JEPA-WAM Couples Latent Transitions With Action Generation
Video world-action models buy better dynamics by generating futures — and then pay for it at every control step. Latent WAMs drop the pixels but often compress the future into a handful of tokens or bolt prediction on the side of the action backbone.
JEPA-WAM, submitted to arXiv on Aug 10, 2026 (arXiv:2608.09381) with work done partly as internships at XYZ Embodied AI, builds a latent WAM inside a frozen V-JEPA 2.1 space. A shared predictor jointly predicts a dense current–future target and feeds dedicated tokens to a flow-matching action expert. At deploy, the target branch is gone.
Design
The transition target is not “future only.” The encoder jointly sees current and future frames and produces a patch-level joint current–future representation so the model learns what stays stable and what moves, with spatial correspondence preserved. The same supervision can be attached to a pretrained VLA (demonstrated on π0.5) without rewriting its perception stack.
Project page: spritewithoutice.github.io/JEPA_WAM.
Numbers
- LIBERO average: 96.7% for JEPA-WAM (0.5B backbone); 97.8% for π0.5+JEPA
- LIBERO-Plus (train on LIBERO only): 79.2% best among methods without large-scale robot-policy pretraining; π0.5+JEPA 86.3% best overall (π0.5 alone 84.5%)
- RoboTwin 2.0 (20 tasks): 79.9% Clean / 36.9% Random for JEPA-WAM; inference 85 ms (11.76 Hz) on the authors’ RoboTwin setup
- Real AgileX COBOT Magic bimanual suite: JEPA-WAM 59.8% ID / 54.2% OOD average score; π0.5+JEPA 90.3% / 84.7%
A Human’s Take
I like that they treat world modeling as backbone shaping, not as a free future video the action head has to babysit. The 79.2% without giant robot pretraining is the number I keep — it says good temporal structure can be stolen from V-JEPA instead of only from teleop fleets. Next receipt I want is open code and multi-lab replications on the same OOD splits.