Gemini Robotics 2.5: Google DeepMind’s New Robot AI Model

Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, calling it the company's most capable vision-language-action (VLA) model yet. Across 200 distinct tasks, DeepMind reports 30% to 100% higher task-success rates and 20% to 30% faster task completion than the prior Gemini Robotics 2 model, alongside new physical-safety steering and a jailbreak-resistance benchmark the company is open-sourcing.

What is Gemini Robotics 2.5?

Gemini Robotics 2.5 is Google DeepMind's latest vision-language-action (VLA) model — a system that takes in camera images and natural-language instructions and outputs the low-level motor commands a robot needs to complete a physical task, such as moving a hand to grip a mug or adjusting grip pressure mid-motion. DeepMind, in the blog post authored by Carolina Parada, describes it as the company's "most capable vision-language-action (VLA) model to date," built to widen the range of skills robots can learn and to shorten how long it takes them to learn each one.

VLA models are the current frontier approach for general-purpose robot control: instead of hand-coding motion for each individual task, a single trained model is meant to generalize across many tasks and objects it has not seen in exactly that combination before. DeepMind's robotics work splits into two related models with different jobs. Gemini Robotics 2.5 is the "acting" model — it directly drives a robot's joints and grippers. Gemini Robotics-ER ("embodied reasoning") sits a layer above it, handling higher-level planning, perception, and workflow tracking, then handing specific manipulation steps down to Gemini Robotics 2.5 to execute. The Mercedes-Benz kitting pilot described below uses exactly this split: Robotics-ER tracks the overall sorting workflow, while Robotics 2.5 controls the arms and hands.

The model is designed to work across different robot embodiments rather than one fixed hardware design. DeepMind says it is rolling out to robotics partners now, starting with Apptronik's humanoid platform and Franka's two-armed research platform, rather than shipping as an open public API at launch.

Performance gains vs. Gemini Robotics 2

DeepMind evaluated Gemini Robotics 2.5 against its predecessor, Gemini Robotics 2, across 200 distinct manipulation tasks. The headline figures: success rates rose 30% to 100% depending on the task, and robots completed tasks 20% to 30% faster.

The gains were not evenly spread. Some tasks that Gemini Robotics 2 essentially failed at — a 0% success rate — became reliable under 2.5. Illustrative examples from DeepMind's own benchmark chart include closing the lid of a toolbox (0% to 93%), sweeping trash into a dustpan (0% to 52%), packing pastries with tongs (5% to 100%), and folding pants (0% to 40%). A task like unhanging a mug and placing it on a table improved from a 29% baseline to 75%.

These numbers come directly from DeepMind's published benchmark chart for the announcement; the company has not released a full breakdown of all 200 tasks. The pattern across the examples DeepMind did publish is notable: several tasks went from a 0% success rate under Gemini Robotics 2 — effectively a task the robot could not do at all — to a majority or near-total success rate under 2.5. DeepMind frames this as evidence that the newer model is not simply faster at tasks it already handled, but has picked up entirely new categories of dexterous manipulation, from tying and untying fabric to using tools like tongs.

The 20%–30% speed improvement is a separate axis from success rate: a task can be completed more reliably and more quickly at the same time, which matters for any deployment where a robot is working a shift alongside human coworkers on a timed line, such as the Mercedes-Benz kitting pilot below.

Physical-safety steering

Beyond raw task performance, DeepMind is emphasizing safety behavior around obstacles and people. Gemini Robotics 2.5 uses what the company calls inference-time reasoning guidance — the model reasons about physical hazards as it acts, rather than relying solely on pre-trained caution. DeepMind reports this produced safe task completion that is 33% better around obstacles and 43% better around nearby people, compared with the prior model.

For a VLA model meant to operate in working factories alongside employees — as the Mercedes-Benz pilot below does — that steering behavior is as central to the pitch as speed or success rate. A robot that is faster and more capable but no safer around people is not deployable on an active production line; DeepMind's framing ties the safety numbers directly to the real-world pilot results it is publicizing alongside them.

DeepMind has not published the exact test protocol behind the 33%/43% figures — such as how many trials, which obstacle types, or how "nearby people" was defined — so those numbers should be read as DeepMind's own reported internal evaluation rather than an independently audited safety certification.

ASIMOV-Jailbreak: a new open-source safety benchmark

Alongside the model, DeepMind introduced ASIMOV-Jailbreak, an open-source benchmark built to test whether a robotics model can be tricked into carrying out unsafe physical actions through disguised or indirect prompts. The benchmark contains hundreds of prompts that attempt to hide unsafe requests across dozens of different scenes.

DeepMind says its Gemini 4 Argon model — which also underpins the simulation and reasoning side of this robotics stack — achieved 99.5% jailbreak resistance on the benchmark, compared with 82% for the previous baseline model, Gemini Robotics ER 2. DeepMind has published the ASIMOV-Jailbreak dataset on Hugging Face so other labs can run the same evaluation independently rather than relying only on DeepMind's own reported score. For background on the broader Gemini 4 model family that Argon belongs to, see our explainer on Gemini 4 Argon, Carbon and every variant.

The reasoning behind a dedicated jailbreak benchmark for robotics is different from chatbot jailbreak testing: a disguised unsafe prompt aimed at a text model might produce an offensive sentence, but a disguised unsafe prompt aimed at a robot could, in principle, result in a physical action in the real world. DeepMind's "hundreds of prompts... across dozens of scenes" framing suggests the benchmark varies both the wording of the request and the physical environment the robot is placed in, to check whether a model's safety behavior holds up outside of a single test scene.

