Alan C. Schultz and John J. Grefenstette
Autonomous vehicles will require both projective planning and reactive components in order to perform robustly. Projective components are needed for long-term planning and replanning where explicit reasoning about future states is required. Reactive components allow the system to always have some action available in real-time, and themselves can exhibit robust behavior, but lack the ability to explicitly reason about future states over a long time period. This work addresses the problem of learning reactive components (normative action models) for autonomous vehicles from simulation models. Two main thrusts of our current work are described here. First, we wish to show that behaviors learned from simulation are useful in the actual physical system operating in the real world. Second, in order to scale the technique, we demonstrate how behaviors can be built up by first learning lower level behaviors, and then fixing these to use as base cornportents of higher-level behaviors.