CoRL 2024 Papers

Learning-based control policies are widely used in various tasks in the field of robotics and control. However, formal (Lyapunov) stability guarantees for learning-based controllers with nonlinear dynamical systems are challenging to obtain. We propose a novel control approach, namely Control with Patterns (CWP), to address the stability issue over data sets corresponding to nonlinear dynamical systems. For data sets of this kind, we introduce a new definition, namely exponential attraction on data sets, to describe nonlinear dynamical systems under consideration. The problem of exponential attraction on data sets is converted to a pattern classification one based on the data sets and parameterized Lyapunov functions. Furthermore, D-learning is proposed as a method for performing CWP without knowledge of the system dynamics. Finally, the effectiveness of CWP based on D-learning is demonstrated through simulations and real flight experiments. In these experiments, the position of the multicopter is stabilized using only real-time images as feedback, which can be considered as an Image-Based Visual Servoing (IBVS) problem.

Learned language-conditioned robot policies often struggle to effectively adapt to new real-world tasks even when pre-trained across a diverse set of instructions. We propose a novel approach for few-shot adaptation to unseen tasks that exploits the semantic understanding of task decomposition provided by vision-language models (VLMs). Our method, Policy Adaptation via Language Optimization (PALO), combines a handful of demonstrations of a task with proposed language decompositions sampled from a VLM to quickly enable rapid nonparametric adaptation, avoiding the need for a larger fine-tuning dataset. We evaluate PALO on extensive real-world experiments consisting of challenging unseen, long-horizon robot manipulation tasks. We find that PALO is able of consistently complete long-horizon, multi-tier tasks in the real world, outperforming state of the art pre-trained generalist policies, and methods that have access to the same demonstrations.

We present High-Density Visual Particle Dynamics (HD-VPD), a learned world model that can emulate the physical dynamics of real scenes by processing massive latent point clouds containing 100K+ particles. To enable efficiency at this scale, we introduce a novel family of Point Cloud Transformers (PCTs) called Interlacers leveraging intertwined linear-attention Performer layers and graph-based neighbour attention layers. We demonstrate the capabilities of HD-VPD by modeling the dynamics of high degree-of-freedom bi-manual robots with two RGB-D cameras. Compared to the previous graph neural network approach, our Interlacer dynamics is twice as fast with the same prediction quality, and can achieve higher quality using 4x as many particles. We illustrate how HD-VPD can evaluate motion plan quality with robotic box pushing and can grasping tasks. See videos and particle dynamics rendered by HD-VPD at https://sites.google.com/view/hd-vpd.

The ability for robots to perform efficient and zero-shot grasping of object parts is crucial for practical applications and is becoming prevalent with recent advances in Vision-Language Models (VLMs). To bridge the 2D-to-3D gap for representations to support such a capability, existing methods rely on neural fields (NeRFs) via differentiable rendering or point-based projection methods. However, we demonstrate that NeRFs are inappropriate for scene changes due to its implicitness and point-based methods are inaccurate for part localization without rendering-based optimization. To amend these issues, we propose GraspSplats. Using depth supervision and a novel reference feature computation method, GraspSplats can generate high-quality scene representations under 60 seconds. We further validate the advantages of Gaussian-based representation by showing that the explicit and optimized geometry in GraspSplats is sufficient to natively support (1) real-time grasp sampling and (2) dynamic and articulated object manipulation with point trackers. With extensive experiments on a Franka robot, we demonstrate that GraspSplats significantly outperforms existing methods under diverse task settings. In particular, GraspSplats outperforms NeRF-based methods like F3RM and LERF-TOGO, and 2D detection methods. The code will be released.

Learning to plan for multi-step, multi-manipulator tasks is notoriously difficult because of the large search space and the complex constraint satisfaction problems. We present Generative Factor Chaining (GFC), a composable generative model for planning. GFC represents a planning problem as a spatial-temporal factor graph, where nodes represent objects and robots in the scene, spatial factors capture the distributions of valid relationships among nodes, and temporal factors represent the distributions of skill transitions. Each factor is implemented as a modular diffusion model, which are composed during inference to generate feasible long-horizon plans through bi-directional message passing. We show that GFC can solve complex bimanual manipulation tasks and exhibits strong generalization to unseen planning tasks with novel combinations of objects and constraints. More details can be found at: https://sites.google.com/view/generative-factor-chaining

