Research-collaboration proposal: URML primitives as V-JEPA 2-AC conditioning + V-JEPA 2 predictions as a URML safety lane (RFC for maintainer comment)

[idoco2003]

Hello Meta FAIR V-JEPA 2 maintainers,

URML is an open-source, vendor-neutral language for describing robot intent (Apache 2.0). Spec, validator, and reference runtimes live at https://github.com/URML-MARS/URML.

RFC-0052 proposes a urml-vjepa2-bridge package with two complementary integration vectors:

1. URML primitive sequences as V-JEPA 2-AC action conditioning input.
2. V-JEPA 2 predictions as a URML predictive-safety lane.

The world-model angle is unique among URML's AI/ML-layer outreach: every other RFC in our current wave wraps a policy that emits actions; V-JEPA 2 predicts future state.

This is framed as a research-collaboration proposal, not a feature request. The bridge is proposal-only; no code lands until your team weighs in.

Full RFC:
https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0052-meta-fair-vjepa2.md

Specific questions in the RFC focus on the action-conditioning encoding (V-JEPA 2-AC, Droid fine-tuning) and the predictive-safety lane shape (how to consume V-JEPA 2 predictions in a substrate-neutral safety envelope).

URML is independent open-source work; nothing on our side is gated by your response.

Thanks for any time you can spare.

Ido Yahalomi
URML maintainer
greenvh@gmail.com