Improving humanoid locomotive performance with learnt approximated dynamics via Gaussian processes for regression