Model-free Reinforcement Learning-based Locomotion for Bipedal Robot

Project Introduction

In this project, we developed a series of locomotion controllers for the bipedal robot XR4 in Mujoco using model-free reinforcement learning. These locomotion controllers enable XR4 to walk and run on different terrains, as well as demonstrate agile football dribbling skills.

Demonstrations

Step in Place

Demonstration of basic locomotion ability of a bipedal robot - stepping in place.

Walk in Different Terrains

Demonstrations of neural network locomotion policy that can adapt to different terrains. The forward walking speed of the bipedal robot is set to a fixed 1 m/s.

Run

Demonstrations of natural forward running ability.

Football Dribble

Demonstrations of a continuous football dribbling task combining locomotion ability and simple object interaction.