Physical AI: Gemini Robotics 2 Gives Robots Whole-Body Control — And ER 2 Is Already Public

If you have been tracking new AI tools in 2026, the most consequential launch of the past week did not happen in a chat window. On 30 July, Google DeepMind introduced Gemini Robotics 2 — an intelligence layer for physical machines — and on 2 August the reporting wave hit, with roboticists dissecting demos of a humanoid walking, crouching, stretching and tidying a cluttered room without task-specific programming. Its companion model, Gemini Robotics ER 2, is already in public preview through the Gemini API and Google AI Studio, which means this is not a sealed research artefact. You can call part of it today.

This article explains what the three new models actually do, why "whole-body control" is the technical headline, what the multi-robot collaboration demo proves and does not prove, how to try the embodied reasoning model yourself, and what the release means for anyone buying or building AI tools in the next twelve months.

What Google DeepMind Announced

The DeepMind announcement covers three models that are designed to operate as one system:

• Gemini Robotics 2 — a vision-language-action model (VLA). It converts camera input and natural-language instructions directly into motor commands. This is the model that controls full humanoids from feet to fingertips, as well as conventional bi-arm robots and grippers.
• Gemini Robotics ER 2 — an embodied reasoning model (a VLM, not a VLA). It talks to humans, interprets the physical scene, plans multi-step tasks lasting several minutes, and then hands motor execution down to a VLA. Announced separately on the Google blog, it is the piece you can call from the Gemini API right now.
• Gemini Robotics On-Device 2 — a lightweight VLA optimised to run locally on the robot, with no round trip to a data centre.

The design choice worth internalising is the split. Earlier robot-learning systems tried to make one network do both jobs: understand the goal and generate the joint torques. Separating them means the slow, expensive reasoning can run at the pace of planning while the fast, cheap control loop runs at the pace of physics. DeepMind explicitly notes that ER 2 can "think" about the next step while the robot is still executing the current one — a design pattern that agent developers will recognise immediately from software orchestration.

Why Whole-Body Control Is the Real Headline

Almost every impressive robot manipulation demo of the past three years was a table-top demo. A fixed base, two arms, a bounded workspace. That constraint is not laziness — it removes balance from the problem. The moment a robot has legs, every reach becomes a stability question: shifting weight to extend an arm can put the machine on the floor.

Gemini Robotics 2 claims to reason across the whole kinematic chain, coordinating legs, torso, arms and fingers as a single system. DeepMind's demonstrations show a humanoid stepping into position, crouching to reach a low shelf, and stretching to place an object high — motions that require the model to plan a posture, not just a grasp.

Independent coverage from The Robot Report notes the model driving a five-fingered, 22-degree-of-freedom hand on Apptronik's Apollo 2 humanoid. That number matters more than the marketing language around it. Twenty-two degrees of freedom in one hand is a far harder control problem than a parallel-jaw gripper, and it is the difference between "can move a box" and "can handle the objects that exist in a kitchen."

Cross-embodiment transfer in hours, not months

The second technical claim deserves as much attention as the first: DeepMind says the model adapts to entirely new robot bodies in a few hours. Historically, transferring a learned policy from one platform to another meant retraining, because the mapping from intent to joint angles is hardware-specific.

If cross-embodiment adaptation genuinely takes hours, the economics of the robotics industry change shape. Hardware makers stop needing in-house foundation-model teams, and the intelligence layer becomes a purchased component — exactly the pattern that played out in software when teams stopped training their own language models. It is the same platform logic we described in our guide to AI agents and workflow automation, applied to machines with actuators.

Multi-Robot Collaboration: What the Demo Actually Shows

The most quoted capability is multi-robot collaboration. In DeepMind's framing, several robots in a shared space can communicate, recognise each other's differing physical strengths, and delegate sub-tasks autonomously to finish a mission that no single machine could complete alone.

Read that carefully. The claim is task delegation between heterogeneous robots, not a swarm of identical units running the same policy. A wheeled robot with a long reach and a humanoid with dexterous hands have genuinely different capability profiles, and a planner that can assign work accordingly is doing something closer to project management than to path planning.

What the demos do not establish is reliability under adversarial conditions: unexpected humans in the workspace, hardware faults mid-task, or conflicting instructions from two operators. Those are the failure modes that decide whether a system reaches a factory floor, and no public benchmark meaningfully covers them yet.

The 91.3% number, in context

Coverage from Pulse 2.0 highlighted ER 2 reaching 91.3% on a moment-finding task — locating the precise instant in a continuous video feed when something relevant happened. It is an unglamorous metric that quietly unlocks a lot.

Continuous video understanding is what lets a robot track its own progress: it can tell that the cup is now on the shelf, that the previous attempt failed, and that it is time to move on. DeepMind calls this success tracking, and describes it as a step change over ER 1.6. In practical terms, it converts "run the script and hope" into a closed loop with retries — the same shift that made software agents usable.

