Title: Co-observation-based Approaches to Mitigating Disturbances in Multi-Robot Systemst
Presenter: Kacper Wardega
Advisor: Wenchao Li (ECE)
Chair: David Starobinski (ECE)
Committee: Roberto Tron (ME), Christos Cassandras (SE, ECE)
Abstract: Many emerging robotic applications rely on cooperation between humans and robots, with robots following automatically generated plans to
achieve specific tasks. The problem of generating these multi-agent motion plans is known as multi-agent pathfinding (MAPF). Given their
safety requirements, it is important to understand and mitigate disturbances within such systems, such as those caused by motion delays,
compromised robots, or inter-robot interference.
In typical multi-robot deployments, the system execution is monitored by comparing robot self-reports of localization with the expected motion
plan. As a result, fault-tolerance mechanisms in centralized MAPF settings, as well as deadlock-avoidance in decentralized MAPF settings
rely both on the trustworthiness of the robot localization self-reports and on real-time performance of the communication between robots.
Our preliminary work demonstrates that such systems are vulnerable to plan-deviation attacks, whereby compromised robots diverge from their assigned paths to enter forbidden areas while trying to conceal their movement by mis-reporting their location. We found that by leveraging
robot co-observations to increase the ability of the system to detect plan-deviation attacks and horizon-limiting announcements to limit the amount of information available to the adversary, we can guaranteeably prevent attacks for a set of independent attackers.
This Ph.D. Thesis proposes to further characterize the impact of unreliable communication and untrustworthy robots within multi-robot systems. We propose mitigation mechanisms based on robot co-observations to lessen the requirement on real-time communications and fully-trusted
robots. Future work centers on the introduction of co-observation-based execution policies for robust execution and observable traffic law map decomposition in decentralized settings. |
