Abstract
Dexterous in-hand manipulation often relies on multiple fingers or palm support to stabilize an object. Manipulating a tiny object only between the index fingertip and thumb is more challenging because small changes at either contact can quickly lead to slip or object loss. We study continuous multi-directional rotation of a tiny sphere held between two tactile fingertips. To address the partial observability of object motion and contact evolution, we use a teacher–student framework. A privileged teacher is trained with reinforcement learning using object and contact information, while a gravity curriculum gradually increases the task difficulty. The learned behavior is transferred through interactive policy distillation to a student that uses a history of proprioceptive and dual-fingertip tactile observations. The high-resolution tactile maps are summarized into a compact representation for history-based control. Simulation results demonstrate sustained multi-directional manipulation by the privileged teacher, with the distilled student retaining this behavior using only proprioceptive and dual-fingertip tactile histories.
Qiang Li
Professor and header of AG
My research interests include underwater robot, collaborative robots, humanoid robots.