Google has announced two physical-AI models: Genie 4, a world model that generates steerable scenes with sound, is limited to trusted testers, while Gemini Robotics ER 2 is in public preview for Google Cloud customers. The contrast matters: one is a striking research-facing demonstration behind a gate, the other a robot-planning model developers can start trying.
Google DeepMind Watch analysis
What happened
The announcements came on 27 October, according to OrcaRouter’s account. Genie 4 can generate audio alongside scenes and respond to instructions while running. Google has not published specifications such as resolution, frame rate, latency or price, and OrcaRouter says access is limited to a group of trusted testers.
Gemini Robotics ER 2 is available in public preview through Google Cloud’s Model Garden and Agent Studio. It is designed to plan tasks over tools such as robot-control and navigation systems, and to track progress so a robot can retry when a step fails. OrcaRouter reports Google’s figures of 91.3% accuracy for pinpointing critical video frames and safety scores of 97.9% for instruction following and 93.0% for human proximity. Those are vendor-reported results, not independent evaluations.
Key findings
- Genie 4 creates scenes with sound
It can generate ambient audio, music and dialogue alongside video, and accept instructions while a scene is running. - Genie 4 remains tester-only
No public price or technical specification is provided in the account, leaving developers without a clear way to assess or budget for it. - ER 2 plans across robot tools
Developers can connect low-level control interfaces, while the model plans tasks and monitors progress. - ER 2 is the accessible preview
Google Cloud customers can try it through Model Garden and Agent Studio, although the account gives no rate card.
Why it matters
Together, the models suggest two complementary pieces of a physical-AI stack: generated environments for simulated experience, and a planner that can orchestrate actions on robots. That is a promising direction, but the practical gap is already visible. ER 2 has a public preview route; Genie 4, the more novel world model, does not.
The safety figures need careful reading. OrcaRouter attributes them to Google, and says physical safeguards such as emergency stops and barriers remain the operator’s responsibility. Strong scores for model reasoning are not a substitute for those protections.
Our read
The most useful news is not simply that Google has two new model names. Developers can try ER 2 now, while Genie 4 remains out of reach for most of them. Google’s account of both is intriguing, but missing specifications and independent measurements mean the demos are a starting point, not a procurement case. We’d begin with the public preview and keep the emergency stop very much in the room.
What to watch
- Whether Google publishes Genie 4 specifications, pricing or broader access.
- Whether ER 2 gets a rate card and a general-availability date.
- How the models perform in independent tests and real robot workflows.
- Whether the two systems are integrated beyond the examples described.
Discussion spark: For developers working on physical AI, should Google prioritise opening Genie 4 to more testers, or make ER 2’s performance and pricing easier to evaluate first?
Sources and evidence
- Daily AI Brief, Oct 28: Genie 4 Arrives Gated as Gemini Robotics ER 2 Opens in Preview – OrcaRouter (29 October 2026, 14:43 UTC)
not affiliated with, endorsed by, or operated by Google or Google DeepMind