DDPG-MPC: A Safe and Efficient Hybrid Path Planning Algorithm for Robots

Local path planning has always been an important problem in the field of robotics, aiming to make the control robot plan the collision-free optimal or sub-optimal path through local information. In this paper, a new safe hybrid local path planning method Deterministic Policy Gradient with Model Predictive Control (DDPG-MPC) is proposed to realize obstacle-free navigation for mobile robots. By combining reinforcement learning with model predictive control, the learnability and rapidity of reinforcement learning are preserved under the constraints of model predictive control. In order to realize more reliable and efficient local path planning, this method first uses DDPG method to train an obstacle-avoiding agent, which can generate the next action according to the current situation forward reasoning and hand it to the robot. Then, this paper uses the designed switcher to automatically switch between DDPG and MPC algorithms according to the constraints that the current robot needs to meet to realize local path planning. The experimental results show that the proposed hybrid local path planning algorithm DDPG-MPC is superior to pure DDPG in safety and MPC in efficiency.