DTVAD:End-to-End Autonomous Vehicle Planning Using Reinfocement Learning

Reinforcement learning (RL) based end-to-end autonomous driving faces significant challenges due to the sim2real gap, primarily because simulation environments are often insufficiently realistic. However, collecting data from the real world is highly risky, as RL requires exploration of potentially dangerous behaviors to learn which situations should be avoided. Additionally, the loss function of RL may hinder the optimization of the perception module. In this paper, we propose DTVAD, an RL-based end-to-end autonomous driving framework that directly leverages real-world datasets, specifically the nuScenes dataset, and employs a decision transformer as the planning module. Notably, we randomly sample 4,096 trajectories as the action space and utilize the Fréchet distance to measure the similarity between trajectories. For each sample in the nuScenes dataset, we select one of the top 70 most similar trajectories as the action for the corresponding state. The discrepancy between the selected action and the recorded trajectory may introduce potential collisions, enabling the agent to learn which behaviors should be avoided. And we use Decision Transformer(DT) as our planning head. We evaluate the effectiveness of DTVAD on the nuScenes dataset through open-loop test.