Zian Wang, Wenjie Huang, Zejian Deng, Yiming Shu, Jiahui Xu, Yong Wang, Shen Li, Dongpu Cao, Chen Sun ✉
SAFE-AD is a research prototype for socially-aware and risk-aware reinforcement learning in interactive autonomous driving. The central idea is to use a physics-informed propagated risk field as a structured intermediate representation for RL tactical planning. Instead of penalizing only instantaneous scalar risk, SAFE-AD models how risk propagates through traffic and maps this field to ego safety, surrounding-vehicle exposure, and social externality.
The preliminary PDE-governed risk-field model is based on DRIFT.
- Propagated risk field: models spatial-temporal traffic risk instead of only instantaneous ego risk.
- PINN surrogate: learns a differentiable approximation of the PDE-governed risk field.
- Risk-aware RL: appends field-derived risk features to the policy observation.
- Social-aware reward shaping: penalizes imposed risk, backward disturbance, jerk, abrupt steering, and unsafe close interactions.