Deep reinforcement learning of robotic manipulation for whip targeting

This paper aims at a deep reinforcement learning
(DRL) controller for fast (< 1.5s) manipulation of a flexible
tool (i.e., whip) to hit a target in 3D space. The controller
consists of a DRL algorithm for optimizing joint motions,
and a proportional-derivative (PD) mechanism for tracking
the optimized motions. Their objective is to minimize the
distance between the whip-end-tip and the target. The proposed
controller was validated in a 7-DOF robot arm by comparing
four DRL algorithms in the physical simulator MuJoCo. It
shows that the proximal policy optimization (PPO) outperforms
others by obtaining the maximum average reward. Notably,
PPO can still effectively interact with the environment under
sparse or even unrewarding conditions, making it a robust
choice for complex and dynamic tasks. Our work provides
preliminary knowledge of DRL applications to fast robotic arm
control in flexible object manipulation.