Environmentally Adaptive Control Policies for Expensive Systems

Running experiments on physical robotic systems typically requires significant resources (personnel, time, and money), restricting the number that can be run during training. Additionally, these systems typically operate over a range of external conditions and can be highly non-linear. We focus on the problem of learning a globally optimal policy to adapt controllers for such systems based on the value of these external conditions in order to maximize expected performance. We propose two novel myopic search methods for this problem, and present results comparing these algorithms with various other approaches. Finally, we use these methods to train both simulated and physical snake robots to automatically adapt to changing terrain.