What Gemini Robotics 2.5 Is and Who Can Use It
Gemini Robotics 2.5 is the action model in Google DeepMind's Gemini Robotics family, and DeepMind calls it its most capable VLA model to date.
DeepMind's announcement and model page publish no application programming interface (API), downloadable weights or public price for Gemini Robotics 2.5.
Access depends on partner status and technical role. DeepMind lists three research partners, and Ars Technica reports they have Gemini Robotics 2.5. Developers who want to test DeepMind's reasoning tools can use the companion preview model without a partnership, while robot buyers cannot yet place an order.
| Group | What they can use | How to get it | Source |
|---|---|---|---|
| Listed research partners (Agile Robots, Apptronik, Boston Dynamics) | Gemini Robotics 2.5 through the partnership (DeepMind lists them; Ars Technica reports they have it; DeepMind's release shows it on Apptronik's Apollo 2) | Partnership with Google DeepMind | DeepMind model page; Ars Technica |
| Trusted testers (DeepMind says it works with 100+) | Early access | Waitlist form on DeepMind's model page | DeepMind model page |
| Developers | Gemini Robotics ER 2 preview (the reasoning model, not 2.5) | Google AI Studio | DeepMind model page |
| Robot buyers | Nothing to order; no public price | None published | DeepMind announcement |
Which Robot Makers Get Gemini Robotics 2.5
Google DeepMind's official model page lists three research partners: Agile Robots, Apptronik, and Boston Dynamics. Ars Technica reports that Gemini Robotics 2.5 is already in the hands of partners including these three firms.
DeepMind's announcement says Gemini Robotics 2.5 works with many robot types, including Apptronik's Apollo 2 humanoid and the Franka Duo. On Apollo 2, the model controls a five-finger, 22 degree-of-freedom SharpaWave hand. On the Franka Duo, the model directs two-finger parallel grippers.
Boston Dynamics is one of DeepMind's three listed research partners and Ars Technica reports it has 2.5, but DeepMind's own release does not show 2.5 on a Boston Dynamics robot. The Bot Scout covers Boston Dynamics' humanoid in its Boston Dynamics Atlas review. DeepMind's announcement also thanks the Agile, Apptronik, and Mercedes-Benz teams, with Mercedes-Benz participating as an industrial pilot partner rather than a machine manufacturer.
What Is New Compared with Gemini Robotics 2
DeepMind highlights three changes in Gemini Robotics 2.5: more dexterous skills, higher success with faster execution, and safety steering at inference time.
The earlier releases were Gemini Robotics in March 2025, Gemini Robotics 1.5 in September 2025 and Gemini Robotics 2 in July 2026.
DeepMind evaluated Gemini Robotics 2.5 across 200 distinct tasks, including serving food with tongs, tying a bow on a gift box, and pipetting a liquid. DeepMind claims that success rates increased by 30% to 100% and execution speeds increased by 20% to 30% compared to previous baselines, varying by task type. However, DeepMind does not publish the underlying per-task baseline numbers, which prevents readers from knowing which specific tasks saw 30% gains and which reached 100%.
| Model | Announced | What changed | Availability |
|---|---|---|---|
| Gemini Robotics 2 | July 30, 2026 | Whole-body control; introduced ER 2 and On-Device 2 | ER 2 on Google AI Studio and in private preview on Gemini Enterprise Agent Platform; VLA and On-Device models to early-access partners |
| Gemini Robotics 2.5 | November 10, 2026 | Evaluated on 200 tasks for dexterity; 30% to 100% higher success and 20% to 30% faster than previous baselines, depending on task type; safety steering at inference time | Rolling out to DeepMind's robotics partners |
The Mercedes-Benz Kitting Pilot
DeepMind conducted a proof-of-concept trial with Apptronik and Mercedes-Benz on active manufacturing lines in Kecskemét, Hungary, and Sindelfingen, Germany. In this trial, an Apollo 2 unit performed automotive kitting by transferring metal brackets from bulk storage bins into specific sequencing containers. DeepMind reported a 96% task success rate during this evaluation.
In this deployment workflow, the Gemini Robotics ER model tracks high-level assembly progress while Gemini Robotics 2.5 controls the robot's arms and hands. DeepMind's announcement quotes Apptronik Chief Executive Officer (CEO) Jeff Cardenas. The Bot Scout's Apptronik Apollo review covers the robot that ran the pilot.
This trial was an experimental proof-of-concept rather than an ongoing production installation. In plain terms, a 96% success rate means about 4 in 100 attempts failed in the pilot. Ars Technica reports that DeepMind director Kanishka Rao described resolving the remaining failure gap as a steep road.
How Gemini Robotics 2.5 and ER 2 Split the Work
The Gemini Robotics ecosystem separates motor execution from scene understanding. Gemini Robotics 2.5 operates as the vision-language-action (VLA) engine that produces physical arm movements and gripper positions. Gemini Robotics ER 2 functions as the embodied reasoning (ER) model, handling spatial perception, multi-step planning, and assembly tracking. A third model, Gemini Robotics On-Device 2, runs locally on the robot.
DeepMind's model page claims the Gemini Robotics family can adapt to any bi-arm robot in just a few hours. To train and validate these control policies, DeepMind uses Gemini 4 Argon as a simulation engine. Argon runs thousands of synthetic practice trials using digital twins generated from camera scenes, pairing them with physical demonstrations and formulating hypotheses when a physical task fails.
According to the Google Developers Blog, Gemini 4 Argon achieved an 87% score on the Embodied Reasoning Question Answering (ERQA) benchmark and controlled a Boston Dynamics Spot robot during an evaluation demo. Ars Technica reports that Gemini Robotics 2.5 runs on cloud servers, so its gains come with latency. DeepMind's announcement does not say where the model is hosted. The Bot Scout's guide to robot foundation models explains how these models work in physical machines.
What the ASIMOV-Jailbreak Safety Benchmark Measures
Alongside the model release, DeepMind introduced ASIMOV-Jailbreak as an open-source evaluation benchmark designed to assess whether robots can resist manipulative or unsafe instructions. The benchmark sits on Hugging Face under a Creative Commons (CC) BY 4.0 license. The repository contains evaluation scripts for Physical Safety & Multi-Turn Jailbreak, with access gated behind user acceptance terms.
DeepMind reports that the benchmark contains hundreds of adversarial prompts that disguise hazardous requests across dozens of distinct physical scenes. In DeepMind's tests, Gemini 4 Argon achieved a 99.5% jailbreak resistance score, compared to 82% for the earlier Gemini Robotics ER 2 baseline. DeepMind also reports that using inference-time reasoning guidance increased safe task completion by 33% around obstacles and 43% around nearby people.
Ars Technica's report gives differing numbers, reporting 1,500 test prompts, a 98.8% resistance score for Argon, and a 73.5% score for standalone ER 2. DeepMind's own official publication lists the 99.5% and 82% figures. DeepMind's 99.5%, 82%, 33% and 43% figures come from its own evaluations, and neither DeepMind nor Ars Technica cites a third-party replication.
What Gemini Robotics 2.5 Means and What It Leaves Out
Gemini Robotics 2.5 serves primarily as an indicator of where DeepMind's partner hardware is heading, rather than an operational software layer most organizations can deploy today. DeepMind is rolling it out to its robotics partners, its gains are DeepMind-reported, and the strongest real-world evidence is one 96% kitting proof-of-concept on Apptronik's Apollo 2.
In our reading of DeepMind's announcement and model page, several details remain unpublished. Google DeepMind has published no pricing structure, no public API, no weights for local deployment, no independent third-party replications, and no latency benchmarks for Gemini Robotics 2.5.
Buyers who want a robot they can order today can start with The Bot Scout's list of robots you can buy now. Teams comparing other robot-model labs can read our profile of Physical Intelligence. Readers weighing a different approach can read how OpenAI is betting on general models for robots.
Our assessment would shift if Google DeepMind establishes general availability, releases public API pricing, publishes independently verified replication data, or announces production line deployments on hardware from Agile Robots or Boston Dynamics.
Bottom Line
Hardware makers who want Gemini Robotics 2.5 can join the trusted-tester waitlist on DeepMind's model page. Developers who need planning and perception today can test Gemini Robotics ER 2 in Google AI Studio. Robot buyers have nothing to order yet and can follow Apollo 2 in The Bot Scout's Apptronik Apollo review.
Before planning a project around Gemini Robotics 2.5, check DeepMind's model page for changes to partner access and the waitlist.
FAQs
Is Gemini Robotics open source?
DeepMind released the ASIMOV-Jailbreak benchmark as an open-source evaluation suite under a CC BY 4.0 license on Hugging Face. DeepMind's announcement and model page publish no Gemini Robotics 2.5 weights or public download. Access runs through DeepMind's robotics partners and its trusted-tester waitlist.
Is Gemini Robotics free?
Google DeepMind has not published pricing for Gemini Robotics 2.5 or its partner programs. Developers can access the reasoning model Gemini Robotics ER 2 as a preview in Google AI Studio, but DeepMind has not released commercial rates or licensing fees for deploying the 2.5 action model.
How do I use Gemini Robotics?
Robotics organizations can join the early-access waitlist on DeepMind's model page, and developers can try Gemini Robotics ER 2 in Google AI Studio.
Does Boston Dynamics use Gemini Robotics 2.5?
Boston Dynamics is listed on DeepMind's model page as one of three official research partners, and Ars Technica reports that it has received the software. However, DeepMind's own release does not show Gemini Robotics 2.5 running on a Boston Dynamics robot.
What is Gemini Robotics ER 2?
Gemini Robotics ER 2 is DeepMind's embodied-reasoning model, introduced on July 30, 2026, alongside Gemini Robotics 2. It processes scene perception, spatial understanding, and workflow progression rather than direct motor signals, and it is available in Google AI Studio and in private preview on Gemini Enterprise Agent Platform.
What does Gemini Robotics do?
Gemini Robotics turns camera input and plain-language instructions into robot motor control. The family includes the Gemini Robotics 2.5 action model for motor control, Gemini Robotics ER 2 for task reasoning, and Gemini Robotics On-Device 2 for onboard local execution.
How good is Gemini Robotics?
DeepMind reports that Gemini Robotics 2.5 completed evaluated tasks 20% to 30% faster with 30% to 100% higher success rates compared to earlier baselines, alongside achieving a 96% success rate during an automotive kitting trial. However, DeepMind has not published the per-task baseline figures, and neither DeepMind nor Ars Technica cites an independent test.