Walking of Prismatic Knee Biped Robot Using Reinforcement Learning

Abstract

This paper presents the RL framework for walking of a 6DOF planar bipedal robot with a prismatic knee joint. The prismatic knee joint allows the robot to adapt to varying terrain and maintain stability during dynamic movements. The design is based on a two-legged walking model, where each leg has a prismatic knee joint and two other joints: hip and ankle are revolute. The control we use is Reinforcement Learning based system that utilizes an Open AI Gym MuJoCo environment and stable-baselines-3 as a tool to train the Biped model. The friction, ground reaction forces, velocity and position are considered as observation space for neural network input. The robot was tested for walking on the plane floor and its performance was compared to that of a conventional bipedal robot with an articulated joint biped. The results showed that the biped with a prismatic knee joint achieved greater stability.