Tolerance-Guided Policy Learning for Adaptable and Transferrable Delicate Industrial Insertion

Policy learning for delicate industrial insertion tasks (e.g., PC board\nassembly) is challenging. This paper considers two major problems: how to learn\na diversified policy (instead of just one average policy) that can efficiently\nhandle different workpieces with minimum amount of training data, and how to\nhandle defects of workpieces during insertion. To address the problems, we\npropose tolerance-guided policy learning. To encourage transferability of the\nlearned policy to different workpieces, we add a task embedding to the policy's\ninput space using the insertion tolerance. Then we train the policy using\ngenerative adversarial imitation learning with reward shaping (RS-GAIL) on a\nvariety of representative situations. To encourage adaptability of the learned\npolicy to handle defects, we build a probabilistic inference model that can\noutput the best inserting pose based on failed insertions using the tolerance\nmodel. The best inserting pose is then used as a reference to the learned\npolicy. This proposed method is validated on a sequence of IC socket insertion\ntasks in simulation. The results show that 1) RS-GAIL can efficiently learn\noptimal policies under sparse rewards; 2) the tolerance embedding can enhance\nthe transferability of the learned policy; 3) the probabilistic inference makes\nthe policy robust to defects on the workpieces.\n