What it is
Goal-conditioned JEPA world model that learns direct intent-to-action control, avoiding broad action search at deployment. Predicts actions directly from current state and goal intent in a JEPA world model, training one shared intent-to-action operator on both observed local changes and deployable goals so control can run without CEM/MPPI search.
Key results
- 95.33% direct macro SR in 1 epoch with 0 search
- 2.9-5.5 ms direct-control latency; 23.44x fewer candidates
- Shared encoder: 89.39% direct SR vs 66.17% for LeWM CEM 300x30
Shows world models can answer the inverse control query directly, without broad action search at deployment.