Meta's RoboJEPA: First Scaling Laws for Multi-Embodiment Robotic World Models
apsarathchandar · x · 2026-10-30
RoboJEPA from Meta FAIR, Mila and Polytechnique Montréal is a latent world model for robotics built on the JEPA architecture and trained across 12 robotic embodiments. Key findings:
- Its imagination error (latent rollout error) follows a second-order power law in compute, extrapolating well beyond the scale at which the law is fit;
- Downstream planning performance improves predictably with compute and correlates strongly with imagination error, making it a reliable proxy for real-robot evaluation;
- The model works zero-shot as a robotic agent, planning toward a single goal image to solve long-horizon tasks on real hardware.
The authors call it the first work establishing scaling laws for multi-embodiment world models trained on real robot data; RoboJEPA at 8B params is the largest yet. All checkpoints plus training and deployment code will be released.
Author Artem Zholus also shares a low-cost recipe: collect 30 minutes of demonstrations with a cheap SO-101 robot, run public RoboJEPA inference on a single GPU for a few hours, and fit your own robotic scaling law.
More from Embodied
- Polymarket gives Neuralink 40% odds of hitting $60B valuation this year — Polymarket · 2026-10-30
- Tesla Robotaxi passes 4 million autonomous miles, adding ~900K in under two weeks with zero injuries — XFreeze · 2026-10-30
- Home-cleaning robot charging $30/hour points to a multi-billion-dollar TAM — xmercury_one · 2026-10-30
- Genie 4 sparks debate: robots' GPT moment nears, home humanoid adoption doesn't — FuSheng_0306 · 2026-10-30
- Handwriting Robot Completes Biology Homework With Cursive Text and Ink Diagrams — alvelda · 2026-10-30
- Foldable iPhone Duo already in stock at Paris Apple Store: 512GB at €2,589, pickup today — mitchdeg · 2026-10-30