A Motion Planning Approach of Mobile Robots Based on Multi-Task Deep Reinforcement Learning

Recently, deep reinforcement learning has been applied to various fields, including motion planning for mobile robots. However, when applied to motion planning in dynamic environments, there are still issues such as slow convergence speed and insufficient planning success rate. Besides, during the motion planning process of a robot, conflicts may arise between its goal-oriented task and obstacle avoidance task, resulting in the robot being stuck or having difficulty finding the correct action. In this paper, a motion planning approach based on multi-task deep reinforcement learning is proposed. In the proposed approach, we decompose the motion planning task into two subtasks: goal orientation and obstacle avoidance. Based on the Twin Delayed Deep Deterministic policy gradient algorithm, we apply the multi-task learning method of projecting conflicting gradients to handle the gradient conflicts between sub-tasks. It assists in finding better optimization directions for the actor network, which avoids overestimating the improvement of the dominant task while underestimating the performance of the non-dominant task. We also improve the policy evaluation method of the Twin Delayed Deep Deterministic policy gradient algorithm. In designing the reward function, we incorporate the idea of artificial potential fields to increase the tendency of the robot to move away from obstacles and adapt to dynamic environments. Finally, we design comparative experiments for mobile robot motion planning, demonstrating that the proposed approach can improve the convergence speed of deep reinforcement learning and the success rate of motion planning.