MLLM can efficiently correct failure samples by leveraging interaction experience prompts.

A key limitation of learned robot control policies is their inability to generalize outside their training data. Recent works on vision-language-action models (VLAs) have shown that the use of large, internet pre-trained vision-language models as the backbone of learned robot policies can substantially improve their robustness and generalization ability. Yet, one of the most exciting capabilities of large vision-language models in other domains is their ability to reason iteratively through complex problems. Can that same capability be brought into robotics to allow policies to improve performance by reasoning about a given task before acting? Naive use of "chain-of-thought" (CoT) style prompting is significantly less effective with standard VLAs because of the relatively simple training examples that are available to them. Additionally, purely semantic reasoning about sub-tasks, as is common in regular CoT, is insufficient for robot policies that need to ground their reasoning in sensory observations and the robot state. To this end, we introduce Embodied Chain-of-Thought Reasoning (ECoT) for VLAs, in which we train VLAs to perform multiple steps of reasoning about plans, sub-tasks, motions, and visually grounded features like object bounding boxes and end effector positions, before predicting the robot action. We design a scalable pipeline for generating synthetic training data for ECoT on large robot datasets. We demonstrate, that ECoT increases the absolute success rate of OpenVLA, the current strongest open-source VLA policy, by 28\% across challenging generalization tasks, without any additional robot training data. Additionally, ECoT makes it easier for humans to interpret a policy's failures and correct its behavior using natural language.

A pivotal challenge in robotics is achieving fast, safe, and robust dexterous grasping across a diverse range of objects, an important goal within industrial applications. However, existing methods often have very limited speed, dexterity, and generality, along with limited or no hardware safety guarantees. In this work, we introduce DextrAH-G, a depth-based dexterous grasping policy trained entirely in simulation that combines reinforcement learning, geometric fabrics, and teacher-student distillation. We address key challenges in joint arm-hand policy learning, such as high-dimensional observation and action spaces, the sim2real gap, collision avoidance, and hardware constraints. DextrAH-G enables a 23 motor arm-hand robot to safely and continuously grasp and transport a large variety of objects at high speed using multi-modal inputs including depth images, allowing generalization across object geometry. Videos at https://sites.google.com/view/dextrah-g.

Traditional reinforcement learning control for quadruped robots often relies on privileged information, demanding meticulous selection and precise estimation, thereby imposing constraints on the development process. This work proposes a Self-learning Latent Representation (SLR) method, which achieves high-performance control policy learning without the need for privileged information. To enhance the credibility of our proposed method's evaluation, SLR is compared with open-source code repositories of state-of-the-art algorithms, retaining the original authors' configuration parameters. Across four repositories, SLR consistently outperforms the reference results. Ultimately, the trained policy and encoder empower the quadruped robot to navigate steps, climb stairs, ascend rocks, and traverse various challenging terrains.

Legged robot locomotion on sand slopes is challenging due to the complex dynamics of granular media and how the lack of solid surfaces can hinder locomotion. A promising strategy, inspired by ghost crabs and other organisms in nature, is to strategically interact with rocks, debris, and other obstacles to facilitate movement. To provide legged robots with this ability, we present a novel approach that leverages avalanche dynamics to indirectly manipulate objects on a granular slope. We use a Vision Transformer (ViT) to process image representations of granular dynamics and robot excavation actions. The ViT predicts object movement, which we use to determine which leg excavation action to execute. We collect training data from 100 real physical trials and, at test time, deploy our trained model in novel settings. Experimental results suggest that our model can accurately predict object movements and achieve a success rate ≥ 80% in a variety of manipulation tasks with up to four obstacles, and can also generalize to objects with different physics properties. To our knowledge, this is the first paper to leverage granular media avalanche dynamics to indirectly manipulate objects on granular slopes. Supplementary material is available at https://sites.google.com/view/grain-corl2024/home.

In this study, we consider the problem of predicting task success for open-vocabulary manipulation by a manipulator, based on instruction sentences and egocentric images before and after manipulation. Conventional approaches, including multimodal large language models (MLLMs), often fail to appropriately understand detailed characteristics of objects and/or subtle changes in the position of objects. We propose Contrastive $\lambda$-Repformer, which predicts task success for table-top manipulation tasks by aligning images with instruction sentences. Our method integrates the following three key types of features into a multi-level aligned representation: features that preserve local image information; features aligned with natural language; and features structured through natural language. This allows the model to focus on important changes by looking at the differences in the representation between two images. We evaluate Contrastive $\lambda$-Repformer on a dataset based on a large-scale standard dataset, the RT-1 dataset, and on a physical robot platform. The results show that our approach outperformed existing approaches including MLLMs. Our best model achieved an improvement of 8.66 points in accuracy compared to the representative MLLM-based model.

