High-speed highway scene prediction based on driver models learned from demonstrations

One of the key factors to ensure the safe operation of autonomous and semi-autonomous vehicles in dynamic environments is the ability to accurately predict the motion of the dynamic obstacles in the scene. In this work, we show how to use a realistic driver model learned from demonstrations via Inverse Reinforcement Learning to predict the long-term evolution of highway traffic scenes. We model each traffic participant as a Markov Decision Process in which the cost function is a linear combination of static and dynamic features. In particular, the static features capture the preferences of the driver while the dynamic features, which change over time depending on the actions of the other traffic participants, capture the driver's risk-aversive behavior. Using such a model for prediction enables us to explicitly consider the interactions between traffic participants while keeping the computational complexity quadratic in the number of vehicles in the scene. Preliminary experiments in simulated and real scenarios show the capability of our approach to produce reliable, human-like scene predictions.