Google DeepMind unveils Gemini Robotics 2.5 with ASIMOV-Jailbreak safety benchmark

Google DeepMind has rolled out Gemini Robotics 2.5, the newest version of its AI platform for physical machines. The company says its robots are now getting close to matching humans at real-world tasks.

The release landed on November 10, 2026. It comes with a proof-of-concept inside Mercedes-Benz factories and an open-source benchmark called ASIMOV-Jailbreak. The benchmark measures whether robots can be talked into doing something dangerous.

How Gemini Robotics 2.5 works

The planner is Gemini Embodied Reasoning, a vision-language model. It takes a spoken or written instruction, watches a live video feed, and works out a strategy for completing the task.

The mover is the vision-language-action model, or VLA. This component turns the plan into concrete physical motions. DeepMind describes Gemini Robotics 2.5 as its most advanced VLA system so far.

The upgrades target four areas: breadth of tasks, learning speed, dexterity, and reliability. On dexterity, DeepMind showed robots tying bows and using tools.

The Mercedes-Benz test and the safety benchmark

In a proof-of-concept with Mercedes-Benz, Gemini Robotics 2.5 ran on Apptronik’s Apollo 2 humanoid robot. The system reached a 96% success rate on pick-and-pack operations.

The safety half of the announcement is ASIMOV-Jailbreak. This open-source benchmark tests how robots respond to unsafe commands disguised as harmless ones. The benchmark grows out of DeepMind’s broader ASIMOV family of tests, which focus on semantic safety for embodied AI.

According to evaluations reported with the release, Gemini 4 Argon, which was used in training the model, achieved 99.5% jailbreak resistance. The previous baseline was 82%.

Background: a fast release cadence

DeepMind has been building out Gemini Robotics in steady increments. Its predecessor, Gemini Robotics 2, arrived in mid-2026.

Beyond Mercedes-Benz and Apptronik, DeepMind’s partner list also includes Agile Robots. The company says it is aiming the technology at both factory work and everyday settings. It also highlights better steerability, meaning operators get more control over how the robot behaves, along with its safety improvements.

What this means

For hardware makers like Apptronik and Agile Robots, a capable general-purpose model from Google lowers a big barrier. If DeepMind supplies that layer, hardware companies can concentrate on the machines themselves.

Making ASIMOV-Jailbreak open source lets other labs and robot makers test their own systems against the same standard. That could shape how the industry talks about robot safety, and it gives regulators and buyers a shared yardstick when they ask how easily a machine can be manipulated.