Google Physical AI: ER 2, On-Device 2 and Partners

Google's latest physical AI portfolio has two distinct model layers. Gemini Robotics ER 2 is a public-preview vision-language model for high-level reasoning, tool use and progress monitoring. Gemini Robotics On-Device 2 is a locally running vision-language-action model that generates numerical robot actions and is currently limited to trusted testers. Robot demonstrations with Boston Dynamics, Apptronik and Franka show how those ideas can reach different bodies.

The useful way to read this strategy is as a stack, not as a single product announcement. Google is developing models that can understand an environment, decide which capability to call and translate instructions into movement. Our physical AI see-think-act guide explains that loop; this article maps Google's current products, previews and research evidence onto it.

Google's physical AI stack in one view

At the top of the stack, ER 2 interprets goals and scenes, builds plans, calls tools and checks whether a task is progressing. At the action layer, On-Device 2 turns instructions and sensor context into numerical actions on a specific robot. Under both sit robot controllers and independent safety systems that handle timing, limits and emergency behavior.

The three layers should not be collapsed into the word intelligence. A model can reason well and still fail because a gripper cannot reach an object. An action policy can move smoothly while misunderstanding the goal. A reliable system needs explicit interfaces between reasoning, action, control and safety.

Component	Role	Output	Availability on August 6, 2026
Gemini Robotics ER 2	High-level embodied reasoning and orchestration	Text and function calls	Public preview via Gemini API and Google AI Studio
Gemini Robotics On-Device 2	Local robot action policy	Numerical robot actions	Trusted testers and waitlist
Apptronik Apollo 2 and Franka F3 Duo	Multi-robot research demonstration	Physical execution through their own interfaces	Demonstration, not a Google robot product
Boston Dynamics Spot	Mobile manipulation demonstration	Navigation and arm execution through Spot APIs	Research demo, not an announced commercial integration

ER 2 is the high-level reasoning and supervision layer

Gemini Robotics ER 2 is based on Gemini 3.5 Flash and accepts text, images, video and audio. It returns text, including structured function calls when the application provides tools. That output can tell a robot application to navigate, inspect or manipulate, but ER 2 does not directly emit joint commands. The application remains responsible for checking arguments and routing approved calls to the correct machine.

The emphasis is continuous video reasoning. A robot can report its changing view while the model tracks whether a step succeeded or a retry is needed. Google's reported progress and moment-finding results are defined evaluations, not universal field success rates. Read our Gemini Robotics ER 2 explainer for details.

On-Device 2 brings numerical actions closer to the robot

Gemini Robotics On-Device 2 addresses a different constraint: robots cannot always depend on a stable cloud connection. The model is designed to run locally, using text, camera images and robot proprioception to produce numerical actions. Local inference can reduce network dependence and keep the action loop close to the machine, although total latency and safety still depend on the hardware and controller.

Google says the model can adapt to a new robot with fewer than 200 demonstrations, or a few hours of data collection in its examples. This is a Google-reported capability, not a guarantee. The model card notes limits on out-of-distribution tasks and high-degree-of-freedom machines. Our On-Device 2 guide covers the evidence.

NIST robotic test facility — Physical AI evaluation depends on the site, sensors, procedures and hardware as well as the model. This is not a Google facility. Source: NIST / A. Eustis via Wikimedia Commons. License: Public domain, U.S. Government work.

Robot partners test whether one model can cross embodiments

Google's partner demonstrations are evidence about portability, not a catalog of Google-made robots. With Apptronik Apollo 2 and Franka F3 Duo, ER 2 showed high-level task decomposition across machines with different physical strengths. With Boston Dynamics Spot, it selected and monitored navigation and manipulator functions while the robot fetched a snack in response to a natural-language request.

These examples matter because warehouses, laboratories and service environments contain heterogeneous equipment. A common reasoning layer could choose among approved capabilities without requiring users to learn every device command. Each robot still needs an adapter, a defined tool schema and its own certified controls. A successful demo does not make the underlying models interchangeable with every arm, quadruped or humanoid.

