Responsibility-Based Socially Compatible Driving Behavior Modeling Verified by Hierarchical Multi-Agent Inverse Reinforcement Learning

Autonomous vehicles (AVs) offer a promising glimpse into a future where transportation is smarter, safer, and more streamlined. Nevertheless, as AVs continue to interact with conventional vehicles (CVs), the potential for increased complexities and challenges cannot be overlooked, such as the frozen robot problem. This study proposes a regret-based model for motion planning responsibilities, encompassing self-respect and courtesy for conflicting personal interests. By incorporating these reciprocal responsibilities, socially compatible driving behaviors are promoted, and uncertainties in behavior are also reduced. A Self-Respect-Courtesy (SR-C) plane is further introduced, illustrating the interaction intensity and tendency. To navigate the trade-offs of responsibilities in varying situations, the concept of environmental niche is provided. Niches help to characterize the outcomes of specific actions with the resulting conditions to fulfill responsibilities. Finally, a hierarchical multi-agent inverse reinforcement learning algorithm is designed to calibrate the proposed model with NGSIM highway lane-changing cases. We found that the proposed model can significantly improve the calibration results and reduce the predictions error of mandatory lane changes by up to 20%. Moreover, the cross-entropy error also significantly decreases in a stable stage, indicating that responsible actions can safely reduce the behavior uncertainties of interactions. Our research revealed that drivers prioritize courtesy responsibility in discretionary lane changes with more consistency, whereas their self-respect preferences are stronger but show more variability in mandatory lane changes. These findings provide valuable insights into the underlying mechanism of interactions.