Figure 4.

This figure illustrates the learning performance of the two movement representations, DMPs and PMPs, for the one-dimensional via-point task. Illustrated are mean values and standard deviations over 15 runs after CMA policy search. In addition, we also compare to the PI² approach (Theodorou et al., 2010) which we could only evaluate for the DMP approach. Without noise the final costs of the two representations are similar if CMA policy search is used (A). In the second example (B) we use zero-mean Gaussian noise with σ = 20 for the controls. In this setup we needed to average each performance evaluation for CMA over 20 roll-outs. For both setups the PMPs could considerably outperform the DMPs in terms of learning speed. For the noisy setup the PMPs could additionally produce policies of much higher quality as they can adapt the variance of the trajectories to the task constraints. PI² could not find as good solutions as the CMA policy search approach in both setups.