Getting the Ball Rolling: Learning a Dexterous Policy for a Biomimetic Tendon-Driven Hand with Rolling Contact Joints

TL;DR

This paper presents the Faive Hand, a biomimetic, tendon-driven dexterous hand with rolling contact joints designed for accessible manufacturing and RL-ready simulation. By integrating rolling joints into a GPU-accelerated simulation (IsaacGym) and employing EKF-based joint sensing, the authors train a closed-loop policy that achieves zero-shot transfer to the real robot for in-hand sphere rotation. Key contributions include modeling rolling joints in simulation, mapping joint commands to tendon actuators, and demonstrating sim2real transfer under domain randomization. The work highlights the potential of high-DoF, human-like hands paired with scalable RL pipelines, while noting challenges in sensing accuracy and system identification that limit more complex tasks. Overall, the Faive Hand offers a practical platform for advancing dexterous manipulation research with RL in real-world settings.

Abstract

Biomimetic, dexterous robotic hands have the potential to replicate much of the tasks that a human can do, and to achieve status as a general manipulation platform. Recent advances in reinforcement learning (RL) frameworks have achieved remarkable performance in quadrupedal locomotion and dexterous manipulation tasks. Combined with GPU-based highly parallelized simulations capable of simulating thousands of robots in parallel, RL-based controllers have become more scalable and approachable. However, in order to bring RL-trained policies to the real world, we require training frameworks that output policies that can work with physical actuators and sensors as well as a hardware platform that can be manufactured with accessible materials yet is robust enough to run interactive policies. This work introduces the biomimetic tendon-driven Faive Hand and its system architecture, which uses tendon-driven rolling contact joints to achieve a 3D printable, robust high-DoF hand design. We model each element of the hand and integrate it into a GPU simulation environment to train a policy with RL, and achieve zero-shot transfer of a dexterous in-hand sphere rotation skill to the physical robot hand.

Figures (6)

• Figure 1: (a) The GPU-based parallelized simulation environment simulating 4096 robot hands in parallel to train a RL policy. (b) The trained policy being deployed on the tendon-driven robot hand with rolling contact joints.
• Figure 2: (a) The Faive Hand has a rolling contact joint design with tendons and ligaments mimicking that of a human finger. (b) Overview of each component of the Faive Hand.
• Figure 3: (a) The Faive Hand can grasp a payload of up to 10kg, demonstrated here with a dumbbell. (b) The motion of the rolling contact joint in the MuJoCo simulator and on the real robot hand, which do not rotate around a fixed axis.
• Figure 4: Overview of the RL training framework for achieving dexterous manipulation on the tendon-driven robot hand with rolling contact joints. After training the policy within a simulation environment, the actor network is transferred to the real robot.
• Figure 5: Training curve evolution for the policies, trained with and without domain randomization (DR), and for a reversed target rotation direction. We took the mean of 9 training rounds for each approach. An area of $\pm\sigma$ is shown around both plots.