Adaptive and Learning Robots

Abstract

In the past, robots have been required to be explicitly programmed to carry out tasks. Learning robots differ from this notion by allowing them to learn from experience, which allows them to get better the more they perform a task. This chapter will covers the basics of learning in robotics and give an overview of how machines can learn skills via reinforcement learning (RL). Topics that will be discussed include states, actions, rewards, and policy to better help understand how agents can learn from experience to create rewarding behaviours. Practical implementations will also be discussed, such as training robots in simulation before placing them into the real world. Important topics such as training in stochastic environments will also be discussed to help the reader better understand some of the difficulty that comes when training in these environments. To help exemplify some of these topics, Google Everyday Robot will be used as a case study to discuss some of the successes and failures of learning robots in the real world. By reviewing some of the issues with training in the real world vs. simulation, learners will gain a better understanding of some of the limitations. By the end of this chapter, learners will have a basic understanding of how robots can learn from their environment. Learners will be able to use this information to help train robots to best behave in uncertain environments.