Federated Gaussian Process Learning via Pseudo-Representations for Large-Scale Multi-Robot Systems

Sanket A. Salunkhe and George P. Kontoudis

Proceedings of the 25th International Conference on Autonomous Agents and Multiagent Systems 2026

Paper Google Scholar

A scalable federated Gaussian-process framework that exchanges compact pseudo-representations for accurate modelling in large-scale multi-robot systems.

Abstract

Large multi-robot systems need scalable ways to model complex environments under computation and communication constraints. This paper introduces pxpGP, a distributed Gaussian-process framework that shares compact pseudo-representations instead of raw local data. The method combines sparse variational inference with a consensus optimization procedure, and is evaluated on synthetic and real-world datasets for both centralized and decentralized multi-robot networks.

Key contributions

Introduces a federated Gaussian-process framework for large multi-robot systems.
Shares compact pseudo-representations to reduce communication and computation demands.
Evaluates centralized and decentralized variants on synthetic and real-world data.

BibTeX

@inproceedings{salunkhe2026federated,
  title={Federated Gaussian Process Learning via Pseudo-Representations for Large-Scale Multi-Robot Systems},
  author={Salunkhe, Sanket A. and Kontoudis, George P.},
  booktitle={Proceedings of the 25th International Conference on Autonomous Agents and Multiagent Systems},
  pages={2392--2400},
  year={2026}
}