Google DeepMind Unveils Gemini Robotics 2.5 With 96% Factory Success

Google DeepMind Gemini Robotics 2.5 unveiled

Google DeepMind pulled back the curtain on November 10, 2026, unveiling Gemini Robotics 2.5, their most capable vision-language-action model yet for driving physical machines. The whole point is to shove robots out of the lab and into the messy, unpredictable real world, where they can tackle a much wider range of jobs with better speed and dependability. Carolina Parada, who wrote the announcement, notes the model expands both the breadth of skills a robot can pick up and the pace it learns them, calling it a genuine leap forward in physical intelligence.

This new system stands on the shoulders of earlier work that brought whole-body smarts to platforms like humanoids, and DeepMind gives a nod to Gemini 4 Argon for speeding up the underlying research. They claim gains across three big pillars: dexterity, reliability, and safety. On the dexterity side, robots can now handle significantly more tasks with finer precision, think manipulating everyday tools or wrestling with delicate, flexible objects that usually trip up automated systems. They tested the model on over two hundred distinct tasks, a list that includes serving food with tongs, tying a bow on a gift box, and pipetting liquid. That mix covers advanced tool use, wrestling with deformable materials, and the gentle touch needed for fragile items, which the announcement frames as a step toward human-level dexterity.

The performance jumps are nothing to sneeze at. Compared to old baselines, DeepMind reports success rates thirty to one hundred percent higher depending on the task, with cycle times twenty to thirty percent faster. The model natively speaks to multiple robot bodies, including Apptronik's Apollo 2 humanoid sporting the five-finger, twenty-two-degree-of-freedom SharpaWave hand, plus the Franka Duo platform with two-fingered parallel grippers. That hardware agnosticism means the same brain can drive different bodies without bespoke retraining for each one.

Gemini Robotics 2.5 Mercedes-Benz factory testing

Reliability got its most public stress test via a proof-of-concept on live Mercedes-Benz lines in Kecskemét, Hungary, and Sindelfingen, Germany. There, Apollo 2 humanoids tackled automotive kitting, the repetitive, back-breaking work of sorting individual parts from bulk bins into sequenced trays in the exact order each car needs down the line. An embodied reasoning model, Gemini Robotics ER, tracks the workflow while Gemini Robotics 2.5 drives the arms and hands, correcting for gaps in the bins on the fly. DeepMind says the pilot hit a ninety-six percent success rate, a figure they suggest could free human technicians from monotonous prep to focus on complex build tasks.

Gemini Robotics 2.5 safety improvements

Safety improvements tag along with the performance gains. As robots move into everyday spaces, the company says it has boosted physical safety steerability and rolled out a new benchmark to test against adversarial jailbreaks specific to robotics. That addresses a growing worry that vision-language-action models could be tricked into unsafe physical behaviors, a risk fundamentally different from text-only vulnerabilities.

Gemini Robotics 2.5 partner rollout timeline

The system is now shipping to DeepMind's robotics partners. For developers and teams building AI-driven automation, the release signals that general-purpose robot control is inching toward commercial viability in unstructured settings. The combo of broader task coverage, higher success rates, faster cycles, and multi-platform support slashes the engineering overhead that historically made every new robot app a custom project. Meanwhile, the safety benchmark and steerability features give integrators a clearer framework for risk assessment before deploying near people.

Teams weighing whether to adopt or build on this stack should watch partner integration timelines and the forthcoming technical docs detailing API surfaces, fine-tuning procedures, and the safety benchmark methodology. The Mercedes pilot proves end-to-end autonomy in a live factory isn't a theoretical milestone anymore, but that ninety-six percent figure also means one in twenty-five picks still fails, a rate needing further work for true lights-out ops. Developers should plan for human-in-the-loop supervision during initial rollouts and treat the new adversarial benchmark as a baseline, not a guarantee, when designing safety architectures for their own applications.

Frequently asked questions

When did Google DeepMind announce Gemini Robotics 2.5?

Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, unveiling it as their most capable vision-language-action model for driving physical machines in real-world environments.

What success rate did Gemini Robotics 2.5 achieve in the Mercedes-Benz factory pilot?

In live Mercedes-Benz production lines in Hungary and Germany, Apollo 2 humanoids powered by Gemini Robotics 2.5 achieved a 96 percent success rate on automotive kitting tasks, sorting parts from bulk bins into sequenced trays.

Which robot platforms are compatible with Gemini Robotics 2.5?

The model natively supports multiple robot bodies including Apptronik's Apollo 2 humanoid with the 22-degree-of-freedom SharpaWave hand and the Franka Duo platform with two-fingered parallel grippers, enabling hardware-agnostic deployment without bespoke retraining.

How does Gemini Robotics 2.5 improve on previous robot control models?

Compared to prior baselines, Gemini Robotics 2.5 delivers 30-100 percent higher success rates depending on task, 20-30 percent faster cycle times, and was tested on over 200 distinct tasks including tool use, deformable object manipulation, and delicate handling like pipetting.

What safety features does Gemini Robotics 2.5 include for real-world deployment?

DeepMind added physical safety steerability controls and released a new adversarial benchmark specifically for robotics jailbreaks, addressing risks that vision-language-action models could be tricked into unsafe physical behaviors-a distinct threat from text-only vulnerabilities.

Source: Google DeepMind