MAT: Multi-Fingered Adaptive Tactile Grasping via Deep Reinforcement\n Learning

摘要

Vision-based grasping systems typically adopt an open-loop execution of a\nplanned grasp. This policy can fail due to many reasons, including ubiquitous\ncalibration error. Recovery from a failed grasp is further complicated by\nvisual occlusion, as the hand is usually occluding the vision sensor as it\nattempts another open-loop regrasp. This work presents MAT, a tactile\nclosed-loop method capable of realizing grasps provided by a coarse initial\npositioning of the hand above an object. Our algorithm is a deep reinforcement\nlearning (RL) policy optimized through the clipped surrogate objective within a\nmaximum entropy RL framework to balance exploitation and exploration. The\nmethod utilizes tactile and proprioceptive information to act through both fine\nfinger motions and larger regrasp movements to execute stable grasps. A novel\ncurriculum of action motion magnitude makes learning more tractable and helps\nturn common failure cases into successes. Careful selection of features that\nexhibit small sim-to-real gaps enables this tactile grasping policy, trained\npurely in simulation, to transfer well to real world environments without the\nneed for additional learning. Experimentally, this methodology improves over a\nvision-only grasp success rate substantially on a multi-fingered robot hand.\nWhen this methodology is used to realize grasps from coarse initial positions\nprovided by a vision-only planner, the system is made dramatically more robust\nto calibration errors in the camera-robot transform.\n