Fast-JEPA speeds latent world model planning 2.41x with 96.5% success rate
ylecun · x · 2026-11-21
Yann LeCun amplified Fast-JEPA, a new paper from a multi-institution team (Wuhan University, HIT, HUST and others, with internship work at DeepCybo) tackling the success-latency tradeoff in latent world model planning.
The problem: latent world models compare candidate action sequences via recursive rollout, which is expensive at test time.
Key results:
- Builds a goal-directed plan to guide search and directly predicts outcomes of complete candidates, avoiding recursive state rollout during refinement
- Across PushT, Reacher, Cube and TwoRoom: 96.50% average success rate at just 10.88 ms, 2.41x faster than INTACT
- Introduces DynaGoal, a benchmark for planning under dynamic out-of-distribution conditions, execution deadlines and controlled response delays; Fast-JEPA scores 73.83% on DynaGoal-Shift and 87.67% on DynaGoal-Deadline
Paper and code are both available.
More from Embodied
- Robot Arm Learns to Rack Vials: 0%→100% Grasp Rate After 27 Minutes on One GPU — MaziyarPanahi · 2026-11-21
- John Deere acquires Orchard Robotics, extending AI vision push into orchard automation — DynamicWebPaige · 2026-11-21
- Robot arm teaches itself to rack vials: one GPU, 27 min, 1010/1024 success — MaziyarPanahi · 2026-11-21
- Wayve self-driving demo handles complex Tokyo streets with zero interventions — alexgkendall · 2026-11-21
- Robot can restock the grocery store? Employees post chaotic photo in reply — MatthewChang · 2026-11-21
- FOSDEM 2027 Robotics & Simulation devroom opens call for talks, deadline Nov 23 — facontidavide · 2026-11-21