Approximate policy transfer applied to simulated Bongo Board balance

Developing global policies for humanoid robots using dynamic programming is difficult because they have many degrees of freedom. We present a formalism whereby a value function for a humanoid robot can be approximated using the known value functions of similar systems. These similar systems can include approximate models of the robot with reduced dimensionality or trajectories derived from human motion capture data. Once an approximate value function is known, a local controller is used to compute control signals. The approximate value function provides information about the global strategies that should be used to solve the task. The local controller provides complementary information about the robots dynamics. We present an implementation of this strategy and simulation results generated by this implementation.