The Application of Artificial Intelligence in Robotic Object Manipulation

Abstract

Robotic object manipulation remains one of the most researched areas, as current control-based methods lack adaptability, only being able to perform highly specialized tasks. However, no simple solution is currently available to help resolve this problem. Although there has been a myriad of studies conducted in the field, there are still multiple approaches to the problem and a solution that is easy to implement would enable robots to act more intelligent. This perspective paper argues that reinforcement learning techniques, paired with simulation to real life transfer methods and supplemental machine learning models, would allow for robots to learn generalized manipulation skills as well as constantly learn from each interaction for future application. Furthermore, this paper explores the structure and design of rewards, diversification of objects inside of an environment, data acquisition strategies, and anticipated challenges. By framing manipulation as a task to constantly learn from, the creation of a robust and general purpose manipulation system to become more than just a hypothetical concept.