From operational to tactical driving: A hybrid learning approach for autonomous vehicles

Research in intelligent transportation systems has increased dramatically in recent years, with the main goals to improve road safety and increase transportation capacity. This paper presents our research work that aims at realizing general driving skill learning capability in autonomous vehicles, with the foreseeable benefit of achieving human-like flexibility and robustness in complex dynamic environments. The challenge is to develop intelligent vehicles endowed with strategic, tactical, and operational skills, which can competently drive in real-world traffic conditions. In our approach, operational driving skills such as lane following, U-turn, reverse parking, etc. are modeled as approximate decision-making rules mapping sensory input to control output. The system automatically captures human expertise by extracting the rules from example. Tactical driving proficiency, on the other hand, is realized using stochastic learning S-model automata, which determine in real-time from sensory data which maneuver to perform given incomplete information about the rapidly changing traffic environment.