CoRL 2024 Papers

This paper tackles the challenging robotic task of generalizable paper cutting using scissors. In this task, scissors attached to a robot arm are driven to accurately cut curves drawn on the paper, which is hung with the top edge fixed. Due to the frequent paper-scissor contact and consequent fracture, the paper features continual deformation and changing topology, which is diffult for accurate modeling.To deal with such versatile scenarios, we propose ScissorBot, the first learning-based system for robotic paper cutting with scissors via simulation, imitation learning and sim2real. Given the lack of sufficient data for this task, we build PaperCutting-Sim, a paper simulator supporting interactive fracture coupling with scissors, enabling demonstration generation with a heuristic-based oracle policy. To ensure effective execution, we customize an action primitive sequence for imitation learning to constrain its action space, thus alleviating potential compounding errors. Finally, by integrating sim-to-real techniques to bridge the gap between simulation and reality, our policy can be effectively deployed on the real robot. Experimental results demonstrate that our method surpasses all baselines in both simulation and real-world benchmarks and achives performance comparable to human operation with a single hand under the same conditions.

We present JointMotion, a self-supervised pre-training method for joint motion prediction in self-driving vehicles. Our method jointly optimizes a scene-level objective connecting motion and environments, and an instance-level objective to refine learned representations. Scene-level representations are learned via non-contrastive similarity learning of past motion sequences and environment context. At the instance level, we use masked autoencoding to refine multimodal polyline representations. We complement this with an adaptive pre-training decoder that enables JointMotion to generalize across different environment representations, fusion mechanisms, and dataset characteristics. Notably, our method reduces the joint final displacement error of Wayformer, HPTR, and Scene Transformer models by 3%, 8%, and 12%, respectively; and enables transfer learning between the Waymo Open Motion and the Argoverse 2 Motion Forecasting datasets.

In recent years, the transformer architecture has become the de-facto standard for machine learning algorithms applied to natural language processing and computer vision. Despite notable evidence of successful deployment of this architecture in the context of robot learning, we claim that vanilla transformers do not fully exploit the structure of the robot learning problem. We propose Body Transformer (BoT), an architecture that exploits the robot embodiment by providing an inductive bias that guides the learning process. We represent the robot body as a graph of sensors and actuators, and rely on masked attention to pool information through the architecture. The resulting architecture outperforms the vanilla transformer, as well as the classical multilayer perceptron, with respect to task completion, scaling properties, and computational efficiency when representing either imitation or reinforcement learning policies.

Many approaches to robot learning begin by inferring a reward function from a set of human demonstrations. To learn a good reward, it is necessary to determine which features of the environment are relevant before determining how these features should be used to compute reward. In particularly complex, high-dimensional environments, human demonstrators often struggle to fully specify their desired behavior from a small number of demonstrations. End-to-end reward learning methods (e.g., using deep networks or program synthesis techniques) often yield brittle reward functions that are sensitive to spurious state features. By contrast, humans can often generalizably learn from a small number of demonstrations by incorporating strong priors about what features of a demonstration are likely meaningful for a task of interest. How do we build robots that leverage this kind of background knowledge when learning from new demonstrations? This paper describes a method named ALGAE which alternates between using language models to iteratively identify human-meaningful features needed to explain demonstrated behavior, then standard inverse reinforcement learning techniques to assign weights to these features. Experiments across a variety of both simulated and real-world robot environments show that ALGAElearns generalizable reward functions defined on interpretable features using only small numbers of demonstrations. Importantly, ALGAE can recognize when features are missing, then extract and define those features without any human input -- making it possible to quickly and efficiently acquire rich representations of user behavior.

Humanoid robots are a key focus in robotics, with their capacity to navigate tough terrains being essential for many uses. While strides have been made, creating adaptable locomotion for complex environments is still tough. Recent progress in learning-based systems offers hope for robust legged locomotion, but challenges persist, such as tracking accuracy at high speeds and on uneven ground, and joint oscillations in actual robots. This paper proposes a novel training framework to address these challenges by employing a two-phase training paradigm with reinforcement learning. The proposed framework is further enhanced through the integration of command curriculum learning, refining the precision and adaptability of our approach. Additionally, we adapt DreamWaQ to our humanoid locomotion system and improve it to mitigate joint oscillations. Finally, we achieve the sim-to-real transfer of our method. A series of empirical results demonstrate the superior performance of our proposed method compared to state-of-the-art methods.

