Anomalously acting agents: the deployment problem

Abstract

The detection of intentionally antagonistic behavior in robot swarms brings about challenges that exceed identifying merely erroneous behavior. We investigate a data-based approach to recognize antagonistic behavior in robots executing a deployment task. The task requires a robot swarm of variable size and start positions to optimally distribute within an arbitrary convex surveillance area. Combining a long short-term memory neural network and a normalizing flow, our approach learns to approximate the probability of a robot action. Thus, actions with low probability density values can be categorized as anomalous. The applicability of the proposed approach is validated on simulated runs containing benevolent, antagonistic, and erroneous robots. Both antagonistic and erroneous robots are detected with an accuracy of more than 90 percent.