Learning Wrist Policies for Anthropomorphic Soft Power Grasping in Handle and Door Manipulation

In this work, we advance robotic grasping by incorporating wrist compliance in a unified hand–arm system inspired by human limb coordination. This integration improves grasping reliability and robustness through impedance and force learning in robotic arms. The compliant wrist system effectively compensates for uncertainties in object position and orientation. Employing a combined impedance-force control approach, we address diverse grasping and manipulation tasks in simulation. Successfully transferring the learned policy to a service humanoid mobile robot enables the seamless execution of grasping and opening tasks for various doors and handles without additional learning, using both fully actuated and underactuated robotic hands. Remarkably, our robust strategies yielded only one failure in 30 trials for the underactuated hand, even with up to 8 cm translation normal to the handle and $33^\circ$ rotation errors, and no failures for the fully actuated one with up to 12 cm translation and $30^\circ$ rotation. This significantly outperforms state-of-the-art end-to-end reinforcement learning approaches. Furthermore, we successfully tested and validated our approach across various constrained everyday tasks in different environments. Our proposed framework represents an advancement in the learning and execution of power grasping with compliant manipulation, achieving practically relevant performance.