Legged navigation is typically examined within open-world, off-road, and challenging environments. In these scenarios, estimating external disturbances requires a complex synthesis of multi-modal information. This underlines a major limitation in existing works that primarily focus on avoiding obstacles. In this work, we propose TOP-Nav, a novel legged navigation framework that integrates a comprehensive path planner with Terrain awareness, Obstacle avoidance and close-loop Proprioception. TOP-Nav underscores the synergies between vision and proprioception in both path and motion planning. Within the path planner, we present a terrain estimator that enables the robot to select waypoints on terrains with higher traversability while effectively avoiding obstacles. In the motion planning level, we construct a proprioception advisor from the learning-based locomotion controller to provide motion evaluations for the path planner. Based on the close-loop motion feedback, we offer online corrections for the vision-based terrain and obstacle estimations. Consequently, TOP-Nav achieves open-world navigation that the robot can handle terrains or disturbances beyond the distribution of prior knowledge and overcomes constraints imposed by visual conditions. Building upon extensive experiments conducted in both simulation and real-world environments, TOP-Nav demonstrates superior performance in open-world navigation compared to existing methods.

Developing agents that can perform complex control tasks from high-dimensional observations is a core ability of autonomous agents that requires underlying robust task control policies and adapting the underlying visual representations to the task. Most existing policies need a lot of training samples and treat this problem from the lens of two-stage learning with a controller learned on top of pre-trained vision models. We approach this problem from the lens of Koopman theory and learn visual representations from robotic agents conditioned on specific downstream tasks in the context of learning stabilizing control for the agent. We introduce a Contrastive Spectral Koopman Embedding network that allows us to learn efficient linearized visual representations from the agent's visual data in a high dimensional latent space and utilizes reinforcement learning to perform off-policy control on top of the extracted representations with a linear controller. Our method enhances stability and control in gradient dynamics over time, significantly outperforming existing approaches by improving efficiency and accuracy in learning task policies over extended horizons.

In this work, we propose a distributed hierarchical locomotion control strategy for whole-body cooperation and demonstrate the potential for migration into large numbers of agents. Our method utilizes a hierarchical structure to break down complex tasks into smaller, manageable sub-tasks. By incorporating spatiotemporal continuity features, we establish the sequential logic necessary for causal inference and cooperative behaviour in sequential tasks, thereby facilitating efficient and coordinated control strategies. Through training within this framework, we demonstrate enhanced adaptability and cooperation, leading to superior performance in task completion compared to the original methods. Moreover, we construct a set of environments as the benchmark for embodied cooperation.

Humans possess a remarkable talent for flexibly alternating to different senses when interacting with the environment. Picture a chef skillfully gauging the timing of ingredient additions and controlling the heat according to the colors, sounds, and aromas, seamlessly navigating through every stage of the complex cooking process. This ability is founded upon a thorough comprehension of task stages, as achieving the sub-goal within each stage can necessitate the utilization of different senses. In order to endow robots with similar ability, we incorporate the task stages divided by sub-goals into the imitation learning process to accordingly guide dynamic multi-sensory fusion. We propose MS-Bot, a stage-guided dynamic multi-sensory fusion method with coarse-to-fine stage understanding, which dynamically adjusts the priority of modalities based on the fine-grained state within the predicted current stage. We train a robot system equipped with visual, auditory, and tactile sensors to accomplish challenging robotic manipulation tasks: pouring and peg insertion with keyway. Experimental results indicate that our approach enables more effective and explainable dynamic fusion, aligning more closely with the human fusion process than existing methods.

Existing methods for safe multi-agent control using logic specifications like Signal Temporal Logic (STL) often face scalability issues. This is because they rely either on single-agent perspectives or on Mixed Integer Linear Programming (MILP)-based planners, which are complex to optimize. These methods have proven to be computationally expensive and inefficient when dealing with a large number of agents. To address these limitations, we present a new scalable approach to multi-agent control in this setting. Our method treats the relationships between agents using a graph structure rather than in terms of a single-agent perspective. Moreover, it combines a multi-agent collision avoidance controller with a Graph Neural Network (GNN) based planner, models the system in a decentralized fashion, and trains on STL-based objectives to generate safe and efficient plans for multiple agents, thereby optimizing the satisfaction of complex temporal specifications while also facilitating multi-agent collision avoidance. Our experiments show that our approach significantly outperforms existing methods that use a state-of-the-art MILP-based planner in terms of scalability and performance.

