Publication: Action Generalization in Humanoid Robots Through Artificial Intelligence With Learning From Demonstration
Loading...
Identifiers
Publication date
2021-06
Defense date
2021-09-22
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
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.
Description
Mención Internacional en el título de doctor
Keywords
Action generalization, Learning from demonstration, Humanoid robots TEO, Continuous goal-directed actions, Evolutionary algorithms, Online evolved trajectories style, Transfer Reinforcement Learning, Deep reinforcement learning, TD3 neural style transfer, Neural policy style transfer