Google DeepMind's Gemini Robotics 2 Brings Whole-Body Intelligence and Multi-Robot Collaboration

Three models, one intelligence layer

On July 30, 2026, Google DeepMind announced Gemini Robotics 2, a trio of models designed to serve as the "intelligence layer" for physical AI. The lineup breaks down into three distinct roles:

• Gemini Robotics 2 (VLA) — a vision-language-action model that converts visual and language input into motor commands. This is the model that actually moves the robot. It can control anything from a two-armed tabletop platform to a full humanoid with 22-degree-of-freedom hands.
• Gemini Robotics ER 2 — the "embodied reasoning" model that acts as the robot's high-level brain. It understands the physical space, plans multi-step tasks lasting several minutes, communicates with humans, and coordinates with the VLA to execute actions. This is available today in Google AI Studio.
• Gemini Robotics On-Device 2 — a slimmed-down VLA that runs locally on robotic hardware, no cloud connection needed. Adapts to new robot bodies in under 200 examples and a few hours of training.

The ER 2 model replaces the ER 1.6 that launched in April, and the improvement is substantial: it consistently outperforms its predecessor across tool orchestration, progress tracking, and safety benchmarks.

What the VLA model can actually do

Let's skip the marketing and look at the numbers. DeepMind tested Gemini Robotics 2 on the Apptronik Apollo 2 humanoid with two different hand setups:

• Whole-body manipulation (Apollo with Inspire hands): 68.4% pick up from table, 45.7% pick up from floor, 76.3% pick up from shelf. These are reasonable numbers for a research model — competent enough to show the architecture works, but not ready for unsupervised deployment.
• Multi-finger dexterity (Apollo with SharpaWave 22-DOF hands): 92% on unscrewing a light bulb (the easiest task), but only 36% on screwing one in, 44% on tying a trash bag, and 32–40% on dustpan and ziplock tasks. Fine motor skills remain the hard problem.

Where it shines is on a simpler gripper platform. The Franka Duo — built by Munich-based Franka Robotics — hit 74.2% on general pick-and-place, 78.9% on diverse tool kitting, and 89.6% on precise insertion tasks. If your use case involves moving and inserting objects with a two-armed industrial platform, this model is already competitive.

The new capability worth paying attention to is whole-body control. Previous Gemini Robotics versions could only handle the upper body. The new model can make a humanoid walk, crouch, bend, and reach — dynamically balancing its centre of gravity rather than reaching from a static position. In the demo video, the Apollo 2 walks to a table, picks up a watering can, walks to a shelf unit, and places it in the correct bin. It's slow — no getting around that — but it's the first time we've seen this level of end-to-end whole-body coordination from a single model checkpoint.

The ER 2 reasoning model: the brain behind the movement

If the VLA is the muscle, the ER 2 is the brain. It's not a motor controller — it's a vision-language model that watches the robot's video feed, tracks progress through multi-step tasks, and decides what the VLA should do next. Think of it as the supervisor that keeps a 5-minute task from turning into a confused loop.

The benchmarks tell a clear story:

• Progress classification: ER 2 scores 57.4% accuracy on 5-level progress tracking (e.g., is the coffee 0-20% poured or 60-80% done?). That beats ER 1.6 and competing frontier models, though there's clearly room to grow.
• Moment-finding: 91.3% accuracy with a mean absolute distance of 0.96 seconds from the true event. This is the "stop pouring when the cup is full" problem, and the model delivers sub-second precision with 4× the execution speed of larger model categories. Crucially, this runs fast enough to operate in real physical time — sub-second latency is a hard requirement for real-world robotics.

ER 2 also introduces multi-robot collaboration. Different robot types — say, an Apptronik Apollo 2 humanoid and a Franka Duo tabletop arm — can share semantic understanding of a task, divide the labour, and hand off subtasks between each other. No single robot form factor can do everything, so this is a pragmatic solution.

Developers can try ER 2 right now via the Gemini API, Google AI Studio, or the Gemini Enterprise Agent Platform (private preview). Code examples for the Live API integration — including a working setup with Boston Dynamics Spot fetching objects on natural-language commands — are available on GitHub.

