In this paper, reinforcement learning is applied to coordinate, in a decentralized fashion, the motions of a pair of hydraulic actuators whose task is to firmly hold and move an object along a specified trajectory under conventional position control. The learning goal is to reduce the interaction forces acting on the object that arise due to inevitable positioning errors resulting from the imperfect closed-loop actuator dynamics. Each actuator is therefore outfitted with a reinforcement learning neural network that modifies a centrally planned formation constrained position trajectory in response to the locally measured interaction force. It is shown that the actuators, which form a multiagent learning system, can learn decentralized control strategies that reduce the object interaction forces and thus greatly improve their coordination on the manipulation task. However, the problem of credit assignment, a common difficulty in multiagent learning systems, prevents the actuators from learning control strategies where each actuator contributes equally to reducing the interaction force. This problem is resolved in this paper via the periodic communication of limited local state information between the reinforcement learning actuators. Using both simulations and experiments, this paper examines some of the issues pertaining to learning in dynamic multiagent environments and establishes reinforcement learning as a potential technique for coordinating several nonlinear hydraulic manipulators performing a common task.
A Deep Reinforcement Learning based Homeostatic System for Unmanned Position Control
