In this experiment, both the domain and the generative model are noisy. Noise is introduced by applying a uniformly distributed noise to one randomly selected joint at each time step. The noise is not “well-behaved” as it is not gaussian, and not zero-mean. I believe the results look better than the previous post, where the domain was noisy, but the generative model was not.
Cross-Entropy open loop planning applied to noisy humanoid walking