Action Generalization In Humanoid Robots Through Artificial Intelligence With Learning From Demonstration [Online]
Action Generalization is the ability to adapt an action to different contexts and environments. In humans, this ability is taken for granted. Robots are yet far from achieving the human level of Action Generalization. Current robotic frameworks are limited frameworks that are only able to work in the small range of contexts and environments for which they were programmed. One of the reasons why we do not have a robot in our house yet is because every house is different. In this thesis, two different approaches to improve the Action Generalization capabilities of robots are proposed. First, a study of different methods to improve the performance of the Continuous Goal-Directed Actions framework within highly dynamic real world environments is presented. Continuous Goal-Directed Actions is a Learning from Demonstration framework based on the idea of encoding actions as the effects these actions produce on the environment. No robot kinematic information is required for the encoding of actions. This improves the generalization capabilities of robots by solving the correspondence problem. This problem
is related to the execution of the same action with different kinematics. The second approach is the proposition of the Neural Policy Style Transfer framework. The goal of this framework is to achieve Action Generalization by providing the robot the ability to introduce Styles within robotic actions. This allows the robot to adapt one action to different contexts with the introduction of different Styles. Neural Style Transfer was originally proposed as a way to perform Style Transfer between images. Neural Policy Style Transfer proposes the introduction of Neural Style Transfer within robotic actions. The structure of this document was designed with the goal of depicting the continuous research work that this thesis has been. Every time a new approach is proposed, the reasons why this was considered the best new step based on the experimental results obtained are provided. Each approach can be studied separately and, at the same time, they are presented as part of the larger research project from which they are part. Solving the problem of Action Generalization is currently a too ambitious goal for
any single research project. The goal of this thesis is to make finding this solution one step closer.