Google DeepMind Rolls Out Gemini Robotics 2.5 to Robot Partners
Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026, a robotics AI model tuned to learn faster in real-world settings.
2 sources 13h ago
Reporting from 2 sources: GameBusiness.jp, GIGAZINE.
Why it matters
DeepMind is now shipping a robotics model whose reliability claim rests on a car factory line and whose safety claim rests on a jailbreak benchmark, which puts the pitch closer to deployment evidence than to demo footage.
Key Facts
- Google DeepMind announced Gemini Robotics 2.5 on November 10, 2026.
- Gemini Robotics 2.5 was evaluated on about 200 tasks, including plating food with tongs and pipetting liquids.
- DeepMind based its reliability claims on field trials conducted on a Mercedes-Benz production line.
- Gemini Robotics 2.5 reached 99.5% jailbreak resistance, against 82% for Gemini Robotics ER 2.
- Gemini 4 Argon was used in developing Gemini Robotics 2.5.
The company reports gains in dexterity, reliability based on field trials on a Mercedes-Benz production line, and safety, including 99.5% jailbreak resistance against an 82% baseline for Gemini Robotics ER 2. The model is being rolled out to DeepMind's robot partners.
The evaluation set ran to roughly 200 tasks, among them plating food with tongs, tying ribbons on gift boxes, and pipetting liquids. DeepMind frames those tasks as steps toward human-level dexterity, built from tool operation, work with deformable objects, and handling of delicate items.
Safety work got its own numbers. Inference-time guidance steers the robot away from dangerous objects in real time, and DeepMind reports the safety classification function raised safe task completion by 33% around obstacles and 43% around nearby people. Development also drew on Gemini 4 Argon, which the company says helped build virtual training spaces for robots from a single image.
Written with AI by Yomimono from the 2 cited sources below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google