A Reinforcement Learning Approach in Assignment of Task Priorities in Kinematic Control of Redundant Robots

Based on the recent advances of Deep Reinforcement Learning (DRL) and promising results, in this paper, we propose a framework for strict priority assignment in the context of kinematic control of redundant robots. The presented method focuses on redundant robots performing multiple concurrent tasks with potentially conflicting requirements and learns how to re-assign task priorities to ensure critical tasks get executed through smooth transitions. A Deep Q-Network (DQN) reinforcement learning agent is trained to assign the proper strict priorities to a stack of predefined kinematic control tasks (e.g., position control, orientation control, obstacle avoidance control, etc.) in a varying environment. Furthermore, to address the discontinuities in the control law due to the changes in the task priorities, a smoothing algorithm is proposed to ensure continuous reference velocities to the robot's joints. The proposed method is generic and extendable to a higher number of tasks and can be used when a reordering, swapping, addition, or deletion of tasks is required. The effectiveness of the proposed method is shown in simulation on a 5-DoF planar manipulator and a 7-DoF planar bipedal robot. The results show that the DRL agent is successful in assigning the correct hierarchy of tasks at each robot's state based on the global goal of the robot.