Reinforcement Learning on a Legged Robot

Lew, Louis Zun Kang (2022) Reinforcement Learning on a Legged Robot. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text LOUIS LEW ZUN KANG_Thesis.pdf Restricted to Registered users only Download (2MB)

Abstract

Legged locomotion has been a challenging task due to the required advanced control for legged robots. In common practice, engineers have to specifically design the controller for a particular robot. However, most of the controllers do not work well in dynamic environments and required a huge amount of resources and time to be properly developed and tested. Reinforcement learning (RL) is an alternative approach that can be taken. For years, engineers and AI scientists have discussed that RL will be the future of autonomous robots by learning different kinds of skills and locomotions. The main objectives of this project are to demonstrate the approach of using reinforcement learning (RL) for a quadruped robot to learn and perform basic movements which include standing, walking and self turning. RL models are being designed and developed with different approaches and Proximal Policy Optimization (PPO) algorithm to learn as efficiently as possible by figuring out the optimal policy. The Webots robot simulator is used to create a virtual environment for the robot to learn. Multiple models are trained and concluded that the combination of the inverse kinematics control method and the Cheetah walking pattern has the best results. It has significantly better cumulative rewards and much more consistent performance in comparison to the others. Standing and turning on the spot actions are also successfully performed by the agent but come with limitations. In conclusion, the objectives of this project are achieved and proved that such an approach is doable to be deployed in a physical robot.