Training a Legged Robot to Walk Using Machine Learning and Trajectory Control for High Positional Accuracy

Training a Legged Robot to Walk Using Machine Learning and Trajectory Control for High Positional Accuracy

Amit Biswas, Neha N. Chaubey, Nirbhay Kumar Chaubey
Copyright: © 2023 |Pages: 16
DOI: 10.4018/978-1-6684-8171-4.ch006
OnDemand:
(Individual Chapters)
Available
$37.50
Current Special Offers
No Current Special Offers
TOTAL SAVINGS: $37.50

Abstract

Legged robots are a class of biologically inspired robots that use articulated leg mechanisms for locomotion. Legged motion is very complex and requires specialized actuation mechanisms and complicated motion control systems to operate. Traditional legged robots were controlled by purely physics-based, however, recent developments of artificial intelligence (AI), and machine learning (ML) techniques have opened new opportunities to train locomotion skills in a legged robot in a much more efficient way than the traditional physics-based controllers. In this chapter, the authors study how machine learning techniques are used to train quadruped robots in basic locomotion skills, evaluate training accuracy, training speed, and also discussed performance, simulation environment, trajectory control, and how the authors achieved accurate tracking of trajectories. Furthermore, this chapter delves into the details of the actual quadruped robot that the authors built to evaluate locomotion policies and some challenges that were faced in building the real robot.
Chapter Preview
Top

Several methods of controlling a robot’s motion have been tried on several different robots. For example, the MIT Cheetah robot generates simple reference trajectories and performs model predictive control to get desired contact forces and then uses Jacobian transpose control to realize them, as discussed by Tuomas Haarnoja in 2018. The ANYmal robot, described by M. Hutter in 2016, plans footholds based on the inverted pendulum model. These methods of control work well but require considerable knowledge of the locomotion task to perform and also a deep understanding of the robot’s dynamics. These requirements are often a limiting factor. A much better approach will be to use a method that can work well without prior knowledge of the robot dynamics or the locomotion task. Reinforcement learning-based training promises to achieve that. Tuomas Haarnoja in 2018 demonstrated that using reinforcement learning it is possible to learn locomotion skills without an explicit model of the robot dynamics.

Complete Chapter List

Search this Book:
Reset