Knowledge Transfer via Dynamic Policy Fusion Between Autonomous Driving Agents

Autonomous driving in mixed and dynamic environments requires the ability to learn from diverse traffic scenarios, specifically unsignalized intersections, while adapting to new ones with minimal exploration. In this work, we address the challenge of transferring experience between autonomous driving agents in a Fine-tuned Transfer Learning framework. We consider a setting where a source agent, an expert vehicle previously trained to handle a specific unsignalized intersection, shares its learned policy with a target agent which serves as the ego vehicle and has never encountered the new environment. Rather than relying solely on random exploration, the target agent dynamically integrates the source agent's policy parameters without any additional learning process. We propose a novel policy sharing framework relying on a fusion mechanism for Q-Learning policies between source and target agents. Simulation results in SUMO demonstrate improved efficiency, safer interactions, less collisions and more consistent decision-making compared to related approaches. This work highlights the potential of context-aware policy sharing for accelerating adaptation and enhancing safety in collaborative autonomous driving systems.