Learning Robotic Skills through Reinforcement Learning

摘要

The application of modern robots has increased dramatically in recent years making systems more productive and efficient. Some of the traditional control approaches used for teaching movement to robots are PID (Proportional Integral Derivative) and LQR (Linear Quadratic Regulator) feedback controllers. These controllers have proven to help keep a system in a stable condition. However, as the model dynamics become more complex and sensitive, these techniques become more difficult to implement. Thus, traditional control techniques for adaptive robotic locomotion with multiple degrees of freedom render the task of maintaining a stable state impossible. This paper addresses the problem of traditional control approaches and proposes reinforcement learning for handling complex states as RL (Reinforcement Learning) is a robust framework and can tackle these problems by testing sample states in a simulated environment and will optimize the robot in each state to achieve for the maximum reward. To train the proposed robot in a realistic environment, this study applies the exploration and exploitation approach, as well as algorithms like curriculum learning, shape reward, input and reward normalization, etc. The findings of this study will generate novel ideas for locomotive robotic motion control and will encourage the use of artificial intelligence in robotic and other automated machines.