The article introduces a learnable Perception-Action-Communication (LPAC) architecture for decentralized navigation in robot swarms, focusing on the coverage control problem. The proposed solution involves a convolution neural network (CNN) that processes localized perception of the environment, a graph neural network (GNN) that enables communication between robots, and a multi-layer perceptron (MLP) that computes robot actions. The system is trained using imitation learning with a centralized algorithm that is aware of the entire environment. The results show that LPAC models outperform standard decentralized and centralized coverage control algorithms and can generalize to different environments.
Publication date: 12 Jan 2024
Project Page: https://github.com/KumarRobotics/CoverageControl
Paper: https://arxiv.org/pdf/2401.04855