This paper introduces SoftManiSim, a novel simulation framework for multi-segment continuum manipulators. Existing continuum robot simulators often rely on simplifying assumptions, such as constant curvature bending or ignoring contact forces, to meet real-time simulation and training demands. To bridge this gap, we propose a robust and rapid mathematical model for continuum robots at the core of SoftManiSim, ensuring precise and adaptable simulations. The framework can integrate with various rigid-body robots, increasing its utility across different robotic platforms. SoftManiSim supports parallel operations for simultaneous simulations of multiple robots and generates synthetic data essential for training deep reinforcement learning models. This capability enhances the development and optimization of control strategies in dynamic environments. Extensive simulations validate the framework's effectiveness, demonstrating its capabilities in handling complex robotic interactions and tasks. We also present real robot validation to showcase the simulator's practical applicability and accuracy in real-world settings. To our knowledge, SoftManiSim is the first open-source real-time simulator capable of modeling continuum robot behavior under dynamic point/distributed loading. It enables rapid deployment in reinforcement learning and machine learning applications. This simulation framework can be downloaded from https://github.com/MohammadKasaei/SoftManiSim.

Manipulation without grasping, known as non-prehensile manipulation, is essential for dexterous robots in contact-rich environments, but presents many challenges relating with underactuation, hybrid-dynamics, and frictional uncertainty. Additionally, object occlusions in a scenario of contact uncertainty and where the motion of the object evolves independently from the robot becomes a critical problem, which previous literature fails to address. We present a method for learning visuotactile state estimators and uncertainty-aware control policies for non-prehensile manipulation under occlusions, by leveraging diverse interaction data from privileged policies trained in simulation. We formulate the estimator within a Bayesian deep learning framework, to model its uncertainty, and then train uncertainty-aware control policies by incorporating the pre-learned estimator into the reinforcement learning (RL) loop, both of which lead to significantly improved estimator and policy performance. Therefore, unlike prior non-prehensile research that relies on complex external perception set-ups, our method successfully handles occlusions after sim-to-real transfer to robotic hardware with a simple onboard camera.

We present OmniH2O (Omni Human-to-Humanoid), a learning-based system for whole-body humanoid teleoperation and autonomy. Using kinematic pose as a universal control interface, OmniH2O enables various ways for a human to control a full-sized humanoid with dexterous hands, including using real-time teleoperation through VR headset, verbal instruction, and RGB camera. OmniH2O also enables full autonomy by learning from teleoperated demonstrations or integrating with frontier models such as GPT-4. OmniH2O demonstrates versatility and dexterity in various real-world whole-body tasks through teleoperation or autonomy, such as playing multiple sports, moving and manipulating objects, and interacting with humans. We develop an RL-based sim-to-real pipeline, which involves large-scale retargeting and augmentation of human motion datasets, learning a real-world deployable policy with sparse sensor input by imitating a privileged teacher policy, and reward designs to enhance robustness and stability. We release the first humanoid whole-body control dataset, OmniH2O-6, containing six everyday tasks, and demonstrate humanoid whole-body skill learning from teleoperated datasets. Videos at the anonymous website [https://anonymous-omni-h2o.github.io/](https://anonymous-omni-h2o.github.io/)

We address the challenge of acquiring real-world manipulation skills with a scalable framework. We hold the belief that identifying an appropriate prediction target capable of leveraging large-scale datasets is crucial for achieving efficient and universal learning. Therefore, we propose to utilize 3D flow, which represents the future trajectories of 3D points on objects of interest, as an ideal prediction target. To exploit scalable data resources, we turn our attention to human videos. We develop, for the first time, a language-conditioned 3D flow prediction model directly from large-scale RGBD human video datasets. Our predicted flow offers actionable guidance, thus facilitating zero-shot skill transfer in real-world scenarios. We deploy our method with a policy based on closed-loop flow prediction. Remarkably, without any in-domain finetuning, our method achieves an impressive 81\% success rate in zero-shot human-to-robot skill transfer, covering 18 tasks in 6 scenes. Our framework features the following benefits: (1) scalability: leveraging cross-embodiment data resources; (2) wide application: multiple object categories, including rigid, articulated, and soft bodies; (3) stable skill transfer: providing actionable guidance with a small inference domain-gap.