Given the high cost of collecting robotic data in the real world, sample efficiency is a consistently compelling pursuit in robotics. In this paper, we introduce SGRv2, an imitation learning framework that enhances sample efficiency through improved visual and action representations. Central to the design of SGRv2 is the incorporation of a critical inductive bias—$\textit{action locality}$, which posits that robot's actions are predominantly influenced by the target object and its interactions with the local environment. Extensive experiments in both simulated and real-world settings demonstrate that action locality is essential for boosting sample efficiency. SGRv2 excels in RLBench tasks with keyframe control using merely 5 demonstrations and surpasses the RVT baseline in 23 of 26 tasks. Furthermore, when evaluated on ManiSkill2 and MimicGen using dense control, SGRv2's success rate is 2.54 times that of SGR. In real-world environments, with only eight demonstrations, SGRv2 can perform a variety of tasks at a markedly higher success rate compared to baseline models.

Multi-sensor fusion is crucial for accurate 3D object detection in autonomous driving, with cameras and LiDAR being the most commonly used sensors. However, existing methods perform sensor fusion in a single view by projecting features from both modalities either in Bird's Eye View (BEV) or Perspective View (PV), thus sacrificing complementary information such as height or geometric proportions. To address this limitation, we propose ProFusion3D, a progressive fusion framework that combines features in both BEV and PV at both intermediate and object query levels. Our architecture hierarchically fuses local and global features, enhancing the robustness of 3D object detection. Additionally, we introduce a self-supervised mask modeling pre-training strategy to improve multi-modal representation learning and data efficiency through three novel objectives. Extensive experiments on nuScenes and Argoverse2 datasets conclusively demonstrate the efficacy of ProFusion3D. Moreover, ProFusion3D is robust to sensor failure, showing strong performance when only one modality is available.

The ability to grasp objects in-the-wild from open-ended language instructions constitutes a fundamental challenge in robotics. An open-world grasping system should be able to combine high-level contextual with low-level physical-geometric reasoning in order to be applicable in arbitrary scenarios. Recent works exploit the web-scale knowledge inherent in large language models (LLMs) to plan and reason in robotic context, but rely on external vision and action models to ground such knowledge into the environment and parameterize actuation. This setup suffers from two major bottlenecks: a) the LLM's reasoning capacity is constrained by the quality of visual grounding, and b) LLMs do not contain low-level spatial understanding of the world, which is essential for grasping in contact-rich scenarios. In this work we demonstrate that modern vision-language models (VLMs) are capable of tackling such limitations, as they are implicitly grounded and can jointly reason about semantics and geometry. We propose \texttt{OWG}, an open-world grasping pipeline that combines VLMs with segmentation and grasp synthesis models to unlock grounded world understanding in three stages: open-ended referring segmentation, grounded grasp planning and grasp ranking via contact reasoning, all of which can be applied zero-shot via suitable visual prompting mechanisms. We conduct extensive evaluation in cluttered indoor scene datasets to showcase \texttt{OWG}'s robustness in grounding from open-ended language, as well as open-world robotic grasping experiments in both simulation and hardware that demonstrate superior performance compared to previous supervised and zero-shot LLM-based methods.

In recent years, instruction-tuned Large Multimodal Models (LMMs) have been successful at several tasks, including image captioning and visual question answering; yet leveraging these models remains an open question for robotics. Prior LMMs for robotics applications have been extensively trained on language and action data, but their ability to generalize in different settings has often been less than desired. To address this, we introduce LLARVA, a model trained with a novel instruction tuning method that leverages structured prompts to unify a range of robotic learning tasks, scenarios, and environments. Additionally, we show that predicting intermediate 2-D representations, which we refer to as *visual traces*, can help further align vision and action spaces for robot learning. We generate 8.5M image-visual trace pairs from the Open X-Embodiment dataset in order to pre-train our model, and we evaluate on 12 different tasks in the RLBench simulator as well as a physical Franka Emika Panda 7-DoF robot. Our experiments yield strong performance, demonstrating that LLARVA — using 2-D and language representations — performs well compared to several contemporary baselines, and can generalize across various robot environments and configurations.