RoboJEPA: Scaling Robotic Latent World Models
RoboJEPA, an 8B JEPA world model on 12 embodiments, reports compute scaling laws for robotic planning.
Researchers present RoboJEPA, a Joint Embedding Predictive Architecture world model trained on real-robot data spanning 12 embodiments. Its imagination error follows a second-order power law in compute and correlates strongly with downstream planning performance, providing a proxy for real-robot evaluation. At 8 billion parameters, it is described as the largest JEPA predictor trained to date and can plan zero-shot from a single goal image on real hardware. All checkpoints, training code, and robot deployment code are released.
- RoboJEPA is trained on real data from 12 robotic embodiments.
- Imagination error follows a second-order power law in compute.
- The 8B JEPA predictor is the largest reported to date.
- Zero-shot goal-image planning is shown on real hardware.
- Checkpoints plus training and deployment code are released.
From the original article
1 min read at huggingface.co
Latent world models have shown a remarkable ability to predict future states and to plan in the real world. In practice, however, we lack a principled way to estimate how their capabilities scale with model size, data, and compute, an open problem that slows progress in the field.
By Artem Zholus, Nicolas Beltran-Velez, Jianhao Yuan · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.