Offline learning at scale has led to breakthroughs in computer vision, natural language processing, and robotic manipulation domains. However, scaling up learning for legged robot locomotion, especially with multiple skills in a single policy, presents significant challenges for prior online reinforcement learning (RL) methods. To address this challenge, we propose DiffuseLoco, a novel, scalable framework that leverages diffusion models to directly learn from offline multimodal datasets with a diverse set of locomotion skills. With design choices tailored for real-time control in dynamical systems, including receding horizon control and delayed inputs, DiffuseLoco is capable of reproducing multimodality in performing various locomotion skills, zero-shot transferred to real quadruped robots and deployed on edge computes. Through extensive real-world benchmarking, DiffuseLoco exhibits better stability and velocity tracking performance compared to prior RL and non-diffusion-based behavior cloning baselines. This work opens new possibilities for scaling up learning-based legged locomotion control through the scaling of large, expressive models and diverse offline datasets.

Home robots performing personalized tasks must adeptly balance user preferences with environmental affordances. We focus on organization tasks within constrained spaces, such as arranging items into a refrigerator, where preferences for placement collide with physical limitations. The robot must infer user preferences based on a small set of demonstrations, which is easier for users to provide than extensively defining all their requirements. While recent works use Large Language Models (LLMs) to learn preferences from user demonstrations, they encounter two fundamental challenges. First, there is inherent ambiguity in interpreting user actions, as multiple preferences can often explain a single observed behavior. Second, not all user preferences are practically feasible due to geometric constraints in the environment. To address these challenges, we introduce APRICOT, a novel approach that merges LLM-based Bayesian active preference learning with constraint-aware task planning. APRICOT refines its generated preferences by actively querying the user and dynamically adapts its plan to respect environmental constraints. We evaluate APRICOT on a dataset of diverse organization tasks and demonstrate its effectiveness in real-world scenarios, showing significant improvements in both preference satisfaction and plan feasibility.

Enabling robots to autonomously perform hybrid motions in diverse environments can be beneficial for long-horizon tasks such as material handling, household chores, and work assistance. This requires extensive exploitation of intrinsic motion capabilities, extraction of affordances from rich environmental information, and planning of physical interaction behaviors. Despite recent progress has demonstrated impressive humanoid whole-body control abilities, they struggle to achieve versatility and adaptability for new tasks. In this work, we propose HYPERmotion, a framework that learns, selects and plans behaviors based on tasks in different scenarios. We combine reinforcement learning with whole-body optimization to generate motion for 38 actuated joints and create a motion library to store the learned skills. We apply the planning and reasoning features of the large language models (LLMs) to complex loco-manipulation tasks, constructing a hierarchical task graph that comprises a series of primitive behaviors to bridge lower-level execution with higher-level planning. By leveraging the interaction of distilled spatial geometry and 2D observation with a visual language model (VLM) to ground knowledge into a robotic morphology selector to choose appropriate actions in single- or dual-arm, legged or wheeled locomotion. Experiments in simulation and real-world show that learned motions can efficiently adapt to new tasks, demonstrating high autonomy from free-text commands in unstructured scenes. Videos and website: hy-motion.github.io//

Vision Language Models (VLMs) for robotics have received significant attention in recent years. As a VLM can understand robot observations and perform complex visual reasoning, it is regarded as a potential universal solution for general robotics challenges such as manipulation and navigation. However, previous robotics VLMs such as RT-1, RT-2, and ManipLLM have focused on directly learning robot actions. Such approaches require collecting a significant amount of robot interaction data, which is extremely costly in the real world. Thus, we propose A3VLM, an object-centric, actionable, articulation-aware vision language model. A3VLM focuses on the articulation structure and action affordances of objects. Its representation is robot-agnostic and can be translated into robot actions using simple action primitives. Extensive experiments in both simulation benchmarks and real-world settings demonstrate the effectiveness and stability of A3VLM.

Vision foundation models trained on massive amounts of visual data have shown unprecedented reasoning and planning skills in open-world settings. A key challenge in applying them to robotic tasks is the modality gap between visual data and action data. We introduce differentiable robot rendering, a method allowing the visual appearance of a robot body to be directly differentiable with respect to its control parameters. Our model integrates a kinematics-aware deformable model and Gaussians Splatting and is compatible with any robot form factors and degrees of freedom. We demonstrate its capability and usage in applications including reconstruction of robot poses from images and controlling robots through vision language models. Quantitative and qualitative results show that our differentiable rendering model provides effective gradients for robotic control directly from pixels, setting the foundation for the future applications of vision foundation models in robotics.

