Stable In-Hand Manipulation With Finger-Specific Multi-Agent Shadow Critic Consensus and Information Sharing

Deep Reinforcement Learning (DRL) has shown its capability to solve the high degrees of freedom in control and the complex interaction with the object in the multi-finger dexterous in-hand manipulation tasks. Current DRL approaches lack behavior constraints during the learning process, leading to aggressive and unstable policies that are insufficient for safety-critical in-hand manipulation tasks. The centralized learning strategy also limits the flexibility to fine-tune each robot finger's behavior. This work proposes the Finger-specific Multi-agent Shadow Critic Consensus (FMSC) method, which models the in-hand manipulation as a multi-agent collaboration task where each finger is an individual agent and trains the policies for the fingers to achieve a consensus across the critic networks through the Information Sharing (IS) across the neighboring agents and finger-specific stable manipulation objectives based on the state-action occupancy measure, a general utility of DRL that is approximated during the learning process. The methods are evaluated in two in-hand manipulation tasks on the Shadow Hand. The results show that FMSC+IS converges faster in training, achieving a comparable success rate and much better manipulation stability than conventional DRL methods.