Prediction in the context of a human-inspired posture control model

Abstract

Maintaining an upright stance is a challenging task in both humans and humanoid robots in that it is complicated by inherently unstable mechanics and requires controlling a noisy, inaccurate, multi-DoF system. Hence, learning and prediction are often involved in humanoid control. This appears to apply similarly in humans where continuous control with sensory feedback avoids falling. This work presents an attempt of integrating a learned predictor into a humanoid posture control. A neurorobotics approach was used, in which the concept of a bio-inspired modular control is tested on a 14 DOF humanoid robot platform. In particular, the paper shows how to address, in a closed loop system, the problem that sensory feedback tends to create a correlation between the state of the system and its controlled inputs. This complicates the training of machine learning models, or identification procedures in general, because the feedback makes the effect of the noise, which is unknown, on the system output dependent on the input. Using a bio-inspired modular control of the robot, a linear predictor based on on-line learning, integrated in the control through a bio-inspired schema, has been compared with the method of a Smith predictor. The on-line predictor compared favorably with the Smith predictor in the task of balancing during voluntary movements.