Meta's 8B RoboJEPA shows robot world models have scaling laws, trained on 15k hours of video
ylecun · x · 2026-10-08
Meta FAIR and Mila released RoboJEPA, an 8B-parameter JEPA world model trained on 15,022 hours of robot video across 23 public datasets and 12 robot embodiments (6,692 hours with actions) — reportedly the largest JEPA predictor to date. It imagines the future in V-JEPA 2.1's feature space and plans toward a single goal image.
Key findings:
- A scaling law fit on 22M–2B models accurately predicts 4B and 8B results.
- Skills unlock in order with compute: moving the arm, holding objects, avoiding obstacles; pushing objects emerges only past 10^22 FLOPs.
- On a real Franka with no task fine-tuning, the 8B model grasps 67% vs π0.5's 5% (goal-image vs text, not apples-to-apples); π0.5 still leads pick-and-place 53% vs 27%.
The paper suggests robot world models finally follow predictable scaling, with direct implications for embodied AI training investments.
More from Embodied
- Undergrad attends first IROS, gets hands-on with new humanoid robots — heatherknight · 2026-10-08
- Desktop robots MOSS and Vita star in an 80-minute robotics seminar at Thammasat University — StewartalsopIII · 2026-10-08
- Nvidia bets billions on physical AI, expanding Halos safety system to robots — Ars Technica AI · 2026-10-08
- Singapore debuts real-life Rock 'Em Sock 'Em robot boxing match — DJiafei · 2026-10-08
- LIBERO-MAX: mid-task changes cut success rates for all 14 robot policies by 11-25.7 points — DJiafei · 2026-10-08
- KAIST's FastOPD cuts VLA robot inference latency by 78% via on-policy distillation — kaist-ai · 2026-10-08