Anticipatory Driving for a Robot-Car Based on Supervised Learning

Prediction and Planning are essential elements of successful human driving, making them equally important for autonomously driving systems. Many approaches achieve planning based on built-in world-knowledge. However, we show how a learning-based system can be extended to planning, needing little a priori knowledge. A car-like robot is trained by a human driver by constructing a database, where look ahead sensory information is stored together with action sequences. From that we achieve a novel form of velocity control, based only on information in image coordinates. For steering we employ a two-level approach in which database information is combined with an additional reactive controller. The result is a trajectory planning robot running at real-time, issuing steering and velocity control commands in a human manner.