Bayesian Optimization for Whole-Body Control of High-Degree-of-Freedom Robots Through Reduction of Dimensionality

This letter aims to achieve automatic tuning of optimal parameters for whole-body control algorithms to achieve the best performance of high-DoF robots. Typically, the control parameters at a scale up to hundreds are often hand-tuned yielding sub-optimal performance. Bayesian optimization (BO) can be an option to automatically find optimal parameters. However, for high-dimensional problems, BO is often infeasible in realistic settings as we studied in this letter. Moreover, the data is too little to perform dimensionality reduction techniques, such as principal component analysis or partial least square. We hereby propose an alternating BO algorithm that iteratively learns the parameters of sub-spaces from the whole high-dimensional parametric space through interactive trials, resulting in sample efficiency and fast convergence. Furthermore, for the balancing and locomotion control of humanoids, we developed techniques of dimensionality reduction combined with the proposed ABO approach that demonstrated optimal parameters for robust whole-body control.