Adapt the Driving Policy to Local Traffic before Entering the New Area

Autonomous vehicles (AVs) may get failures when driving in a foreign area, e.g., a new city, since the driving policy has not been designed for the local traffic characteristics. Thus, it is necessary to carefully adjust the driving policy when driving into a new area. Due to the safety and efficiency issues thorough feedback adjustment. The work proposes the active environment adaption method for AV planning using reinforcement learning. It can adjust the driving policy to local traffic with traffic data before entering the new area. By extract the local traffic driving characteristic and forming virtual environment data, the driving policy will be adapted for better initial performance in the new area. This work first uses the conditional variational autoencoder (CVAE) method to extract the local traffic characteristics from the local vehicle's driving data. The extracted model can partially represent the future trajectories of the vehicles in this area. This model will then form a virtual environment by generating vehicles data to obey the local traffic distribution. The final policy will be trained using a reinforcement learning framework, e.g., deep Q learning. Our method is demonstrated by adapting the longitudinal driving policy between different local traffic. Our adaptive policy improve the performance compared to default policy and maintains similar performance to artificial adjustments in the target local traffic.