The application provides a
reinforcement learning path planning model training method fusing
potential field prior guidance. Firstly, local
observation data of a vehicle in an unstructured environment is acquired, and environment state characteristics are constructed; on this basis, an Actor-Critic policy network based on a TD3
algorithm is constructed. In view of the problem that a target cannot be reached in a traditional artificial
potential field method, the repulsive
force function is improved, a target distance adjustment factor is introduced into the repulsive
force function, and a mapping model from a virtual
potential field force to a kinematic control instruction of the vehicle is established. A
hybrid action exploration strategy is further constructed, the output action of the policy network and the artificial potential field guided action are linearly weighted and fused, and a dynamic weight self-adaptive attenuation
mechanism based on the
simulated annealing idea is designed, so that a smooth transition from rule guidance to autonomous decision is realized. In the network training process, a behavior
cloning (BC) auxiliary
loss function is introduced, and the update direction of the policy network is constrained by using the potential field guided action. The application can effectively alleviate the problems of
cold start difficulty and large exploration
blindness in the early stage of deep
reinforcement learning training, and has the advantages of fast convergence speed, high
obstacle avoidance safety and smooth path planning.