One Future, Every Robot: Label-Efficient Collective-State Prediction with Decentralized JEPA

Researchers introduced CS-JEPA, a decentralized swarm prediction model that lets robots infer a shared future state from local observations and a small amount of messaging.

It is pretrained without global labels and then evaluated with few labeled episodes, outperforming a raw-future reconstruction baseline on prediction error and inter-robot agreement.

The method remains robust across topology and size shifts, including ring topologies, mutual-kNN settings, and unseen swarm-size families.

A follow-up also showed better planning-related value estimation, suggesting a scalable and label-efficient path for multi-robot shared-state learning.