Inverse Reinforcement Learning in Automatic Driving Decision

Abstract

With the urgent need of automatic driving on urban roads, autonomous unmanned system must complete the driving task considering safety, efficiency and comfort. For the planning and decision-making module, reinforcement learning can learn human strategies in a human-like manner. However, the reward function is difficult to be determined manually, and inverse reinforcement learning (IRL) can find a reasonable reward function that explains the human strategy. In this paper, the machine learning method on unmanned system is studied, and the IRL based on maximum entropy is introduced to learn the reward function. Experiments on the real-world nuScenes dataset is implemented by setting the features of reward function that conforms to urban environmental constraints. Finally, a reasonable reward function is obtained, which demonstrates the weights of the features can describe the trajectory of unmanned vehicle under the urban road.