VLA-JEPA Has Arrived

I’m doing graduate research on robot foundation models and Physical AI.

Vision-Language-Action (VLA) models still have real challenges — they struggle to understand physical laws, and they’re weak on long-horizon tasks, among other things. According to one source I came across, real-world VLA deployment success rates are apparently quite low — which matches my own sense of things as well. There are of course cases being deployed via PoCs, but I take that figure to be specifically about full production deployments, though the source and methodology behind it aren’t something I’ve verified myself.

Now, to help address VLA’s weaknesses, VLA-JEPA has arrived. For background on JEPA, see this article: ai.meta.com

That said, VLA-JEPA seems to differ from JEPA2 in the following way. As I understand it, VLA-JEPA is a new kind of VLA that incorporates JEPA’s approach into VLA pretraining.

	V-JEPA2	VLA-JEPA
Purpose	World model	VLA pretraining
Input	Video	Video + robot data
Output	Future latent	Action policy
Planner	Yes	No
Policy	No	Yes
Use case	Planning	Manipulation

For now, having read this, I’m going to check the code and try running it on my own GPU server next.

sophon.at

---

This is an English translation of the original Japanese post: VLA-JEPAが出た