Graph Neural Networks for Multi-Robot Coordination

Published:

The core question: How can locally communicating robots share structured information across the team to perform environmental monitoring more effectively?

The Problem with Local Rules

Classical coverage control (Lloyd’s algorithm) work well when robots sensor ranges can cover the entire environment. But in realistic scenarios—robots with limited sensing ranges, non-uniform starting positions, sparse regions of interest—purely local gradient-based approaches get stuck in poor configurations.

The insight: robots could do much better if they shared relevant information with each other across the network. But what information should they share, and how should they use it?

Our Approach: Learning to Communicate

We use Graph Neural Networks (GNNs) to learn communication policies that allow robots to leverage non-local information while remaining fully decentralized:

  • Each robot computes local gradient information from its sensors
  • The GNN transforms and propagates this information through the communication network
  • Multi-hop communication allows robots to “see” beyond their immediate neighbors
  • The policy is trained via imitation learning from a clairvoyant expert

The GNN architecture naturally handles the permutation symmetry in robot teams and scales gracefully to larger teams than seen during training.

Key Results

  • Significantly higher coverage quality compared to Lloyd’s algorithm on challenging non-uniform coverage densities
  • Transfers to new scenarios: policies trained on one density distribution work on others
  • Scales to larger teams: 10-robot policies work on 20+ robot teams
  • Explicit communication benefit: ablation studies confirm the value of multi-hop information sharing

Publications

  • IEEE International Conference on Robotics and Automation (ICRA), 2022
    Coverage Control in Multi-Robot Systems via Graph Neural Networks
    W. Gosrich, S. Mayya, R. Li, J. Paulos, M. Yim, A. Ribeiro, V. Kumar
    arXiv | IEEE

Collaborators

This work was conducted at Penn’s GRASP Lab as part of the DCIST program.