Gemini Robotics 2 is Google DeepMind's new suite of robot foundation models, announced on July 30, 2026, that extends AI control from a robot's tabletop grippers to its entire body — feet to fingertips. It ships as three models: a vision-language-action (VLA) model that turns camera frames and spoken instructions into motor commands, an embodied-reasoning model (ER 2) that plans and supervises multi-minute tasks, and an on-device VLA that runs locally on the robot. The headline claim is that a humanoid can now walk, crouch, bend, manipulate objects and coordinate with a second robot autonomously, and that the stack adapts to an entirely new robot body in hours rather than months.
That last number is the one worth sitting with. The bottleneck in robotics has never really been "can a model pick up a cup." It has been that every new chassis, every new gripper, every new camera placement meant retraining from something close to scratch. If a foundation model can port to a new embodiment in hours, the economics of the entire humanoid industry change — and that is a much bigger story than another lab demo of a robot folding laundry.
This article breaks down what actually shipped, how the three models divide the work, what the demos do and do not prove, and where Gemini Robotics 2 sits against the rest of the physical-AI field in 2026.
Key Takeaways
- Google DeepMind announced Gemini Robotics 2 on July 30, 2026, adding whole-body control, advanced dexterity and multi-robot collaboration to its robot foundation model line.
- Three models ship together: Gemini Robotics 2 (VLA), Gemini Robotics ER 2 (embodied reasoning), and Gemini Robotics On-Device 2 (local VLA).
- The system was demonstrated on Apptronik's Apollo 2 humanoid performing full-body autonomous movement, not just seated arm manipulation.
- ER 2 acts as the planner: it observes the scene, decomposes a task into steps, and makes hundreds of decisions across tasks lasting several minutes.
- DeepMind says the stack adapts to new robot bodies in hours — the most commercially significant claim in the release.
What is Gemini Robotics 2 and what changed?
Gemini Robotics 2 is a family of three robot foundation models from Google DeepMind that convert vision and natural-language instructions into physical motor control across a robot's whole body. The previous generation focused mainly on upper-body manipulation from a fixed base. Version 2 adds locomotion, balance and coordination, so a humanoid can move through an environment and act on it in a single continuous policy rather than handing off between a walking controller and a grasping controller.
According to Google DeepMind's announcement, the model drives full humanoids as well as conventional two-armed robots, working with both multi-finger hands and simple parallel grippers. That hardware-agnosticism matters more than it sounds. Most robot learning research is quietly tied to one arm, one hand, one lab. A policy that transfers across grippers is a policy that a manufacturer can actually buy.
The demo footage published alongside the release shows tasks that are unglamorous on purpose: screwing in a lightbulb, tying a garbage bag, tidying a cluttered room. These are chosen because they require sustained multi-step reasoning plus fine force control — the two things robots historically fail at simultaneously.
How the three models divide the work
The architecture is a deliberate split between thinking and acting, and understanding it explains most of what the demos show.
| Model | Role | Runs where | What it outputs |
|---|---|---|---|
| Gemini Robotics 2 | Vision-language-action policy | Cloud / on-robot compute | Motor commands from camera + instruction |
| Gemini Robotics ER 2 | Embodied reasoning, planning, supervision | Cloud | Step decomposition, progress tracking, instructions to the VLA |
| Gemini Robotics On-Device 2 | Efficient local VLA | Fully on-device | Motor commands with no network dependency |
ER 2 is the piece that gets underrated. As described in Google's ER 2 write-up, it observes the environment, formulates a work procedure, tracks progress, and instructs the action model — making hundreds of decisions during tasks that run for several minutes. That is the difference between a robot that can execute a verb and a robot that can finish a chore.
The on-device variant is the practical one. Latency and connectivity are real constraints on a factory floor, and a policy that degrades gracefully to local inference is far more deployable than one that needs a round trip to a data center for every control step. It is the same architectural instinct you see in running LLMs locally — the cloud model sets the ceiling, the local model sets the floor of what still works when the network does not.
Is Gemini Robotics 2 actually better at dexterity?
Yes on the demonstrations, with an important caveat: dexterity gains are shown, not benchmarked against a public standard. DeepMind reports advanced dexterity across both multi-finger hands and parallel grippers, and Bloomberg's coverage framed the release explicitly around robots "struggling with dexterity" — the honest read is that this is progress on a still-unsolved problem, not a solved one.
Robotics does not yet have a Terminal-Bench or an MMLU. There is no shared, adversarial, third-party evaluation that lets you say model A beats model B by N points on manipulation. That absence is why every robotics launch in 2026 is judged on video, and video is the easiest modality in the world to cherry-pick. When we evaluate claims like this in practice, the questions that matter are: how many takes, what was the success rate across trials, and did the robot recover from failure or did the clip cut?
To DeepMind's credit, the multi-robot collaboration demo is harder to fake than a single-arm clip, because two robots sharing a task have to handle each other's timing errors in real time.
The Apptronik partnership and why the hardware partner matters
The showcase robot is Apptronik's Apollo 2. Per Robotics & Automation News, Apollo 2 running Gemini Robotics 2 performs full-body autonomous movement — walking, crouching, bending, manipulating — while reasoning through complex tasks in real time.
