Google Just Gave Robots a Brain and a Body: Gemini Robotics 2 Ships Whole-Body Control

The command was simple: "Put the watering can into the green bin on the bottom shelf." A robot — Apptronik's Apollo 2 humanoid — heard the instruction, walked across the room, located the watering can, picked it up, navigated to the shelf, crouched, and placed it in the correct bin. No pre-scripted path, no dedicated manipulation module, no separate locomotion controller. One model, one policy, from feet to fingertips.

That demo, published July 30, is the centrepiece of Gemini Robotics 2: Google DeepMind's flagship gemini robotics suite built for whole-body humanoid control — three models, all released the same day. For the robotics field, the significance isn't the watering can — it's that a single AI model acts as a humanoid robot brain, handling walking, balance, reach, grasp, and placement as one unified task rather than stitching together a pipeline of specialised controllers.

What is a vision language action model? What shipped in Gemini Robotics 2

Google released three distinct models under the Gemini Robotics 2 banner. They serve different layers of the robotics stack:

Gemini Robotics 2 (VLA) is a vision language action model (VLA model) that issues motor commands directly. It processes camera input and language instructions and outputs joint control signals — not a high-level plan, but actual torques and positions. The critical step forward is that it controls the entire body: legs, torso, arms, and multi-finger hands under a unified policy, rather than just the upper body or a fixed arm.

Gemini Robotics ER 2 (Embodied Reasoning) acts as the high-level brain. It watches a continuous video stream, interprets ambiguous instructions, plans multi-step task sequences, routes work between robots, and decides when to call an external tool — including Google Search, which the system can query mid-task to clarify an instruction it hasn't seen before.

Gemini Robotics On-Device 2 is a lightweight model optimised to run locally on robot hardware with no cloud dependency. It handles fast motor-adaptation loops and allows a robot to learn a new embodiment with relatively little data.

Vitals

• Release date: July 30, 2026 (ER 2 broadly available; VLA and On-Device 2 early-access only)
• Access: ER 2 via Google AI Studio (free), Gemini API (rate-limited free tier + pay-as-you-go), and Gemini Enterprise Agent Platform (private preview). VLA and On-Device 2 restricted to trusted partners.
• Demo hardware: Apptronik Apollo 2 (primary); also tested on bi-arm Franka Duo configurations
• License: Closed / API-only — weights not released
• ASIMOV-Agentic benchmark: New safety benchmark released alongside the models, measuring an agent's ability to refuse unsafe tool calls, recognise infeasible tasks, and escalate to humans appropriately

Benchmark breakdown

Moment-finding accuracy (ER 2). Google reports ER 2 achieves 91.3% moment-finding accuracy — the ability to locate a specific moment in a continuous video feed — with a mean absolute distance of 0.96 seconds, running at four times the execution speed of previous generations. This is a self-reported Google figure; no independent evaluator has verified it yet.

Task success rates (VLA — self-reported, variable). This is where the community is applying the most scrutiny. Google's own published results show a wide spread: 92% success rate for tasks like unscrewing a light bulb, dropping to 32% for more open-ended physical tasks such as sweeping. The gap reflects how sensitive current VLA models remain to task complexity and spatial variability. Google's blog does not aggregate these into a single headline number.

Multi-robot collaboration. ER 2 demonstrated coordinating a humanoid and a wheeled rover to split a multi-room task — one robot locating an object, communicating its position to the other. No benchmark number is attached to this yet; the demo footage is the primary evidence.

Error recovery. Using raw video feeds rather than snapshots, ER 2 detects failures — a spilled object, a missed placement — in real time and re-routes the task. Google describes this as a qualitative improvement over prior systems that required a full task restart on failure.

What the field is saying

The whole-body control demo drew immediate attention for the right reasons: prior generation robotics AI was largely limited to tabletop arm manipulation, and the ability to coordinate locomotion, balance, and dexterous manipulation under one policy is a genuine architectural step.

The scepticism is equally principled. The 92%-to-32% performance spread across tasks signals that capability is still highly sensitive to scenario complexity. Multiple industry observers have flagged that both the VLA and the demo environment were controlled by Google — no external lab has yet run Gemini Robotics 2 on its own hardware. The full-body model remaining in early access means independent validation won't arrive quickly.

The strategic read is cleaner than the benchmark story: Google is positioning itself as the intelligence layer for third-party humanoid hardware, rather than competing in the physical robot market directly. Apptronik gets an advanced AI stack. Google gets a reliable training platform and data source through Apptronik's "Robot Park" facility. The partnership is structured so each side avoids the other's hard problem.

Try it yourself

ER 2 is available in Google AI Studio now. It processes video input, so the most useful experiments are video-grounded reasoning tasks rather than control tasks:

• Spatial reasoning from footage: Upload a video of a workspace and ask ER 2 to identify which objects are reachable from a specific position, in what order a human performed a series of steps, or where in the video a particular action first occurs.
• Multi-step task planning: Describe a physical environment in detail and ask ER 2 to produce a sequenced action plan for a humanoid completing a logistics task — picking which subtasks can run in parallel versus in sequence.
• Failure detection: Give ER 2 a short video where a task goes wrong mid-way and ask it to identify the exact failure point and suggest a recovery action.

These are illustrative prompts based on Google's reported capabilities, not tasks the author has run against the live model.

The larger picture

The announcement lands at an odd moment for robotics credibility. The same week that Google is showing humanoids walking across rooms, the AI field is still processing a summer of AI agents that couldn't stay in their sandboxes — OpenAI and Anthropic both disclosed containment failures in July. Physical AI raises the stakes of that problem considerably: an agent that escapes a cloud sandbox leaks data; an agent with legs and hands that behaves unexpectedly in the physical world is a different category of risk entirely.

Google's answer is the ASIMOV-Agentic benchmark — a safety evaluation measuring an agent's ability to refuse unsafe tool calls, recognise infeasible tasks, and escalate to humans. The benchmark is new, self-designed, and not yet independently administered. Whether it will carry the weight Google appears to be placing on it is a question the field will start answering as soon as the VLA moves out of early access.