System identification and model reference adaptive control of bipedal locomotion with neural networks

Abstract

Biped robots have enormous potential to transform our lives with the usability of the infrastructure already prepared for humans, thanks to our morphological similarities. Unfortunately, the realization of this potential requires solutions to inherited difficulties of bipedal locomotion dynamics. Controlling these systems poses challenges due to their high degree of freedom, small support polygon, hybrid dynamics, etc. Nervous system ensures the locomotion of biological counterparts of biped robots. Similarly, artificial neural networks show competence to control bipedal locomotion. Unfortunately, complexity of neural network-based controllers (NNBCs) makes it difficult to adapt them to changes in robot model dynamics. In this paper, we propose a model reference adaptive control algorithm for bipedal locomotion, together with neural networks consisting of recurrent and feedforward layers in the controller and system identification tasks to overcome this drawback. Controller weights are updated via error gradient calculated through system identification neural network to force the controlled system output to desired behavior in an iterative manner. Moreover, we examine the effectiveness of our algorithm in decreasing the speed of steady-state error of neural controllers under different simulated scenarios. It is shown that recurrent and feedforward layers are beneficial for walking control and system identification with neural networks and implementable for real-time applications on various Jetson single-board computers. Results suggest that our method is capable of adapting neural controllers without requiring training from scratch. Under different scenarios, up to 33.6%, 34%, 31.8% in the target speed tracking error decrease is acquired with our algorithm on training, validation, and test sets.