Off-line path integral reinforcement learning using stochastic robot dynamics approximated by sparse pseudo-input Gaussian processes: Application to humanoid robot motor learning in the real environment | IEEE Conference Publication | IEEE Xplore