Meta's RoboJEPA: 8B-parameter robot world model reveals first multi-embodiment scaling law
meta · hf · 2026-10-09
Meta released RoboJEPA, a JEPA-based latent world model trained on real-robot data spanning 12 embodiments. Its 'imagination error' follows a second-order power law in compute, enabling out-of-range quality prediction; downstream planning improves predictably with compute, and the model can be deployed zero-shot as a robotic agent planning toward a single goal image on real hardware. At 8B parameters it is the largest JEPA predictor to date, with checkpoints, training and deployment code all open-sourced.
Related event: Meta Releases RoboJEPA, an 8B-Parameter Robotic World Model(3 posts)→
More from Embodied
- Ditto-Bench stress-tests GPT-6 Astra for robot control: simple tasks shine, physics stalls — anand_bhattad · 2026-10-09
- NVIDIA shows frontier AI agents building Omniverse simulations from natural language — nordicinst · 2026-10-09
- NVIDIA Shows Six Projects Building Omniverse Simulations With Frontier AI Agents Like Astra — NVIDIA Blog · 2026-10-09
- Tesla Vision-Only Robotaxi Handles Nighttime Downpours, Silencing Earlier Skeptics — elonmusk · 2026-10-09
- Hark Sees Explosive 48-Hour Growth, Brett Adcock Says Team Scrambling to Fix Infra — adcock_brett · 2026-10-09
- aimotive driving dataset: train labels come from a tracker that sees the future — RexDouglass · 2026-10-09