Google DeepMind rolls out Gemini Robotics 2.5 to partners
Google DeepMind says its new robot-control model helped Apptronik humanoids achieve a 96% success rate sorting parts in a Mercedes-Benz production-line pilot.
Loading… Բեռնվում է…
Google DeepMind says its new robot-control model helped Apptronik humanoids achieve a 96% success rate sorting parts in a Mercedes-Benz production-line pilot.
Sole source: Google DeepMind
Google DeepMind introduced Gemini Robotics 2.5 on 10 November 2026, and said the model is rolling out to robotics partners. The company describes it as a vision-language-action model: software that uses visual information and language to guide a robot’s actions.
The release follows Google DeepMind’s introduction earlier in 2026 of whole-body control for several types of robot, including humanoids. The original Gemini Robotics and its related reasoning model launched on 12 March 2025.
Google DeepMind says it partnered with robot maker Apptronik and Mercedes-Benz to test Apollo 2 humanoids in a proof-of-concept on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany.
In the pilot, the robots autonomously sorted parts from bulk bins into trays in the order needed for each vehicle, achieving a 96% success rate, Google DeepMind says. Its Gemini Robotics ER model tracked the workflow, while Robotics 2.5 controlled the hands and arms that picked metal brackets and placed them in containers. Google DeepMind says transferring such preparation work to robots could let technicians focus on complex build tasks.
Google DeepMind has not specified the trial count, how it defined success or how often people had to intervene. It has not announced a timetable for expanding the pilot.
Google DeepMind says it evaluated the model on roughly 200 tasks, including serving food with tongs, tying a bow and pipetting liquid. It says success rates were 30% to 100% higher than previous baselines, depending on the task, and completion was 20% to 30% faster. The company also names Apollo 2, with its five-finger SharpaWave hand, and Franka Duo, with two-finger parallel grippers, among the platforms that work with the model.
From 8 December 2026, most repositories inactive for a year will have their alerts paused and archived. A push or turning Dependabot alerts back on in repository settings will restore them, while some enforced security policies will prevent the pause.
The company says reasoning guidance helps steer robots around physical hazards as they act. In its evaluations, it says, safe task completion improved by 33% around obstacles and 43% around nearby people.
Separately, Google DeepMind says it is releasing ASIMOV-Jailbreak, an open-source test of resistance to prompts that disguise unsafe requests. It reports 99.5% resistance in the prompt test for Gemini 4 Argon, compared with 82% for its previous Gemini Robotics ER 2 baseline.
Google DeepMind says Gemini 4 Argon also helped develop the new model by generating simulated practice runs and digital copies of real scenes, and by helping researchers analyze failures and test improvements. The company has not announced terms or a date for broader access to Robotics 2.5.
2 min
Companies pledge shutdown mechanisms for models they control as part of a political commitment adopted at a 9 November 2026 meeting with France’s president. The declaration sets no deadlines or sanctions.
2 min
The deployment covers DP World London Gateway, DP World Logistics Park, Ford Dagenham and two Port of Tilbury locations, giving port, logistics and manufacturing tenants dedicated connectivity for industrial applications.
3 min