Evaluation of walk optimisation techniques for the NAO robot

Locomotion performance is a critical component of any humanoid robot application. The procedure of optimising a walk engine has a high cost in both resources and time. The selection of the most appropriate optimisation algorithm, fitness function, and parameter space to maximise the benefit-cost ratio can dramatically improve the performance of the optimisation process. In this paper, we compare different meta-optimised optimisation algorithms, different fitness functions, and two different parameter spaces, in a physics-based simulation. The purpose of the comparison is to select the most appropriate combination to be used in hardware. The combination that yields the greatest increase in walk performance given a fixed expenditure is considered as the best, and is implemented in hardware. We found that Policy Gradient Reinforcement Learning with a fitness function based on the efficiency and a parameter space expanded to include the joint stiffnesses not only performed the best, in terms of improving the walk speed and efficiency, but also in terms of selecting gaits that were more stable. This combination was then applied to the physical NAO, improving the default walk´s speed by 57% and its efficiency by 30%.