KDDRL framework improves AV safety in mixed traffic with dual-driven RL

Autonomous vehicles (AVs) navigating mixed traffic face a stubborn trio of problems: physics-based priors miss the hidden intentions of human drivers, abrupt maneuvers create non-stationary conditions that leave long-tail safety events unexplored, and hybrid action spaces destabilize training because continuous car-following and discrete lane-changing operate on different time scales. A new paper from Jie Fang, Wei Zheng, Mengyun Xu, and Eui-Jin Kim, posted to arXiv (2608.13878) on August 14, 2026, tackles all three with a framework they call KDDRL (Knowledge-Data Dual-Driven Reinforcement Learning).

The core idea is to stop treating perception as passive. KDDRL first uses a conditional deep generative model to synthesize intention-aware future trajectories for surrounding vehicles, converting raw sensor data into predictive states. On top of that, a knowledge-data dual-driven paradigm fuses probabilistic data-driven insights with physics-based constraints, guiding the RL agent to safely explore safety-critical scenarios instead of relying on either source alone. A coupling module then compresses both the intention-aware trajectories and physical constraints into compact shared embeddings, which enables asynchronous multi-timescale optimization—so the continuous action of maintaining a following distance and the discrete action of changing lanes can be trained together without interference.

In dataset-calibrated simulations, KDDRL demonstrated better handling of intention uncertainty, faster training convergence, and superior performance across safety, efficiency, and comfort metrics compared to conventional baselines. This is a promising step toward AVs that can reason proactively about unpredictable human drivers in mixed traffic.