One Future, Every Robot: Label-Efficient Collective-State Prediction with Decentralized JEPA
CS-JEPA gets decentralized robots to predict a shared latent future without pooling their outputs or training them to agree.
Each robot uses its own history plus bounded neighbor messages, but deployment stays decentralized. In an independent replication, agreement improved across every seed and evaluated split while accuracy also rose. Against a raw-future reconstruction baseline, collective-state error fell 28.4% in distribution and 64.4% to 75.6% under topology and swarm-size shifts. Controls and action-conditioned tests suggest the representation still supports receiver-local decisions. HF Daily Papers' note
Each robot uses its own history plus bounded neighbor messages, but deployment stays decentralized. In an independent replication, agreement improved across every seed and evaluated split while accuracy also rose. Against a raw-future reconstruction baseline, collective-state error fell 28.4% in distribution and 64.4% to 75.6% under topology and swarm-size shifts. Controls and action-conditioned tests suggest the representation still supports receiver-local decisions. HF Daily Papers' note
score 4