Meta FAIR open-sources RoboJEPA, an 8B-parameter latent world model for robots

Meta FAIR and collaborators released RoboJEPA, a JEPA-style action-conditioned latent world model for robots, covered on 8 October 2026 after an 7 October arXiv paper. Authors and correspondence include Artem Zholus, Jeannette Bohg (Meta), Nicolas Ballas and Mahmoud Assran.

Scale and training

• Models span 22 million to 8 billion parameters; the 8B model is described as the largest JEPA predictor to date.
• Training used large-scale multi-embodiment robot data spanning 12 embodiments (the paper’s emphasis).
• Imagination error follows a second-order power law in compute, which the authors say enables extrapolation.
• Downstream planning improves with compute, and imagination error correlates with planning performance.

What is released

• Checkpoints, training code and robot deployment code are open-sourced on GitHub (facebookresearch/robo_jepa), with a project site at robojepa.github.io.
• The work reports zero-shot agent planning toward a single goal image on real hardware.

The paper does not justify inventing per-task success percentages beyond what it states; claims about planning gains should be read in the paper’s own evaluations.