KINEMATIC CONTROL OF SNAKE-LIKE ROBOT USING DEEP REINFORCEMENT LEARNING

Abstract

In many industries today around the globe, robots can be seen carrying out different tasks. These robots have capabilities to lift heavy loads, move at a very unbelievable speed, and execute tasks at a high level of pin-point accuracy. But despite this their amazing repertoire of tasks, most robots will find it very difficult to adapt themselves to new and environments that are unfamiliar to them. This could be because human environments are so dynamic and unpredictable and very difficult to be programmed, but rather must be learned firsthand by the robot. The desire to build machines that learn behavior based on the environment presented to them is one of the goals of Reinforcement Learning (RL). Reinforcement learning, an aspect of machine learning which is inspired by behavioral psychology, allows an agent – the learner and decision-maker, to automatically and autonomously discover optimal behavior through trial and error interactions with its environments in an attempt to solve problems. This work developed a Deep Reinforcement Learning based locomotion control model for snake-like robot. The agent proved to be robust for snake-like robot locomotion in training and achieved 95.8% successful episodes in the evaluation scenario. The learned locomotion by the agent is remarkably similar to the slithering gait and performs a similar sinus-like movement of a real snake. Deep Reinforcement Learning approach therefore remains promising for future application with snake-like robot.