The European connection that matters

This story has a significant European dimension that most coverage misses.

First, the hardware partners. Two of the four launch partners are European companies. Franka Robotics (Munich, Germany) built the Franka Duo platform used to validate the gripper dexterity benchmarks — and those 89.6% insertion accuracy numbers were recorded on European hardware. Agile Robots, also Munich-based, is listed as a research partner for next-generation humanoid and general-purpose robotics. These aren't peripheral partnerships; DeepMind cited Franka and Agile Robots alongside Apptronik and Boston Dynamics as core collaborators.

Second, DeepMind itself. Founded in London and still headquartered there, Google DeepMind operates substantial AI research and engineering teams in the UK and across Europe. The robotics safety framework — including the new ASIMOV-Agentic benchmark for safety orchestration — was developed with contributions from the London lab.

Third, availability. Gemini Robotics ER 2 is accessible through Google AI Studio, which serves developers in all EU member states. There's no region lock, and the free tier allows initial experimentation. The VLA and On-Device models require a waitlist application with a trusted tester program for enterprise automation and robotics startups.

What the AI Act means for this

If you're building or deploying AI-powered robots in the EU, the EU AI Act applies here. Physical robots equipped with AI that can make autonomous decisions in the real world fall squarely under the Act's high-risk category. The requirements are non-trivial: mandatory risk assessments, human oversight mechanisms, and conformity assessments before CE marking.

DeepMind's timing is interesting here. They published a detailed safety technical report alongside the model release, documenting performance on safety instruction following and human proximity benchmarks. ER 2 is their safest robotics model yet — it can halt a humanoid when a person is nearby and resume only once the area is clear. The ASIMOV-Agentic benchmark specifically tests the ER model's ability to refuse unsafe tool calls from a VLA and to request human intervention when uncertain. These are precisely the kinds of safeguards that EU regulators will want to see.

For European robotics companies and automation integrators, having safety benchmarks and documentation ready at launch — rather than scrambling to comply after deployment — is a meaningful advantage.

How this stacks up against the field

Google DeepMind isn't alone in embodied AI. Figure AI raised $1.5 billion in March 2026 with its humanoid development. Tesla's Optimus continues to generate marketing material but no published benchmarks. Nvidia's Project GR00T provides an alternative simulation-to-reality pipeline. Meanwhile, the open-source robotics community has made real progress with RL-based approaches on platforms like Unitree and Agility Robotics hardware, though generalisation across embodiments remains a weak point.

Where Gemini Robotics 2 differentiates itself is the unified model architecture: a single VLA checkpoint that can drive three different robot embodiments (Apollo, Franka Duo, On-Device platforms) and an embodied reasoning layer that coordinates between them. Nvidia's approach is more modular and simulation-heavy; Tesla's is vertically integrated into their own hardware. DeepMind's bet is that a shared intelligence layer, adapted to different bodies through lightweight fine-tuning, scales better than building a separate AI stack for each robot model.

The pricing story is incomplete. ER 2 is in free preview on Google AI Studio with no announced rate limits or production pricing. The VLA and On-Device models are in private preview with no published costs. For European companies considering integration, the sensible move is to start prototyping with ER 2 in AI Studio while waiting for commercial terms.

What this means for developers today

If you want to get hands-on with this:

1. Start with ER 2 in AI Studio. It's live, free to experiment with, and the Gemini Live API integration means you can prototype a "conversational robot brain" that plans and reasons about physical tasks without owning any hardware. The getting-started notebook walks through configuration and prompting.
2. Apply for the VLA waitlist if you have hardware (or a hardware partner). The trusted tester program targets both enterprise automation teams and robotics startups — European companies should mention their EU data residency and regulatory compliance needs in the application.
3. Read the safety report. If you're building for the EU market, the ASIMOV-Agentic benchmark and safety metrics are directly relevant to AI Act conformity documentation.

The bottom line: Gemini Robotics 2 is the most complete robotics AI release from a major lab we've seen, and it comes with European hardware partnerships, EU-accessible deployment paths, and publicly documented safety benchmarks. That combination doesn't happen often.