RoboJEPA: Scaling Robotic Latent World Models
RoboJEPA, an 8B JEPA world model trained on 12 robot embodiments, reports compute scaling laws and zero-shot planning.
RoboJEPA is a Joint Embedding Predictive Architecture world model trained on real-robot data spanning 12 embodiments. Its latent rollout error follows a second-order power law in compute, and downstream planning performance improves predictably with compute, with imagination error correlating strongly with real-robot results. The authors show zero-shot deployment as agents that plan toward a single goal image for long-horizon tasks on hardware. At 8 billion parameters it is described as the largest JEPA predictor trained to date, and checkpoints plus training and deployment code are released.
- RoboJEPA trains on real data from twelve robotic embodiments.
- Latent imagination error follows a second-order power law in compute.
- Planning quality on real robots tracks that imagination error.
- An 8B checkpoint plans zero-shot from a single goal image.
- Authors release checkpoints plus training and deployment code.
From the original article
1 min read at arxiv.orgLatent 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 · arxiv.org
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.