A Vision-Based Approach to Real-Time Trust Evaluation within Multiagent Systems

Abstract

As robotic autonomy increases, single-agent systems are evolving into more complex multiagent systems that require more factors for successful implementation. In exchange for this complexity, applications such as autonomous delivery services and reconnaissance swarms become possible when they are infeasible with only one agent. By incorporating Trust into the multiagent system, the complexity of the agent-to-agent relationships can be managed more reliably by accounting for agent defectiveness or abnormality. This research explores ways of obtaining trust values for other agents using their own sensors and implementing them using a novel trust model called Omni Light Trust. The implementation is also deployed on a real multiagent system tasked with mapping an unknown region collaboratively with realistic computational limitations. Results show the relationship trust has with certain perception techniques, mission performance, and time delay. The perception pipeline was found to be the greatest bottleneck to this research as an inline dependency to each iteration of the trust algorithm. Operating in real-time and minimizing time delay are instrumental to getting acceptable performance from the trusted output. Consequently, this research focused a lot on how to solve this bottleneck by surveying the different perception technique pathways one can choose and how they synergize with the trust algorithm.