A Task-Adaptive Deep Reinforcement Learning Framework for Dual-Arm Robot Manipulation

Closed-chain manipulation occurs when several robot arms perform tasks in cooperation. It is complex to control a dual-arm system because it requires flexible and adaptable operation ability to realize closed-chain manipulation. In this study, a deep reinforcement learning (DRL) framework based on actor-critic algorithm is proposed to drive the closed-chain manipulation of a dual-arm robotic system. The proposed framework is designed to train dual robot arms to transport a large object cooperatively. In order to sustain strict constraints of closed-chain manipulation, the actor part of the proposed framework is designed in a leader-follower mode. The leader part consists of a policy trained from the DRL algorithm and works on the leader arm. The follower part consists of an inverse kinematics solver based on Damped Least Squares (DLS) and works on the follower arm. Two experiments are designed to prove the task adaptability, one of which is manipulating an object to a random pose within a defined range, the other is manipulating a delicate structural object within a narrow space.