Implicit Predictive Behavior Cloning for Autonomous Driving Decision-Making in Urban Traffic

Making efficient and safe decisions for autonomous driving in urban traffic is a highly challenging task that is crucial for ensuring driving safety and smooth traffic flow. Behavior cloning and data-driven learning-based models provide promising directions for related research such as decision-making and predicted information reasoning. Unlike explicit behavior cloning strategies, we propose a method based on implicit behavior cloning that is not limited to continuous action spaces and can learn implicit multimodal information in driving environments. Furthermore, our proposed model is built upon the Transformer network, enabling effective utilization of large-scale real-world datasets to learn and predict interactive information. In this study, we utilized the Lyft Level 5 dataset, which is currently the largest and most detailed dataset available for self-driving research, along with a differentiable simulator built based on this dataset for model training and validation. We tested the model in interactive and non-interactive simulation environments, and the results demonstrate that the proposed model outperforms compared baseline models on most evaluation metrics related to safety- or task-oriented performance. This work provides a new insight for the research field of autonomous driving decision-making in urban traffic.