Accession Number:

AD1183861

Title:

Lifelong Learning of Perception and Action in Autonomous Systems

Descriptive Note:

[Technical Report, Final Report]

Corporate Author:

Trustees of the University of Pennsylvania

Personal Author(s):

• Eaton, Eric
• Baveja, Satinder
• Grauman, Kristen
• Konidaris, George
• Learned-Miller, Erik
• Littman, Michael
• Mataric, Maja
• Sha, Fei
• Stone, Peter

Report Date:

2022-11-02

Pagination or Media Count:

234

Abstract:

Under the DARPA Lifelong Learning Machines L2M program, we explored a comprehensive approach to lifelong learning for autonomous systems, addressing fundamental issues of continual learning and transfer across diverse tasks, scalable knowledge maintenance, self-directed learning, and adaptation to changing environments for embodied agents. The key aspects of our L2M approach include continual learning for perception and action, transfer between diverse tasks, scalable lifelong knowledge maintenance, self-directed learning for autonomous discovery, and modeling the non-stationary distribution of tasks. We explored each of these aspects separately, developing various lifelong learning algorithms for classification and reinforcement learning settings. These developed algorithms were then integrated together via modular framework, yielding an L2M system that supports both classification and reinforcement learning tasks. We evaluated the lifelong learning performance of this L2M system using the Johns Hopkins Applied Physics Labs MiniGrid lifelong learning benchmark, and applied this L2M system to integrated perception and action through a robotic scavenger hunt using Matterport 3D, showcasing our L2M systems ability to rapidly learn diverse tasks in unstructured environments and quickly adapt to changes.

Descriptors:

Subject Categories:

Distribution Statement:

[A, Approved For Public Release]