A Simulation-Based Framework for Generating Alerts for Human-Supervised Multi-Robot Teams in Challenging Environments

Abstract

In a multi-agent mission with failures, uncertainty, complex dependencies, and intermittent information flow, the role of human supervisors is stressful and challenging. Alerts based on future mission predictions can be useful to assist the supervisors in responding to mission updates quickly, and devise more effective strategies. Monte-Carlo forward simulations can be used to estimate future mission states and build probability distributions of possible mission outcomes and generate alerts. However, in order to get reasonable estimates, we need a large number of simulations, representing a long-duration multi robot mission. All this needs to be performed within seconds, and therefore traditional physics based robotic simulations are infeasible. We adapt ideas from discrete event simulation paradigm, and present our novel simulation techniques like adaptive time step size, robot grouping, and intelligent time interval selection. Our technique achieves a sufficient level of accuracy in estimating probabilities, thereby generating higher quality alerts, while lowering overall fidelity of the discrete simulations for faster computation. We also provide theoretical insights on error levels in probability estimation using our method, which can guide in choosing appropriate levels of fidelity while maintaining accuracy requirements in different application scenarios. Lastly, we demonstrate sufficiently accurate real-time alert generation for a few representative mission scenarios, where the computational time is in the order of seconds using our adaptive techniques.