Cross-Embodiment Dexterous Grasping with Reinforcement Learning

TL;DR

This work proposes a universal action space based on the human hand's eigengrasps, and simplifies the robot hand's proprioception to include only the positions of fingertips and the palm, offering a unified observation space across different robot hands.

Abstract

Dexterous hands exhibit significant potential for complex real-world grasping tasks. While recent studies have primarily focused on learning policies for specific robotic hands, the development of a universal policy that controls diverse dexterous hands remains largely unexplored. In this work, we study the learning of cross-embodiment dexterous grasping policies using reinforcement learning (RL). Inspired by the capability of human hands to control various dexterous hands through teleoperation, we propose a universal action space based on the human hand's eigengrasps. The policy outputs eigengrasp actions that are then converted into specific joint actions for each robot hand through a retargeting mapping. We simplify the robot hand's proprioception to include only the positions of fingertips and the palm, offering a unified observation space across different robot hands. Our approach demonstrates an 80% success rate in grasping objects from the YCB dataset across four distinct embodiments using a single vision-based policy. Additionally, our policy exhibits zero-shot generalization to two previously unseen embodiments and significant improvement in efficient finetuning. For further details and videos, visit our project page https://sites.google.com/view/crossdex.

Figures (9)

• Figure 1: We propose CrossDex, learning a cross-embodiment policy for dexterous grasping. The learned RL policy can grasp diverse objects with a variety of dexterous hands and transfer to hands not seen during training.
• Figure 2: CrossDex employs a unified observation and action space to facilitate the learning of a universal policy across various dexterous hands. Rather than relying on joint angles specific to each hand, our policy utilizes the positions of the fingertips and palm to discern the spatial relationship between the hand and the object. Actions are represented using eigengrasps from the MANO hand model, which are mapped to position targets of each hand's PD controller through a retargeting process. This design, akin to teleoperation, enables consistent control across different dexterous hands. The policy is trained using reinforcement learning within a cross-embodiment simulation environment built on IsaacGym. To learn a vision-based policy, we substitute the object pose in this pipeline with the object's point cloud.
• Figure 3: Comparison of RL training curves: cross-embodiment learning vs. individual training on the ShadowHand.
• Figure 4: Abaltion study on the eigengrasp action space. "MANO" refers to using raw axis angles of the MANO hand model as actions. "$k$-E" refers to using the first-$k$ eigengrasps as actions. Results show average success rates across five YCB objects.
• Figure 5: Average success rates for each YCB object across various embodiments.