Artificial Consciousness for Improving Reinforcement Learning

Reinforcement learning methods are useful for robot learning, but become slow when robots possess many degrees of freedom. We suggest equipping robots with fast on-board simulators, in order to accelerate learning. Such simulators will resemble forms of consciousness, enabling the robots to perform run-time trials in a simulated world, rather than tediously performing them in practice. We have applied this method to locomotion for a friction-propelled snake-like robot. The simulator on this robot uses an accurate non-linear model of isotropic friction that is fast enough to be executable in real time. Although our original goal was to propose a method for robot programming, the approach appears useful for reinforcement learning in a general context.