$ cat wiki/models/gemini-robotics-er-2.md
Gemini Robotics ER 2
Spec
| Attribute | Value |
|---|---|
| Developer | Google DeepMind |
| Released | 2026-07-30 |
| Announced | 2026-07-30 |
| Context window | unknown |
| Pricing | unknown |
| License | proprietary (API-only) |
| Availability | Gemini API, Google AI Studio (gemini-robotics-er-2-preview); private preview on Gemini Enterprise Agent Platform |
| Full name | Gemini Robotics ER 2 (Embodied Reasoning) |
| Architecture | embodied-reasoning VLM; orchestration layer above a VLA |
Release Date
Announced 2026-07-30, two days after its VLA sibling Gemini Robotics 2 (source).
It is the only publicly available member of the three-model family — the VLA and the on-device VLA are early-access partnership only. That asymmetry is the release's shape: the reasoning layer ships to every developer, the acting layer does not.
Benchmarks
Video understanding (source):
| Measure | Value |
|---|---|
| Continuous progress classification | 57.4% accuracy |
| Moment finding (exact frame of a critical event) | 91.3% accuracy |
| Moment finding, mean absolute distance error | 0.96 seconds |
| DeepMind reports ER 2 competing closely with much larger model categories at **4× the execution | |
| speed**. The comparison category is not named numerically, so this is a claim of relative | |
| standing rather than a measured gap. |
Use Cases
- Task orchestration — planning and sequencing over a VLA acting layer
- Multi-robot collaboration — coordinating more than one robot on a shared task
- Video understanding for robotics — locating the moment a critical event occurs in a stream, to within ~1 second
- Embodied reasoning available to developers without robot hardware access
Safety — ASIMOV-Agentic
Released with ASIMOV-Agentic, described as a new open benchmark for the safety of agentic robotic AI (source). It evaluates whether the embodied-reasoning layer can:
- refuse dangerous commands from the VLA layer below it,
- assess whether a task is physically feasible before attempting it,
- proactively request human intervention when uncertain.
This is a safety benchmark aimed at the layer boundary — whether the reasoning model will decline what the acting model proposes. It arrives in the same week as Frontier Pacing and the TC260 agent-interaction draft recorded on AI Governance, all three of which are about constraining what an agent may do to another agent or to the world.
Compared To
| Model | Role | Availability | Headline measure |
|---|---|---|---|
| Gemini Robotics ER 2 | embodied reasoning | Gemini API | moment finding 91.3% |
| Gemini Robotics ER 1.6 | embodied reasoning | Gemini API | gauge reading 93% |
| Gemini Robotics 2 | VLA (acting) | early access | shelf pick 76.3% |
| ER 1.6 and ER 2 share no benchmark, so the 93% and 91.3% figures measure different things and | |||
| must not be read as a regression | |||
| (source). |