The Three-Layer Stack: Why DeepMind Open-Shipped ER 2 and Gated the VLA
Gemini Robotics 2 is a deliberate three-layer product topology — public ER 2 reasoning, partner-gated VLAs, and on-device adaptation — that lets Google ship orchestration before whole-body control is broadly available.
Google DeepMind's July 30 robotics release is best read as an architecture memo, not a humanoid hype reel. Gemini Robotics 2 splits physical AI into three layers — and the split explains who gets access to what.
Layer 1: Gemini Robotics 2 (VLA) — the motor cortex, gated
The vision-language-action model converts pixels and language into motor commands for whole humanoids and bi-arm platforms. This is where "feet to fingertips" lives — walking, crouching, dexterous picks.
It is also the layer not on the open API. DeepMind routes Gemini Robotics 2 and On-Device 2 to early-access hardware partners.
Layer 2: Gemini Robotics ER 2 — the planner, public
ER 2 is available on Gemini API, Google AI Studio, and private preview on Gemini Enterprise Agent Platform. It watches video, plans multi-minute tasks, and orchestrates lower-level VLAs.
Multi-robot collaboration — wheeled bases handing off to humanoids — lives here.
Layer 3: On-Device 2 — latency and adaptation
On-Device 2 targets edge deployment. DeepMind claims few-hour adaptation with under 200 examples.
Why publish success rates now?
DeepMind disclosed pick rates (76.3% shelf, 45.7% floor) and flagged multi-finger dexterity as still challenging — unusually candid for a launch blog.
Safety orchestration as product surface
ASIMOV-Agentic tests whether ER agents refuse unsafe VLA tool calls and escalate to humans.
### Sources
- Google DeepMind — Gemini Robotics 2 brings whole body intelligence to robots (July 30, 2026)
- Google — Introducing Gemini Robotics ER 2 (July 30, 2026)
- Unite.AI — Google Ships Gemini Robotics ER 2 With Multi-Robot Teamwork (July 30, 2026)