Learning in simulation and transferring the learned policy to the real world has the potential to enable generalist robots. The key challenge of this approach is to address simulation-to-reality (sim-to-real) gaps. Previous methods often require domain-specific knowledge *a priori*. We argue that a straightforward way to obtain such knowledge is by asking humans to observe and assist robot policy execution in the real world. The robots can then learn from humans to close various sim-to-real gaps. We propose TRANSIC, a data-driven approach to enable successful sim-to-real transfer based on a human-in-the-loop framework. TRANSIC allows humans to augment simulation policies to overcome various unmodeled sim-to-real gaps holistically through intervention and online correction. Residual policies can be learned from human corrections and integrated with simulation policies for autonomous execution. We show that our approach can achieve successful sim-to-real transfer in complex and contact-rich manipulation tasks such as furniture assembly. Through synergistic integration of policies learned in simulation and from humans, TRANSIC is effective as a holistic approach to addressing various, often coexisting sim-to-real gaps. It displays attractive properties such as scaling with human effort. Videos and code are available at https://transic-robot.github.io/.

We present InterACT: Inter-dependency aware Action Chunking with Hierarchical Attention Transformers, a novel imitation learning framework for bimanual manipulation that integrates hierarchical attention to capture inter-dependencies between dual-arm joint states and visual inputs. InterACT consists of a Hierarchical Attention Encoder and a Multi-arm Decoder, both designed to enhance information aggregation and coordination. The encoder processes multi-modal inputs through segment-wise and cross-segment attention mechanisms, while the decoder leverages synchronization blocks to refine individual action predictions, providing the counterpart's prediction as context. Our experiments on a variety of simulated and real-world bimanual manipulation tasks demonstrate that InterACT significantly outperforms existing methods. Detailed ablation studies validate the contributions of key components of our work, including the impact of CLS tokens, cross-segment encoders, and synchronization blocks.

In this work, we address the challenging problem of long-horizon goal-reaching policy learning from non-expert, action-free observation data. Unlike fully labeled expert data, our data is more accessible and avoids the costly process of action labeling. Additionally, compared to online learning, which often involves aimless exploration, our data provides useful guidance for more efficient exploration. To achieve our goal, we propose a novel subgoal guidance learning strategy. The motivation behind this strategy is that long-horizon goals offer limited guidance for efficient exploration and accurate state transition. We develop a diffusion strategy-based high-level policy to generate reasonable subgoals as waypoints, preferring states that more easily lead to the final goal. Additionally, we learn state-goal value functions to encourage efficient subgoal reaching. These two components naturally integrate into the off-policy actor-critic framework, enabling efficient goal attainment through informative exploration. We evaluate our method on complex robotic navigation and manipulation tasks, demonstrating a significant performance advantage over existing methods. Our ablation study further shows that our method is robust to observation data with various corruptions.

In this paper, we reimagine volumetric representations through the lens of quadrics. We posit that rigid scene components can be effectively decomposed into quadric surfaces. Leveraging this assumption, we reshape the volumetric representations with million of cubes by several quadric planes, which results in more accurate and efficient modeling of 3D scenes in SLAM contexts. First, we use the quadric assumption to rectify noisy depth estimations from RGB inputs. This step significantly improves depth estimation accuracy, and allows us to efficiently sample ray points around quadric planes instead of the entire volume space in previous NeRF-SLAM systems. Second, we introduce a novel quadric-decomposed transformer to aggregate information across quadrics. The quadric semantics are not only explicitly used for depth correction and scene decomposition, but also serve as an implicit supervision signal for the mapping network. Through rigorous experimental evaluation, our method exhibits superior performance over other approaches relying on estimated depth, and achieves comparable accuracy to methods utilizing ground truth depth on both synthetic and real-world datasets.

We introduce ClutterGen, a physically compliant simulation scene generator capable of producing highly diverse, cluttered, and stable scenes for robot learning. Generating such scenes is challenging as each object must adhere to physical laws like gravity and collision. As the number of objects increases, finding valid poses becomes more difficult, necessitating significant human engineering effort, which limits the diversity of the scenes. To overcome these challenges, we propose a reinforcement learning method that can be trained with physics-based reward signals provided by the simulator. Our experiments demonstrate that ClutterGen can generate cluttered object layouts with up to ten objects on confined table surfaces. Additionally, our policy design explicitly encourages the diversity of the generated scenes for open-ended generation. Our real-world robot results show that ClutterGen can be directly used for clutter rearrangement and stable placement policy training.