Proposal: URML (substrate-neutral robot intent) capability-manifest mapping for Gemini Robotics SDK — multimodal-VLA typed substrate

[idoco2003]

Hi @google-deepmind gemini-robotics-sdk team,

Proposing a URML v0.1 capability-manifest mapping for Gemini Robotics SDK over google-deepmind/gemini-robotics-sdk. URML (Apache-2.0) is a substrate-neutral spec for robot intent: a typed primitive vocabulary plus a Layer-1 capability manifest and a validator that rejects, before any actuator publishes, any program a manifest cannot honor.

Gemini Robotics is a multimodal VLA that emits tool-calls; URML's primitive vocabulary is the natural typed substrate those calls dispatch to. URML's natural-language layer + query_detection primitive compose with Gemini's multimodal tool-call surface in a way structurally similar to the agentic-skills pattern URML engaged in a prior outreach. The high-leverage integration is straightforward: Gemini emits multimodal tool-calls → calls map to typed URML primitives → URML validates against the active manifest → URML dispatches to substrate.

This is proposal-only, part of URML's Move #11 outreach (15 net-new VLA / foundation-model RFCs in this wave). The integration is the highest-leverage URML-side touchpoint in this wave.

Full RFC: https://github.com/URML-MARS/URML/blob/main/docs/rfcs/0145-deepmind-gemini-robotics-sdk-outreach.md

Questions worth maintainer input on:

1. NL-layer substrate declaration manifest fields. Tool-registration convention, multimodal-context declaration, model-access binding?
2. Multimodal-input cross-sensor binding. Should URML's manifest declare that a multimodal VLA consumes specific sensors simultaneously?
3. Bridge home. URML repo (reference/llm-bridge/GeminiRoboticsBridge), google-deepmind/gemini-robotics-urml contributed example, or external?
4. Coordination with the DeepMind broader engagement. Single-entry-point preferred or separate per-repo?
5. Conformance listing. README link to URML's compatible-runtimes registry once a working bridge ships?
6. Anything else.

Ido Yahalomi (URML maintainer, urml.dev, greenvh@gmail.com)

---

AI-assisted prose, maintainer-reviewed before posting (see VIBE.md). Human-only correspondence available on request.