Enhancing UGV Navigation with Adaptive Human Intervention

Project Team

Principal Investigator

Dawn Tilbury, University of Michigan Lionel Robert, University of Michigan

Government

Jon Smereka, U.S. Army GVSC

Industry

Lilia Moshkina, SoarTech

Ahmed Mekky, Gen Sasaki, Mathworks

Student

Wonse Jo (Research Fellow), Seung Hun Lee (MS Student), U of M

Project Summary

Project begins 2024.

The goal of this research is to develop methods that can help semi-autonomous UGVs that are navigating uncertain and complex terrain effectively request and incorporate human input to arrive at their desired destination. We will begin our research using the Artificial Potential Fields (APF) method, whereby the UGVs follow “artificial potentials,” representing a positive charge on the goal and a negative charge on obstacles. Since the APF can exhibit local minima, which can lead to a “stuck” UGV, we will first develop a method to predict a local minimum in advance. We will present this prediction to a human operator, and explore different approaches to request human expertise that can help the UGV avoid the local minimum. As the project moves forward, we will also add a learning capability to the system so that similar local minima can be avoided in the future without requesting human intervention.

RQ1: How can the UGV determine that it is approaching a local minimum? How far in advance (distance, time) of the local minimum can this be detected? How does this depend on the context/environment and capability of the sensors?

RQ2: How does the information on the first-person view (video, vectors, points, etc.) and the global view (which may be poorly-known) help the user suggest useful waypoints to avoid the local minimum? What information is most useful and when?

RQ3: How can a database of human-suggested waypoints in diverse environments be used to create a heuristic or machine-learning model that can inform future automated planning algorithms? What improvement in the automated planning method is enabled? How challenging is it for humans to suggest new waypoints in scenarios the UGV still cannot resolve?

By addressing the three research questions, we aim to develop methods that not only predict a local minimum in advance but also determine when, where, and how human intervention is needed for successful path planning (to avoid being trapped in a local minimum).

#2.21