Latent-space video world model beats V-JEPA 2, DINO-WM and Dreamer on real gas-station drive-offs
richdotca · x · 2026-11-15
MaxSebti trained a video world model that predicts the future in latent space instead of generating pixels, applied to detecting real gas-station drive-offs.
Context on prior work:
- V-JEPA 2 (Meta): learns physics from 1M+ hours of internet video, plans zero-shot on robots
- DINO-WM: predicts on frozen DINOv2 features
- Dreamer: recurrent latents for control
The author's model beat all of them on real drive-off clips while staying within the false-alarm budget. Benchmark charts show markedly higher recall at 10^-3–10^-2 false alarms/hour over 400 normal hours, budget compliance on both 400 normal and 100 busy hours, and near-zero alarms on honest customers.
More from Embodied
- Growing Bans on AI Smart Glasses Show Cultural Response Powering Privacy Fight — CarissaVeliz · 2026-11-15
- China flies 10-ton electric passenger plane with pilot aboard after autonomous tests — MarcoFigueroa · 2026-11-15
- PolyUMI wins Best Device Track at CoRL UMI Arena: open visual-tactile-audio rig — GeorgiaChal · 2026-11-15
- Kosmo and Ropedia release Physical AI dataset: 1,589 3D-scanned environments, 131 hours of egocentric video — liuziwei7 · 2026-11-15
- Berkeley Legend Patterson Predicts Superintelligent Chef Robots Within Two Years — davidpattersonx · 2026-11-15
- Robot dog spotted walking a real dog on a leash in San Francisco — Polymarket · 2026-11-15