Maneuver Planning and Learning: a Lane Selection Approach for Highly Automated Vehicles in Highway Scenarios.

Highway scenarios are highly dynamic environments where several vehicles interact following their own goal, leading to different combinations of scenes that also change over time. Human drivers adapt their driving behavior integrating current information with their former experiences. In a similar way, an autonomous system performing any driving activity should be able to integrate information learned from former interactions. Reinforcement Learning has shown promising results, but it should only be applied to autonomous vehicles if the system is also able to fulfill safety and integrity requirements on a deterministic and reproducible way. This paper presents a planning system that is able to learn over time, always complying to the safety requirements. Our planner integrates several layers interacting with each other, combining the advantages of Reinforcement Learning based systems and reactive systems. We present a planner that ensures driving safety on short horizons and integrates previous experiences to optimize the expected reward. We evaluate our method in simulation comparing different learning techniques. Our results show that the planning system is able to adaptively integrate this experience outperforming rule-based strategies.