The Spot and ER 2 fetch-demo analysis reconstructs the high-level sequence and explains why it should not be described as a Boston Dynamics product launch.

Robot or system	What the demonstration explored	What it did not establish
Apptronik Apollo 2	Humanoid participation in a coordinated multi-robot task	Commercial Google-Apollo integration or general humanoid autonomy
Franka F3 Duo	Precise manipulation as part of an orchestrated workflow	Universal compatibility with every Franka setup
Boston Dynamics Spot	Natural-language planning, Spot API calls and progress monitoring	A new Spot model, a public deployment or certified autonomous fetching
On-Device 2 test platforms	Local action generation and adaptation to different embodiments	General access or guaranteed transfer with fewer than 200 examples

The strategy combines general reasoning with embodiment-specific action

Physical AI faces a tension between generality and precision. Language and vision models can reason across many scenes, but safe movement depends on a robot's geometry, sensors, payload and control frequency. Google's split architecture acknowledges that tension: ER 2 aims to generalize at the semantic and planning level, while an on-device VLA or robot API handles embodiment-specific action.

This also explains why the meaning of a VLA differs from the role of ER 2. A VLA maps observation and instruction to actions. A high-level VLM decides which goal-directed action or tool should be attempted. In a mature application, conventional planning, learned policies and human approval may all sit between those endpoints.

Modularity still creates integration work: developers need consistent coordinate frames, reliable tool results, clear error codes and a shared definition of completion.

Decision card summarizing the key decisions in Google Physical AI: Gemini Robotics ER 2, On-Device 2 and Robot Partners — A Physical AI Lab editorial card reconstructed from official Google sources. Source: Physical AI Lab. License: Owned original.

Availability is uneven across the portfolio

ER 2 is the part most developers can try immediately. Google offers a standard preview model and a streaming preview model through the Gemini API, with free and paid tiers. Enterprise Agent Platform access is described as a private preview. Preview status means interfaces, prices and limits can change, so experiments should isolate model calls behind an adapter.

On-Device 2 is not a public download or generally available API. Google directs interested teams to a trusted-tester waitlist. The partner robot demonstrations also do not create a bundled purchasing route. A company evaluating Spot, Apollo or Franka hardware must still work through the relevant manufacturer and its supported software stack.

The access matrix below is a practical checkpoint for planning.

Need	Current route	Important caveat
Test high-level robot reasoning	ER 2 in Google AI Studio or Gemini API	Public preview; validate privacy, cost and reliability
Build a low-latency streaming supervisor	ER 2 streaming preview through Live API	Not a deterministic motor-control loop
Run Google's new action model locally	Apply to On-Device 2 trusted-tester program	No general public access
Reproduce a partner demonstration	Use documented APIs and contact the relevant vendor	Demo assets and integrations may not be publicly packaged

What Google's announcements do and do not prove

The announcements support a clear direction: Google is investing in multimodal reasoning that can observe task progress, call robot tools and work across embodiments, while also developing a local action model. They show credible research progress and provide a public ER 2 experimentation path. They do not prove that general-purpose robots are ready for unsupervised deployment in open environments.

Safety remains layered. Google's material covers semantic and orchestration risks, but not functional-safety engineering, guarded workspaces or emergency stops. Robot video and audio may include identifiable people, creating notice, consent and minimization obligations. Buyers should start with one bounded workflow and compare it with a deterministic process.

Frequently asked questions

Does Google make the robots shown with Gemini Robotics?

No. Boston Dynamics Spot, Apptronik Apollo 2 and Franka systems come from their respective manufacturers. Google supplies or researches the AI models and orchestration layer used in the demonstrations.

Which Google physical AI model is publicly accessible?

Gemini Robotics ER 2 is in public preview through the Gemini API and Google AI Studio. Gemini Robotics On-Device 2 is limited to trusted testers as of August 6, 2026.

Can ER 2 and On-Device 2 be used together?

They occupy complementary conceptual layers, so a system could use high-level ER 2 planning and an on-device action model. Google has not promised a universal plug-and-play integration; robot-specific adapters, validation and safety controls remain necessary.

Official sources checked

Last checked: August 6, 2026