BiGAIL: Learning Intention-based Driving Policy from Multi-task Demonstrations

With the rapid development of traffic intelligence, autonomous driving technology has gradually attracted the interests of researchers. Behavioral decision-making is one of the most important parts of autonomous driving system (ADS). As a common solution, imitation learning (IL) provides a more natural and intuitive way of learning through the prior knowledge of experts. Generative adversarial imitation learning(GAIL), which is a branch of IL, is often used to learn the driving policy because of its robustness and capacity of handling large-scale problems. However, modal collapse caused by GAIL may make the generated policies lack diversity resulting in the failure of multi-task learning. In the paper, we propose an algorithm named as bidirectional generative adversarial imitation learning (BiGAIL) that allows the agent to learn the map between task intentions and driving policies, so as to achieve the goal of learning intention-based driving policy. Through simulation verification, the agent trained with BiGAIL is able to select the appropriate policy based on the current environment and learn different driving policies from multi-task demonstrations.