Collision-Free Trajectory Planning of Mobile Robots by Integrating Deep Reinforcement Learning and Model Predictive Control

In this paper, we present an efficient approach to real-time collision-free navigation for mobile robots. By integrating deep reinforcement learning with model predictive control, our aim is to achieve both collision avoidance and computational efficiency. The methodology begins with training a preliminary agent using deep Q-learning, enabling it to generate actions for next time steps. Instead of executing these actions, a reference trajectory is generated based on them, which avoids obstacles present on the original reference path. Subsequently, this local trajectory is employed within an MPC trajectory-tracking framework to provide collision-free guidance for the mobile robot. Experimental results demonstrate that the proposed DQN-MPC hybrid approach outperforms pure MPC in terms of time efficiency and solution quality.