GARL: Grasp-Aware Reinforcement Learning for Robust and Dexterous in-Hand Rotation
Abstract
: Dexterous multi-fingered hands on robotic arms are often subject to varying orientations and accelerations, yet most reinforcement learning (RL) methods for in-hand manipulation assume a fixed hand pose, limiting robustness and generalization. Our approach incorporates both object-related and hand-related considerations within a RL pipeline (Proximal Policy Optimization in our case) to achieve desired behaviors for dexterous manipulation tasks. It targets key aspects of multifingered grasps through quantitative measures of guaranteed stability, controllability, and manipulability. This enables a comprehensive analysis of the learned policies, characterizing the safety margin against undesired grasp configurations. It favors finger-gaiting sequences characterized by perturbation-invariant in-hand rotation with up to 12 × fewer training steps and improves generalization to unseen perturbation intensities by 25 × . The approach also enhances task performance by 18 . 1% and reduces average joint torques by 26 . 7%. Object generalization experiments demonstrate that our method provides consistent and reliable training outcomes. These simulation-only results bring the required robustness for future real-world deployment on humanoid robots. Code and Videos are available on our website.