One caveat worth stating plainly: a single-number score on a proprietary evaluation is a signal, not a guarantee. Treat it the way you would treat any vendor benchmark, using the framework in our piece on how to read AI coding benchmarks.

How to Try Gemini Robotics ER 2 Today

You do not need a robot to experiment with the reasoning model. According to DeepMind's model page, ER 2 is in public preview on Google AI Studio and the Gemini API, with private preview on the Gemini Enterprise Agent Platform. Supported capabilities include Search grounding, function calling, code execution, structured output, URL context, and the Live API for text output.

A realistic first experiment: feed it a photograph or short video of a physical space and ask for a plan. Because it returns structured output and calls functions, its plans can drive things other than robot joints — a warehouse task queue, a camera-based inspection routine, a compliance checklist.

1. Start with spatial questions. Ask it to identify objects, describe their relationships and flag hazards. This tests the perception layer before you trust the planning layer.
2. Then ask for a multi-step plan with explicit success criteria per step. Success tracking is the feature; make it declare what "done" looks like.
3. Wire in one tool call. Function calling is where reasoning becomes action, and it is the same integration pattern used by the Model Context Protocol tooling that dominates the agent ecosystem — see our review of MCP servers that ship 67 tools from one endpoint.
4. Log everything. Physical-world evaluation is unforgiving; you want the video, the plan and the outcome for every trial.

What This Means for the AI Tools Market

Three shifts follow from this release, and none of them are about humanoids specifically.

1. The intelligence layer is separating from the hardware

If one model family can drive bi-arm robots, humanoids and grippers with hours of adaptation, then robot makers compete on actuators, cost and reliability while the reasoning is bought in. That is good for hardware startups and difficult for anyone whose differentiation was a proprietary control stack.

2. Physical-world agents inherit software-agent problems

Multi-step plans, tool calls, retries and success criteria are exactly the primitives that software agents use — with the added detail that a bad action has mass. Prompt injection through a physical scene is not science fiction: a printed instruction taped to a box is untrusted input. The OWASP Top 10 for LLM applications reads differently when the executor has arms.

3. On-device inference becomes a compliance feature, not just a latency one

Gemini Robotics On-Device 2 exists because a control loop that depends on network round trips is fragile. It also means camera data need not leave the building, which is the argument that gets robotics past a privacy review. For teams in Europe, high-risk classification under the EU AI Act is a live question for machines operating around people, and local processing simplifies parts of that story.

Who Should Actually Care Right Now

Being honest about the audience is more useful than universal hype.

• Robotics teams with existing hardware. You are the primary audience. The cross-embodiment claim is directly testable on your platform, and it either saves you a year of policy work or it does not.
• Warehouse, logistics and manufacturing operations leads. Worth a pilot conversation, not a purchase order. Ask vendors which intelligence layer they use and what their retry behaviour looks like when a task fails halfway.
• Agent developers with no robots at all. ER 2 is a genuinely interesting planner with video understanding and success tracking. Those capabilities are useful for inspection, monitoring and any workflow where the ground truth is a camera feed.
• Enterprise buyers of general AI tools. Mostly a signal to watch. The relevant lesson is architectural: reasoning and execution are separating into distinct layers, which is how you should structure your own automation too. Our AI for business hub covers that evaluation process.
• Consumers waiting on a household robot. Still waiting. Nothing here is a shipping consumer product, and tidying a room in a demo environment is a long way from a stranger's kitchen.

The Honest Limitations

Four things are missing from the public picture, and each one should temper the enthusiasm.

No failure rates. Demo videos show successes. The number that matters for physical deployment is the rate of task completion across hundreds of unscripted attempts, and it is not published.

No latency or compute figures for the VLA. Whole-body control is a real-time problem. Without published control-loop frequencies, it is impossible to judge how much of the capability survives outside a lab network.

Limited availability. ER 2 is in preview; Gemini Robotics 2 itself is offered through partnerships rather than a self-serve endpoint. Access, not capability, is the near-term constraint for most teams.

Safety framing is early. DeepMind discusses responsibility, but embodied AI safety needs standards and third-party testing, not vendor policy documents. That work is genuinely unfinished across the industry.

The Bottom Line

Gemini Robotics 2 is the clearest sign yet that the frontier-model race has a second front. For two years, competition ran on text, code and video generation. This release argues that the next competitive dimension is the physical world, and that the winning position is the same one that worked in software: own the reasoning layer, make it portable across hardware, and let everyone else build the bodies.

For most readers, the practical action this month is small and concrete. Open Google AI Studio, hand ER 2 a photograph of a real space, and ask it for a plan with success criteria. You will learn more about where embodied AI actually stands in twenty minutes than from a week of demo videos. Then keep following the story in our generative AI news and AI agents and automation hubs, where we track each release against what it can verifiably do.