Report
RoboJEPA's reported scaling laws link compute to robot-planning performance
The RoboJEPA team says its model trained on data spanning 12 robotic embodiments, with predictors reaching 8 billion parameters.
TLDR
The RoboJEPA team reports that its world model’s prediction error falls predictably as training compute increases, while robot-planning performance improves. It also describes zero-shot planning toward a goal image on real robot hardware. The team says it will release model checkpoints and training and deployment code.
Combined views
—
1 Source, first seen ago
— likes— comments— saves— reposts
