D-JEPA: A Decision-Aligned Latent World Model
D-JEPA tries to make latent world models choose the action that works, not just the future that looks closest to the goal.
The paper identifies a “decision-local prediction gap,” where the best-looking predicted future can lead to a worse real outcome than another available option. D-JEPA learns relations among candidate futures from executed outcomes, then refines the model’s latent geometry around action choices. The authors report gains across control, manipulation, pretrained action-producing models, physical robots, and autonomous driving, including 87.89% success on PushT and a 17-point gain on physical robot tasks. HF Daily Papers' note
The paper identifies a “decision-local prediction gap,” where the best-looking predicted future can lead to a worse real outcome than another available option. D-JEPA learns relations among candidate futures from executed outcomes, then refines the model’s latent geometry around action choices. The authors report gains across control, manipulation, pretrained action-producing models, physical robots, and autonomous driving, including 87.89% success on PushT and a 17-point gain on physical robot tasks. HF Daily Papers' note
score 5