On 30 July 2026, Google DeepMind announced Gemini Robotics 2, Gemini Robotics ER 2 and Gemini Robotics On-Device 2. The family expanded robot control to full humanoid motion, dexterous manipulation, multi-robot coordination and rapid adaptation to new bodies.
This release brief was checked against first-party material on 10 August 2026. The date above is the public announcement date, not the date a repository was created or a third-party provider added the model. Where access or weights arrived later, that distinction is recorded below.
Release record
| Field | Verified detail |
|---|---|
| Announcement | 30 July 2026 |
| Availability or weight release | 30 July 2026, with ER 2 in AI Studio, enterprise private preview and VLA access for selected partners |
| Release type | proprietary vision-language-action and embodied-reasoning model family |
| Access | ER 2 public developer access plus private-preview and early-access paths; no downloadable weights |
| Architecture | a whole-body VLA, a vision-language embodied-reasoning agent and an efficient on-device VLA |
| Maximum stated context | task sequences lasting several minutes and hundreds of decisions; token limit not stated in the launch article |
What changed
Google released three named models with distinct deployment roles. Robotics 2 controlled multiple embodiments from one checkpoint; ER 2 planned and tracked long tasks; and On-Device 2 adapted to new bi-arm robots from less than 200 examples. The release also introduced ASIMOV-Agentic for embodied safety orchestration.
The practical comparison is therefore not simply whether Gemini Robotics 2, Gemini Robotics ER 2 and Gemini Robotics On-Device 2 has the largest headline score. Teams need to ask whether its architecture, access terms, latency, tool behaviour and evaluation setup match the workload they actually intend to run. A model can lead one harness while losing on cost, refusal behaviour, multilingual quality or repeatability in another.
Benchmarks worth retaining
| Evaluation | Reported result | How to read it |
|---|---|---|
| Embodiment transfer | new bi-arm robots adapted with less than 200 examples | First-party data-efficiency claim |
| Task horizon | several minutes and hundreds of decisions | Long-horizon embodied-agent capability |
| ASIMOV-Agentic | new safety benchmark introduced | Measures unsafe tool refusal, feasibility and human escalation |
These are release-time results, not independently reproduced guarantees. Several success-rate charts are platform- and task-specific, and access to the VLA checkpoints remains restricted, limiting independent reproduction. Scores should remain attached to the disclosed effort setting, agent harness, tool access, timeout, context-management policy and judge model. Moving a number into a procurement sheet without those conditions creates false comparability.
Architecture and access
Gemini Robotics 2, Gemini Robotics ER 2 and Gemini Robotics On-Device 2 is described as a whole-body VLA, a vision-language embodied-reasoning agent and an efficient on-device VLA with task sequences lasting several minutes and hundreds of decisions; token limit not stated in the launch article of stated context. Its access position at verification time is ER 2 public developer access plus private-preview and early-access paths; no downloadable weights. That wording matters: open weights, source-available weights, an API, a product preview and a research demonstration give adopters very different rights and different levels of reproducibility.
Before deployment, record the exact model identifier or checkpoint, inference stack, quantisation, reasoning setting, region, price schedule and supplier terms. If the release uses a custom licence, read the licence itself rather than relying on the word “open” in launch copy. If it is API-only, preserve the dated documentation and change-notice route because the served snapshot can change without a downloadable artefact.
What an evaluation should test next
For Gemini Robotics 2, Gemini Robotics ER 2 and Gemini Robotics On-Device 2, a credible internal gate should include:
- a frozen set of representative tasks with pass, fail and abstain criteria;
- a matched baseline using the same tools, timeout, prompt budget and reviewer rubric;
- repeated runs to expose variance rather than reporting a single best attempt;
- latency, token use and total task cost alongside task success;
- adversarial, multilingual and long-context cases relevant to the real deployment; and
- rollback evidence showing the previous model can be restored safely.
The wider model change-control guide explains how to keep model, prompt, tool and corpus changes reconstructable. The AI dependency inventory guide covers the release and supplier records needed after deployment.
AIEngine verdict
This is a major physical-AI model family and should not have been omitted. Evaluation must separate high-level reasoning, low-level control, embodiment transfer and safety-stop behaviour.
This is a launch assessment, not a certification. Benchmark leadership is useful evidence of where to test; it is not authorization to place the model in a high-impact workflow without domain evaluation, security review and an accountable owner.
Primary sources
- Google DeepMind Gemini Robotics 2 announcement
- Gemini Robotics ER 2 model card
- Gemini Robotics 2 safety report
Image provenance
Hero image: Google DeepMind official release artwork. The locally served WebP is a crop of the first-party release or model-card asset recorded in the repository provenance manifest.
