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

Researchers have developed a new federated learning framework called pxpGP that dramatically speeds up training for large-scale multi-robot systems. It tackles the crippling cubic computational complexity of traditional Gaussian Process models by using sparse variational inference to create compact 'pseudo-representations'. In tests, pxpGP and its decentralized variant dec-pxpGP outperformed existing distributed GP methods in both hyperparameter estimation and prediction accuracy, especially in large networks. The work was accepted at AAMAS 2026.