Uploaded October 2021 | Updated September 2026, 25 minutes ago
Lockheed Martin Robotics Seminar: Human-Centered Machine Learning for Autonomous Navigation
Maggie Wigness
Computer Scientist
Combat Capabilities Development Command
Army Research Laboratory
Data-driven AI/ML techniques have advanced significantly to automate skills such as detection, target recognition, and mobility. Yet, there are many applications, such as military operation or humanitarian assistance and disaster relief, where it is highly likely that the operating domain will depict some distributional shift from that in which a system was trained. Under these scenarios, the design of AI systems that can be trained or refined quickly, potentially in real-time, becomes critically important to ensure safety and success. I will discuss how we are incorporating learning from human demonstration to address the need for efficient, on-line learning in the field. I will specifically focus on several approaches to learn navigation behaviors for unmanned ground robots using teleoperation demonstrations, allowing for non-expert users to refine ML reward functions with relatively little effort.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
Lockheed Martin Robotics Seminar: Human-Centered Machine Learning for Autonomous Navigation
Maggie Wigness
Computer Scientist
Combat Capabilities Development Command
Army Research Laboratory
Data-driven AI/ML techniques have advanced significantly to automate skills such as detection, target recognition, and mobility. Yet, there are many applications, such as military operation or humanitarian assistance and disaster relief, where it is highly likely that the operating domain will depict some distributional shift from that in which a system was trained. Under these scenarios, the design of AI systems that can be trained or refined quickly, potentially in real-time, becomes critically important to ensure safety and success. I will discuss how we are incorporating learning from human demonstration to address the need for efficient, on-line learning in the field. I will specifically focus on several approaches to learn navigation behaviors for unmanned ground robots using teleoperation demonstrations, allowing for non-expert users to refine ML reward functions with relatively little effort.
For more information on the Maryland Robotics Center see:
https://robotics.umd.edu
