GASP: Generative Adversarial State Policy and Multi-modal Sensing Framework for Dexterous Manipulation

초록

Dexterous manipulation using multi-finger hands requires precise control and adaptability. Previous studies of dexterous manipulation have been conducted via imitation learning (IL). However, training multi-finger robot hands to imitate expert-like behaviors still remains challenging, since traditional IL algorithms suffer from state occupancy mismatches. In this paper, we propose Generative Adversarial State Policy (GASP), a novel IL algorithm that employs a state-based discriminator to explicitly align the policy’s state distribution with demonstrations, thereby imitating expert-like behaviors. We also introduce a task environment that utilizes tactile sensors, which allows for more sensitive feedback on complex deformable objects. We validate our approach in tasks involving both rigid and deformable objects, including a challenging pour task, where it is shown that tactile feedback is critical for precise force control and dynamic adjustments of the multi-finger hand. Various experiments demonstrate that GASP outperforms state-of-the-art methods, with and without tactile sensors, in dexterous manipulation tasks.

키워드