Reinforcement Learning Grasping with Force Feedback from Modelling of Compliant Fingers

Venue: IEEE/ASME Transactions on Mechatronics

Authors: Luke Beddow, Helge Wurdemann, Dimitrios Kanoulas

Pdf: https://dkanou.github.io/publ/J23.pdf

Abstract: Current reinforcement learning approaches for grasping do not consider force feedback combined with compliant grippers, but these features are suited to grocery object grasping, where objects can be bruised by high forces, and exhibit varied weights, textures, and surface irregularities. This work combines force feedback with reinforcement learning on a compliant gripper. We design a three degrees of freedom caging inspired gripper, which can grasp by trapping objects with three compliant fingers and a movable palm. We instrument the fingers with strain gauges for force sensing, then model their bending for simulating grasping at 16.8x real time. Then, we train a reinforcement learning grasping controller based on in-grasp force feedback, with real world transfer achieving 98.0% grasp success rate on training objects, with an average sim2real gap of 3.1%. We demonstrate generalization to 42 novel grocery objects with a success rate of 95.0%, with 80.1% of grasps tolerating a 5N vertical disturbance. In-grasp finger forces averaged 1.4N and palm forces 3.0N. We also validate our method with three finger rigidities, show that our model and in-grasp sensing improve learning and performance, and compare against three baselines.