Apptronik's Apollo 2 and the Mercedes-Benz pilot

The most concrete real-world evidence DeepMind points to is a pilot run with Apptronik's Apollo 2 humanoid robot on active Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 uses a five-finger, 22-degree-of-freedom hand DeepMind refers to as SharpaWave.

The pilot's task was automotive kitting: sorting metal brackets out of bulk bins and into sequenced trays for assembly workers. In this setup, a separate DeepMind model, Gemini Robotics-ER, tracks and plans the overall workflow, while Gemini Robotics 2.5 controls Apollo 2's arms and hands for the physical manipulation. DeepMind reports the pilot reached a 96% success rate.

Apptronik co-founder and CEO Jeff Cardenas framed the pilot as validation of the robot's original purpose: "This is what we built Apollo for: useful work alongside people in real industrial settings." Apptronik and Mercedes-Benz have an existing commercial agreement to pilot Apptronik's humanoid robots in Mercedes-Benz manufacturing facilities, a partnership the two companies first announced in 2024 and have continued to expand as Apptronik's hardware has moved from the original Apollo to Apollo 2. For more on Apptronik's current robot platform, see Apptronik's own announcement of Apollo 2 and its Robot Park training facility.

Running a humanoid robot on a live production line, rather than in a staged demo, raises the stakes for exactly the safety metrics DeepMind is promoting with this release. Both Kecskemét, Hungary, and Sindelfingen, Germany, are established Mercedes-Benz production sites, which means the pilot ran inside an environment with existing line speed, existing workers, and existing safety procedures that the robot had to fit into, rather than a purpose-built robotics lab.

Franka Duo and other hardware partners

Gemini Robotics 2.5 is not limited to humanoid hardware. DeepMind says the model also runs on the Franka Duo platform, which uses simpler two-fingered parallel grippers rather than a humanoid hand. Franka's robots are widely used in robotics research labs for manipulation work, so pairing Gemini Robotics 2.5 with Duo gives DeepMind a testbed that is distinct from Apptronik's full humanoid form factor — a lower-cost, more common lab setup rather than a bipedal robot built for a factory floor.

Supporting both a dexterous humanoid hand and a simpler industrial gripper from the same model is part of DeepMind's pitch that Gemini Robotics 2.5 generalizes across robot embodiments rather than being tuned to one piece of hardware. That is a meaningfully different claim than a model trained and tuned for a single robot: if the same weights can control a 22-degree-of-freedom humanoid hand and a basic parallel gripper, it suggests the underlying model has learned something more general about manipulation rather than memorizing one robot's specific joint geometry.

DeepMind's robotics stack also leans on Gemini 4 Argon as a simulation engine: Argon generates large numbers of varied practice runs and digital twins built from real camera scenes, which DeepMind says helps researchers understand why a robot fails at a task and test fixes both in simulation and on physical hardware before deploying them. That simulation-to-real pipeline is also where DeepMind says Argon's own jailbreak-resistance score of 99.5% was measured, tying the agentic-reasoning side of the robotics stack back to the ASIMOV-Jailbreak benchmark described above.

How this fits Google's broader AI push

Gemini Robotics 2.5 arrives as Google continues to position its Gemini model family as a single underlying system spanning text, video generation, and now physical robot control. The robotics-specific models (Gemini Robotics 2.5 and Gemini Robotics-ER) sit alongside the general-purpose Gemini 4 family, with Argon acting as the connective tissue for simulation and agentic safety evaluation in this announcement. Readers wanting the fuller map of Google's current model lineup can see what is Gemini 4 Argon, Carbon and every variant, and for how Google's generative tools compare to its video model, Gemini Omni vs. Veo 3.1 covers that side of the stack. Competing labs have been shipping their own frontier updates around the same period; see our coverage of Claude Fable 5.2 and Mythos 5.2 for context on how rivals are positioning their latest releases.

Availability and what it means for robotics development

DeepMind says Gemini Robotics 2.5 is rolling out to robotics partners rather than shipping as a general public API at launch. That mirrors how DeepMind introduced earlier Gemini Robotics models: partner hardware makers like Apptronik and Franka get access first, with broader availability following as the safety and performance claims are tested on more real deployments.

For now, the clearest evidence of real-world impact is the Mercedes-Benz kitting pilot's 96% success rate on an active production line — a higher bar than most lab demonstrations, since it involves variable lighting, real parts, and people working nearby. That is also the main reason DeepMind is leading its announcement with a named factory pilot and a named executive quote rather than only lab benchmark numbers: it is trying to show the 30%–100% success-rate gains translate into a real deployment, not just a cleaner demo environment.

DeepMind frames the Apptronik and Franka pairings as proof points for its cross-embodiment generalization claim, not as an exclusive arrangement — the company has said the model is rolling out more broadly to robotics partners beyond the two named in this announcement.

What next

Watch for DeepMind or Apptronik to publish expanded pilot results beyond kitting, and for Franka to detail how Duo-based deployments perform outside controlled demos. The open-sourced ASIMOV-Jailbreak dataset on Hugging Face also gives outside researchers a way to independently check DeepMind's 99.5% jailbreak-resistance figure, which is worth watching for third-party replication in the coming months.

FAQ

See the questions and answers below for quick answers on dates, partners and benchmark numbers.

Written byTheo Park

Theo Park runs the AI desk at Pandromeda. He follows model launches from the frontier labs and the open-weight community, tracks the assistants and developer tools built on them, and explains what each release changes on pricing, capability and safety. His reporting leans on primary sources: model cards, technical reports, API documentation and the companies' own announcements.

More from Theo ParkAll AI coverage