Data collection in robotics is spread across diverse hardware, and this variation will increase as new hardware is developed. Effective use of this growing body of data requires methods capable of learning from diverse robot embodiments. We consider the setting of training a policy using expert trajectories from a single robot arm (the source), and evaluating on a different robot arm for which no data was collected (the target). We present a data editing scheme termed Shadow, in which the robot during training and evaluation is replaced with a composite segmentation mask of the source and target robots. In this way, the input data distribution at train and test time match closely, enabling robust policy transfer to the new unseen robot while being far more data efficient than approaches that require co-training on large amounts of data from diverse embodiments. We demonstrate that an approach as simple as Shadow is effective both in simulation on varying tasks and robots, and on real robot hardware, where Shadow demonstrates over 2x improvement in success rate compared to the strongest baseline.

Intelligent driving assistance can alert drivers to objects in their environment; however, such systems require a model of drivers' situational awareness (SA) (what aspects of the scene they are already aware of) to avoid unnecessary alerts. Moreover, collecting the data to train such an SA model is challenging: being an internal human cognitive state, driver SA is difficult to measure, and non-verbal signals such as eye gaze are some of the only outward manifestations of it. Traditional methods to obtain SA labels rely on probes that result in sparse, intermittent SA labels unsuitable for modeling a dense, temporally correlated process via machine learning. We propose a novel interactive labeling protocol that captures dense, continuous SA labels and use it to collect an object-level SA dataset in a VR driving simulator. Our dataset comprises 20 unique drivers' SA labels, driving data, and gaze (over 320 minutes of driving) which will be made public. Additionally, we train an SA model from this data, formulating the object-level driver SA prediction problem as a semantic segmentation problem. Our formulation allows all objects in a scene at a timestep to be processed simultaneously, leveraging global scene context and local gaze-object relationships together. Our experiments show that this formulation leads to improved performance over common sense baselines and prior art on the SA prediction task.

Robotic grasping in cluttered environments remains a significant challenge due to occlusions and complex object arrangements. We have developed ThinkGrasp, a plug-and-play vision-language grasping system that makes use of GPT-4o's advanced contextual reasoning for grasping strategies. ThinkGrasp can effectively identify and generate grasp poses for target objects, even when they are heavily obstructed or nearly invisible, by using goal-oriented language to guide the removal of obstructing objects. This approach progressively uncovers the target object and ultimately grasps it with a few steps and a high success rate. In both simulated and real experiments, ThinkGrasp achieved a high success rate and significantly outperformed state-of-the-art methods in heavily cluttered environments or with diverse unseen objects, demonstrating strong generalization capabilities.

Timely capturing the dense geometry of the surrounding scene with unlabeled LiDAR data is valuable but under-explored for mobile robotic applications. Its value lies in the huge amount of such unlabeled data, enabling self-supervised learning for various downstream tasks. Current dynamic 3D scene reconstruction approaches however heavily rely on data annotations to tackle the moving objects in the scene. In response, we present LiDARGrid, a 3D opacity grid representation instantly derived from LiDAR points, which captures the dense 3D scene and facilitates scene forecasting. Our method features a novel self-supervised neural volume densification procedure based on an autoencoder and differentiable volume rendering. Leveraging this representation, self-supervised scene forecasting can be performed. Our method is trained on NuScenes dataset for autonomous driving, and is evaluated by predicting future point clouds using the scene forecasting. It notably outperforms state-of-the-art methods in point cloud forecasting in all performance metrics. Beyond scene forecasting, our representation excels in supporting additional tasks such as moving region detection and depth completion, as shown by experiments.

