Quietly Robust Multi-Agent Systems

Project Overview

This research develops multi-agent systems which are robust to disturbances even while they intelligently minimize communications.

World Modeling

Agents must understand the world in which they operate, necessitating at least some global representation of the world. We seek to create embodied world models which will enable agents to predict the current global state given their own sparse data.

Communication Reduction

By predicting more information about the world, we decrease the required communication for coordination of multi-agent systems in partially observable environments. We aim to learn uncertainty-reducing communication paradigms.

Multi-Agent Coordination and Planning

Pairing the predictive world model with uncertainty-reducing communication/action co-planning will improve the robustness and increase the utility of multi-agent systems in harsh environments.

Existing Works

Below are some existing works in this direction: