Google DeepMind launches Gemini Robotics 2.5 with 96% success in Mercedes-Benz pilot
Google DeepMind has released Gemini Robotics 2.5, a new vision-language-action (VLA) model that turns camera input and natural-language instructions directly into motor commands for robot arms, hands and humanoid bodies. The company announced the model on November 10, 2026, describing it as its most capable VLA model to date, and said it is rolling out to its robotics partners rather than to the general public. The headline result comes from outside the lab: humanoid robots running the model completed an automotive parts-kitting task with a 96% success rate on active Mercedes-Benz production lines in Hungary and Germany.
The release arrives a little over three months after Gemini Robotics 2, which Google DeepMind introduced on July 30, 2026 together with the Gemini Robotics ER 2 reasoning model and the lighter Gemini Robotics On-Device 2. Version 2.5 replaces the VLA layer of that family. According to Google DeepMind, it was evaluated across more than 200 distinct tasks and delivers success rates between 30% and 100% higher than the previous baseline, depending on the task category, while finishing tasks 20% to 30% faster.
The announcement also covers a new safety layer. Google DeepMind says inference-time reasoning guidance raised safe task completion by 33% around obstacles and by 43% around nearby people, and it released ASIMOV-Jailbreak, an open-source benchmark of prompts that disguise unsafe requests to robots. Gemini 4 Argon, the company's frontier model, now serves both as a simulation engine for training and as the reference model for refusing unsafe robot tasks, with a reported 99.5% resistance score on the new jailbreak benchmark against 82% for Gemini Robotics ER 2.
··········
GEMINI ROBOTICS 2.5 AND THE GEMINI ROBOTICS MODEL FAMILY.
Google DeepMind's robotics stack is built from separate models with distinct jobs. The VLA model handles low-level control: it receives images and an instruction and outputs the movements of joints, fingers or grippers. The embodied reasoning model, Gemini Robotics ER, sits above it as a planner that understands the scene, breaks a job into steps, tracks progress from video and decides when to move on. A third model, On-Device, is a smaller VLA designed to run locally on robot hardware without a network connection.
Gemini Robotics 2.5 is a new version of the first layer only. Google DeepMind's model page now lists Gemini Robotics 2.5 as its most advanced VLA, alongside Gemini Robotics ER 2 and Gemini Robotics On-Device 2, which keep their July version numbers. In the Mercedes-Benz pilot the planning and control layers worked together, with Gemini Robotics ER tracking the workflow and reasoning about the next step while Gemini Robotics 2.5 controlled the arms and hands.
........
Model | Role | Latest version | Access stated by Google |
Gemini Robotics 2.5 | VLA, vision and language to motor control | Nov 10, 2026 | Rolling out to robotics partners |
Gemini Robotics ER 2 | Embodied reasoning and task planning | Jul 30, 2026 | Gemini API and Google AI Studio (preview) |
Gemini Robotics On-Device 2 | Local VLA for robot hardware | Jul 30, 2026 | Early-access partners |
........
The hardware named for version 2.5 is the same pair Google DeepMind used to present Gemini Robotics 2: Apptronik's Apollo 2 humanoid fitted with the five-finger, 22-degree-of-freedom SharpaWave hand, and the Franka Duo bi-arm platform with two-finger parallel grippers. The July release had also shown the earlier family on Dexmate, SO101 and Trossen platforms through the On-Device model, but the 2.5 announcement does not extend support claims to those robots.
··········
GEMINI ROBOTICS 2.5 BENCHMARKS AGAINST GEMINI ROBOTICS 2.
The figures Google DeepMind published are success rates on specific manipulation tasks, measured on its own evaluation setup. The company's charts compare Gemini Robotics 2 with Gemini Robotics 2.5 on eight examples that require fine control of deformable or awkward objects, such as fabric, ribbon, food and hinged lids. In six of the eight, the earlier model did not complete a single attempt.
........
Task | Gemini Robotics 2 | Gemini Robotics 2.5 | Change (points) |
Close lid of black toolbox | 0% | 93% | +93 |
Packing pastries with tongs | 5% | 100% | +95 |
Untying ribbon bows | 0% | 70% | +70 |
Hanging a cardigan | 0% | 62% | +62 |
Sweep trash into dustpan with brush | 0% | 52% | +52 |
Unhang mug to table | 29% | 75% | +46 |
Pack rice cooker | 0% | 40% | +40 |
Folding pants | 0% | 40% | +40 |
........
A simple average of these eight published results moves from about 4% to 66.5%. That average is our calculation, and it should be read with care, because Google DeepMind selected these tasks to illustrate the new capabilities and did not publish the full results of the 200-task evaluation. The broader claim in the announcement is the 30% to 100% range of improvement across task types, which is a relative figure stated by the company and cannot be checked independently from the published material.
The comparison with the July release helps place the numbers. When Gemini Robotics 2 launched, Google DeepMind reported multi-finger results on Apollo with Sharpa hands of 36% for screwing in a light bulb, 44% for tying a trash bag and 32% for a dustpan task, and openly stated that multi-finger dexterity remained a challenge. The 2.5 charts focus on that same weakness, which is where the gains are largest. Google DeepMind has not said whether the improvement comes from more training data, a larger model, a new architecture or the simulated practice described below, and it has not disclosed parameter counts or latency figures.
··········
GEMINI ROBOTICS 2.5 IN THE MERCEDES-BENZ AND APPTRONIK PILOT.
The industrial result is the part of the announcement most likely to interest manufacturers. Apptronik's Apollo 2 humanoids running Gemini Robotics 2.5 worked on live automotive production lines at Mercedes-Benz plants in Kecskemét, Hungary, and Sindelfingen, Germany. The task was kitting: picking metal brackets out of bulk bins and placing them in ordered trays prepared for each vehicle, a repetitive job that requires the robot to identify parts in clutter, grasp them reliably and sort them correctly.
Google DeepMind reports a 96% success rate for this work. The announcement does not state how many picks or shifts the figure covers, how a failure was defined, whether humans intervened, or what cycle time the robots achieved compared with human workers. Those details decide whether a deployment is economical, so the 96% figure shows that the model can operate in a real factory environment, while saying little yet about productivity. Apptronik's chief executive, Jeff Cardenas, is quoted in the announcement praising the results, and the acknowledgements credit Agile Robots, Apptronik and Mercedes-Benz as partners.
........
Pilot detail | What Google DeepMind disclosed |
Robot | Apptronik Apollo 2 humanoid |
Sites | Mercedes-Benz, Kecskemét (Hungary) and Sindelfingen (Germany) |
Task | Kitting metal brackets from bulk bins into per-vehicle trays |
Model setup | Gemini Robotics ER for workflow, Gemini Robotics 2.5 for arms and hands |
Success rate | 96% |
Volume, duration, cycle time | Not disclosed |
........
··········
GEMINI ROBOTICS 2.5 SAFETY TESTS AND THE ASIMOV-JAILBREAK BENCHMARK.
Physical robots raise a different category of risk from chatbots, because an unsafe instruction ends in movement rather than text. Google DeepMind describes two mechanisms in the new release. The first is inference-time reasoning guidance, in which the system reasons about hazards before acting. The company says this raised safe task completion by 33% in scenes with obstacles and by 43% in scenes with people close to the robot. As with the task results, these are relative improvements reported by the company, and the baseline success rates were not given in the announcement.
The second is a new open-source benchmark, ASIMOV-Jailbreak, which contains hundreds of prompts that try to disguise unsafe requests across dozens of scenes. It extends the ASIMOV-Agentic benchmark that Google DeepMind introduced in July to test whether the reasoning model refuses unsafe tool calls and asks a human for help when uncertain. On ASIMOV-Jailbreak, Gemini 4 Argon scored 99.5% resistance, compared with 82% for Gemini Robotics ER 2. Expressed as the share of disguised requests that got through, that is a fall from 18% to 0.5%, or about 36 times fewer successful jailbreak attempts on this test set. Google DeepMind says Gemini 4 Argon leads its agentic robotics evaluations for refusing unsafe tasks, resolving uncertainty and working near humans.
The announcement does not say whether Gemini 4 Argon now replaces Gemini Robotics ER 2 as the planner in partner deployments, or whether the jailbreak resistance figure applies to a robot running Gemini Robotics 2.5 end to end. Readers comparing the two models can find their general capabilities and API prices in our Gemini 4 Carbon vs Gemini 4 Argon comparison.
··········
HOW GEMINI 4 ARGON TRAINS GEMINI ROBOTICS 2.5 IN SIMULATION.
Collecting real-world robot data is slow and expensive, and Google DeepMind's answer in this release is to use its frontier model as a simulation engine. Starting from a single real camera image of a scene, Gemini 4 Argon can generate thousands of varied simulated practice runs and build a digital twin of the environment. The company says this accelerated training of the new VLA model.
Gemini 4 Argon is also used on the analysis side. Researchers can give it video of a failed robot attempt, ask it to identify the cause, and then test candidate fixes first in simulation and then on physical robots. This ties the robotics program closely to Google's general-purpose models: the same Gemini 4 Argon that the company launched for coding, finance, legal and security work, covered in our report on the Gemini 4 Argon global launch, API pricing and benchmarks, now generates training data for robot control. Google DeepMind has used a similar approach with world models, most recently with the Genie 4 real-time world model.
··········
WHO CAN USE GEMINI ROBOTICS 2.5 AND HOW DEVELOPERS GET ACCESS.
Gemini Robotics 2.5 is not a public product. Google DeepMind says the model is rolling out to its robotics partners, and its model page lists Agile Robots, Apptronik and Boston Dynamics as research partners and says the company is working with more than 100 trusted testers. There is no API endpoint, model ID, price or general release date for the new VLA model. Companies that want early access can join a waitlist through a form on the Gemini Robotics page, and early-stage robotics startups can apply to the Google DeepMind Accelerator, which supports teams building with Gemini Robotics models.
Developers who want to work with the family today have one public entry point: Gemini Robotics ER 2, the reasoning and planning model, which is available in preview through the Gemini API and Google AI Studio under the model ID `gemini-robotics-er-2-preview`, with code examples on GitHub. It can call tools such as Google Search or developer-defined functions, follow task progress from video and hand motor execution to a VLA model. Gemini Robotics ER 2 is also in private preview on Gemini Enterprise Agent Platform. Google has not published pricing for it in the announcement materials.
For robot makers and integrators, the practical picture is therefore split. Planning and scene understanding can be prototyped now with the ER 2 preview, while the improved dexterity of Gemini Robotics 2.5 is reserved for partners working on Apollo 2, Franka Duo and similar hardware. The open-source ASIMOV-Jailbreak benchmark is the one component of the release that any team can download and use immediately to test its own robot instruction pipelines.
··········
GEMINI ROBOTICS 2.5: CONFIRMED FACTS AND OPEN QUESTIONS.
Google DeepMind has documented a new VLA model with large gains on the hardest manipulation tasks it showed in July, an industrial pilot with a 96% success rate on Mercedes-Benz kitting lines, and a safety evaluation in which Gemini 4 Argon resisted 99.5% of disguised unsafe requests. The model works with Apptronik's Apollo 2 humanoid and the Franka Duo platform, and its training used Gemini 4 Argon to generate simulated practice from real scenes.
Several points that would matter to buyers remain undisclosed. Google DeepMind has not published the full 200-task results, the size or latency of the model, the volume and cycle time behind the 96% pilot figure, or a timeline for wider access, pricing or an On-Device version of the 2.5 generation. Until those details appear, Gemini Robotics 2.5 is best understood as a substantial step in robot dexterity for Google's partner network, with Gemini Robotics ER 2 remaining the only part of the family that outside developers can call through an API.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
