Training system for autonomous driving control policy

A unmanned vehicle, comprising a processor, configured to:

• construct a simulator to simulate static factors of power systems of vehicles and driving roads, and dynamic factors of pedestrians, non-motor vehicles, and surrounding vehicles,
• wherein road videos are captured by a road camera, and the dynamic factors are detected in the road videos;
• set a driving objective function in the simulator which is constructed, and a driving control policy of an optimal objective function is searched by using a machine learning algorithm;
• and
• execute a policy transfer to retrain the driving control policy searched out in the simulator according to data acquired by the unmanned vehicle to obtain a retrained driving control policy used for the unmanned vehicle, and
• control movement of the unmanned vehicle based on the retrained driving control policy, wherein
• an execution of the policy transfer comprises:
• running a control policy model in the unmanned vehicle with an autonomous driving control policy model obtained through training in the simulator as a starting point, and updating the control policy model by obtained data;
• executing a control action sequence (a1, a2, a3, . . . , an) on the unmanned vehicle, and collecting perception states (s0, s1, s2, s3, . . . , sn) of all executed action;
• setting an initial state in the simulator as s0, and executing a same action sequence (a1, a2, a3, . . . , an);
• and collecting perception states (s0, u1, u2, u3, . . . , un);
• constructing a transition correction function g in the simulator, inputting an action a from a current state s and the driving control policy π to the g, and outputting a correction action a′ replacing the action a from the g and executing the a′ in an environment, wherein a′=g(s, π(s));
• and
• training the g by using an evolutionary algorithm or a reinforcement learning algorithm to minimize a difference between data from the unmanned vehicle and data from the simulator, wherein Σi(si−ui)2 is minimized;
• and
• after training the g, obtaining the retrained driving control policy π through training in the simulator, and the retrained driving control policy π is directly used for controlling the movement of the unmanned vehicle.