Evaluating autonomous vehicle stacks (AVs) in simulation typically involves replaying driving logs from real-world recorded traffic. However, agents replayed from offline data are not reactive and hard to intuitively control. Existing approaches address these challenges by proposing methods that rely on heuristics or generative models of real-world data but these approaches either lack realism or necessitate costly iterative sampling procedures to control the generated behaviours. In this work, we take an alternative approach and propose CtRL-Sim, a method that leverages return-conditioned offline reinforcement learning to efficiently generate reactive and controllable traffic agents. Specifically, we process real-world driving data through a physics-enhanced Nocturne simulator to generate a diverse offline reinforcement learning dataset, annotated with various reward terms. We then train a return-conditioned multi-agent behaviour model that allows for fine-grained manipulation of agent behaviours by modifying the desired returns for the various reward components. This capability enables the generation of a wide range of driving behaviours beyond the scope of the initial dataset, including adversarial behaviours. We demonstrate that CtRL-Sim can generate diverse and realistic safety-critical scenarios while providing fine-grained control over agent behaviours.

Recognizing and manipulating transparent tableware from partial view RGB image observations is made challenging by the difficulty in obtaining reliable depth measurements of transparent objects. In this paper we present the Transparent Tableware SuperQuadric Network (T$^2$SQNet), a neural network model that leverages a family of newly extended deformable superquadrics to produce low-dimensional, instance-wise and accurate 3D geometric representations of transparent objects from partial views. As a byproduct and contribution of independent interest, we also present TablewareNet, a publicly available toolset of seven parametrized shapes based on our extended deformable superquadrics, that can be used to generate new datasets of tableware objects of diverse shapes and sizes. Experiments with T$^2$SQNet trained with TablewareNet show that T$^2$SQNet outperforms existing methods in recognizing transparent objects, in some cases by significant margins, and can be effectively used in robotic applications like decluttering and target retrieval.

The autonomous driving industry is increasingly adopting end-to-end learning from sensory inputs to minimize human biases in system design. Traditional end-to-end driving models, however, suffer from long-tail events due to rare or unseen inputs within their training distributions. To address this, we propose TOKEN, a novel Multi-Modal Large Language Model (MM-LLM) that tokenizes the world into object-level knowledge, enabling better utilization of LLM’s reasoning capabilities to enhance autonomous vehicle planning in long-tail scenarios. TOKEN effectively alleviates data scarcity and inefficient tokenization by producing condensed and semantically enriched representations of the scene. Our results demonstrate that TOKEN excels in grounding, reasoning, and planning capabilities, outperforming existing frameworks with a 27% reduction in trajectory L2 error and a 39% decrease in collision rates in long-tail scenarios. Additionally, our work highlights the importance of representation alignment and structured reasoning in sparking the common-sense reasoning capabilities of MM-LLMs for effective planning.

Planar-symmetric hands, such as parallel grippers, are widely adopted in both research and industrial fields. Their symmetry, however, introduces ambiguity and discontinuity in the SO(3) representation, which hinders both the training and inference of neural network-based grasp detectors. We propose a novel SO(3) representation that can parametrize a pair of planar-symmetric poses with a single parameter set by leveraging the 2D Bingham distribution. We also detail a grasp detector based on our representation, which provides a more consistent rotation output. An intensive evaluation with multiple grippers and objects in both the simulation and the real world quantitatively shows our approach's contribution.

In the field of locomotion task of quadruped robots, Blind Policy and Perceptive Policy each have their own advantages and limitations. The Blind Policy relies on preset sensor information and algorithms, suitable for known and structured environments, but it lacks adaptability in complex or unknown environments. The Perceptive Policy uses visual sensors to obtain detailed environmental information, allowing it to adapt to complex terrains, but its effectiveness is limited under occluded conditions, especially when perception fails. Unlike the Blind Policy, the Perceptive Policy is not as robust under these conditions. To address these challenges, we propose a MBC:Multi-Brain collaborative system that incorporates the concepts of Multi-Agent Reinforcement Learning and introduces collaboration between the Blind Policy and the Perceptive Policy. By applying this multi-policy collaborative model to a quadruped robot, the robot can maintain stable locomotion even when the perceptual system is impaired or observational data is incomplete. Our simulations and real-world experiments demonstrate that this system significantly improves the robot's passability and robustness against perception failures in complex environments, validating the effectiveness of multi-policy collaboration in enhancing robotic motion performance.