About Site Language

WHDL is viewable in multiple languages. Use the pull-down menu to select a language to view the site.

I changed my language, but I’m still seeing resources in the other languages?

If a resource or text has not been translated into your selected language, it will appear in the initially added language. We are always looking for help translating these resources. If you can help, contact us!

WHDL - 00016168

Robotic Manipulators Using Deep Meta Reinforcement Learning to Catch Objects in Space

Aaron Borger

Resource Types

Academic (Thesis, Dissertation, Etc.)

Resource Languages

English

Institutional Repository

Description / Abstract

Submitted to the Department of Mathematics and Computer Science in partial fulfillment of the requirements for the degree of Bachelor of Arts

This work proposes a method to train an artificial intelligent robot in simulation and then re-train it in the real-world space environment. This method utilizes a meta-learning algorithm that allows an agent to re-train while in space and learn to catch an object with a robotic manipulator after only a few shots. This work seeks to develop the building block for intelligent astro-robotics by learning to catch a ball in space. Subsequent versions may be able to use the same method to learn to grasp uncontrollable, non-uniform objects in space. This effort will be beneficial towards solving multiple challenging problems such as assembling and servicing spacecraft in orbit by learning to manipulate tools and components in zero gravity as well as space debris removal by learning to catch resident space objects. The reinforcement learning algorithm at the core combines the Twin Delayed Deep Deterministic Policy Gradients algorithm (TD3) with the implicit model agnostic meta-learning algorithm (iMAML). This new algorithm allows for efficient few-shot learning with continuous observations, such as the position and velocity of the object as detected by a Mask Region-Based Convolutional Neural Network (MR-CNN), and continuous actions such as the location, orientation, and time of the grasp.

Download this Resource

Download Free from WHDL

Download as a PDF: Robotic Manipulators
File Size Small - 1.1 MB

Go to Download Information

Subject Index

Copyright Notes

This material is used by permission of the rightsholder in accordance with the terms of the relevant content release. As a user, you have permission to use this content and download a copy for non-commercial use only. Downloading multiple quantities of this resource is expressly forbidden.

KeywordsArtificial intelligence, Robotics, Deep learning, Reinforcement learning, Meta learning, Computer vision, Machine learning, Space exploration

Degree AwardedBachelor of Arts

Awarding InstitutionNorthwest Nazarene University

Committee MembersStephen Parke, Barry L. Myers

Supervisor or ChairDale Hamilton

Peer Reviewed?No

Keep Browsing