Home > Published Issues > 2021 > Volume 10, No. 1, January 2021 >

Assessment of Machine Learning of Optimal Solutions for Robotic Walking

Rodrigo Matos Carnier and Yasutaka Fujimoto
Yokohama National University, Yokohama, Japan

Abstract— The generation of optimal solutions for robotic bipedal walking using whole-body dynamics is well known to have a big computational cost, preventing online trajectory generation for optimal control methods that satisfy Pontryagin's Principle and its Conditions of Optimality. However, bipedal walking has fundamental kinematic and dynamic characteristics that shape different solutions for different parameters in similar curves. Such characteristics were previously defined in biomechanical literature as movement primitives. Recently, studies generated parametrized optimal solutions by performing regressions from training data into movement primitives using Machine Learning. The learned solutions were very close to the actual optimal solution. This study evaluates the precision of such strategy by optimizing the gait of a 6 degrees of freedom planar robot using different Cost Functions, in order to understand if the precision of Machine Learning in recreating optimal solutions is impacted by what is being optimized.

Index Terms— robotic bipedal walking, machine learning, optimal control, movement primitives

Cite: Rodrigo Matos Carnier and Yasutaka Fujimoto, "Assessment of Machine Learning of Optimal Solutions for Robotic Walking," International Journal of Mechanical Engineering and Robotics Research, Vol. 10, No. 1, pp. 44-48, January 2021. DOI: 10.18178/ijmerr.10.1.44-48