This is a strategic choice, not just a marketing one. Google DeepMind is positioning itself as the brain layer and letting partners own the body layer. It is the Android playbook applied to robotics: don't manufacture the hardware, make sure every serious manufacturer runs your stack. If the "adapts to a new body in hours" claim holds up under third-party testing, that strategy gets very strong very fast, because the switching cost for a robot OEM collapses.
The counter-position is Tesla, which is building both brain and body in-house, and the Chinese humanoid makers, who are competing primarily on cost per unit. Those are three genuinely different bets on where the value in physical AI accumulates.
How Gemini Robotics 2 fits the broader AI hardware race
Robot foundation models are compute-hungry in a way that connects directly to the infrastructure story we have been tracking. Training a whole-body VLA means video-scale multimodal data and a lot of accelerator time, and inference for the cloud-side ER 2 model is a per-robot, per-second cost, not a per-query one.
That puts robotics squarely inside the same squeeze driving the custom AI chip revolution and South Korea's $880B AI chip investment plan. A fleet of a thousand humanoids running continuous multimodal inference is a data center workload with legs attached. The on-device model is not just a latency optimization — it is a cost strategy, and the labs that get local inference right will be able to deploy fleets that cloud-only competitors cannot afford to run.
There is also a safety dimension that the announcement handles quietly. An LLM that misreads an instruction produces a bad paragraph. A whole-body policy that misreads an instruction produces 60 kilograms of moving metal. The industry's evaluation practices have not caught up with that asymmetry, and this is exactly the same structural gap we flagged in AI browser agents and their security holes: capability is shipping faster than the harness that verifies it.
The gap nobody is talking about: fleet learning economics
Here is the angle missing from most coverage. "Adapts to new bodies in hours" is being read as a research result. It is really a business model result.
If adaptation is cheap, then every robot deployed becomes a data source that improves the shared policy, and the marginal value of each additional deployment rises rather than falls. That is a compounding loop, and it is the same loop that made cloud LLMs defensible. Whoever gets to a large, diverse, continuously-operating fleet first does not just have more robots — they have a policy nobody else can replicate without matching the fleet.
Which means the number to watch over the next 12 months is not benchmark scores. It is deployed robot-hours running Gemini Robotics 2 in unstructured environments. DeepMind has not published that figure. Until it does, treat the capability claims as promising and the moat claims as unproven.
Frequently Asked Questions
What is Gemini Robotics 2?
Gemini Robotics 2 is Google DeepMind's robot foundation model suite announced on July 30, 2026, which gives humanoid and two-armed robots whole-body control, advanced dexterity and multi-robot coordination. It includes a vision-language-action model, an embodied reasoning model called ER 2, and an efficient on-device variant.
How is Gemini Robotics 2 different from the first version?
The first generation focused on upper-body manipulation from a fixed base. Gemini Robotics 2 adds whole-body control, so a humanoid can walk, crouch and bend while manipulating objects under a single policy. It also adds multi-robot teamwork and a local on-device model.
Which robots run Gemini Robotics 2?
The flagship demonstration platform is Apptronik's Apollo 2 humanoid. DeepMind states the model can also drive conventional two-armed robots using either multi-finger hands or standard parallel grippers, and can adapt to new robot bodies in hours.
Can Gemini Robotics 2 run without an internet connection?
Partially. Gemini Robotics On-Device 2 is an efficient VLA optimized to run locally on robot hardware, so basic action execution continues offline. The ER 2 reasoning model, which handles long-horizon planning, is the cloud-side component.
Is Gemini Robotics 2 available to developers?
Access is currently through Google DeepMind's robotics partner program rather than a general public API. Compare that to open-weight releases like Thinking Machines' Inkling, where anyone can download the weights — robotics models remain far more gated than language models in 2026.
Does this mean humanoid robots are ready for factories?
Not yet at scale. Verified commercial humanoid deployments in 2026 still number in the hundreds to low thousands of units globally, and most require vendor engineering support on site. Gemini Robotics 2 improves the software ceiling; hardware supply chains and integration cost remain the binding constraints.
The verdict
Gemini Robotics 2 is the most important robotics release of 2026 so far, and the reason is architectural rather than acrobatic. Splitting embodied reasoning from action execution, and shipping a local model alongside the cloud one, is the design that makes fleet deployment financially plausible. The whole-body demos are impressive; the hours-not-months embodiment transfer is the part that could actually reorganize an industry.
The honest caveat: robotics has no shared benchmark, so every claim here rests on DeepMind's own video and DeepMind's own numbers. Treat it as a strong signal, not a settled result, and watch for third-party reproductions on non-Apptronik hardware over the next two quarters.
If you are tracking where the compute for all of this comes from, read our breakdown of the custom AI chip revolution next — the brains in these robots are downstream of that fight.
The robots finally learned to walk and think at the same time. Now the industry has to prove they can do it for eight hours a shift, on someone else's factory floor, without a Google engineer standing nearby.
Top comments (0)