Safe RL for Islands & Bridges: A Topologically Isomorphoic Problem to In-hand Manipulation

Dexterous in-hand manipulation can be thought of traversing narrow regions of stable state space while moving between relatively more stable grasp poses. Mimicing this structure, we construct a topologically isomorphic problem, islands and bridges, wherein we connect large islands by narrow bridges to correspond to stable and unstable state-space respectively. Here, we demonstrate that using a shield to ensure safe actions signficantly improves sample complexity for goal based navigation. Our results suggest safe reinforcement learning for enabling model free learning for in-hand manipulation.