Trust-Calibrated Meta-Learning for Adaptive Multi-Robot Motion Planning under Temporal Logic Specifications in Human-Robot Collaborative Bounding Overwatch

Project begins 2024.

Bounding overwatch is a fundamental movement technique that involves a collaborative teaming mechanism where the overwatch element provides security for a non-overwatch element. That is, the continued movement of the team depends on the ability of the overwatch element to provide coverage (respond to any threats). Therefore, any group movement will require the team to adapt to how the environment may impact the overwatch element’s coverage capability.

We will answer three fundamental research questions in this project:

Q1: How to effectively deploy robot assets and team robots with humans to adapt to environment/mission changes in real-time and provide enhanced overwatch through multiple viewpoints?

Q2: How to understand necessary factors that impact trust and evaluate the trust(worthiness) of the overwatch robots for multi-viewpoint bounding overwatch under dynamically changing team configuration in offroad environments?

Q3: How to automatically discover trust-calibrated multi-robot motion plans that can satisfy and adapt to complex mission requirements given the environment/mission situation in real-time?

This is the first attempt to utilize robot assets to provide multiple overwatch viewpoints and human-robot teaming for enhanced bounding overwatch in off-road environments. Our proposed work will establish a better understanding of trust impact factors for bounding overwatch and general offroad motion tasks. We will also develop trust-calibrated robot motion plans with linear temporal logic reward shaping based on meta-RL to achieve formally verified robot motion plans that can adapt to environmental changes in real-time. Besides the novel multi-viewpoint human-robot teaming framework, this project seeks to overcome two restrictions in basic knowledge from ARC 1.33 and brings about the following fundamental innovations: (1) Our previous work for multi-robot systems (MRS) symbolic motion planning in 1.33 assumes fixed environments modeled by a Markov decision process (MDP), which is not able to cope with changes and adapt in real-time. In this project, we seek to investigate novel meta-RL algorithms for MRS with a distribution of tasks modeled by a set of MDPs and efficient RL training approaches that utilize the learning history to realize near real-time adaptation; (2) In 1.33, we developed trust-seeking robot motion plans in order to find bounding paths that have the maximal overall trust and hence human’s willingness to accept the protection from the overwatch team, which however may lead to the overtrust situation.

In this project, we will develop a novel bidirectional trust calibration mechanism based on new computational trust(worthiness) models for MRS and feedback of trustworthiness and situational awareness information leveraging the existing literature on trust calibration.

#2.22