Gemini Robotics 2.5 brings robots closer to real industrial tasks

Google DeepMind introduces Gemini Robotics 2.5, a robot control model tested in a Mercedes-Benz factory. The tool improves precision, but safety and latency remain concrete limitations.

Google DeepMind has unveiled Gemini Robotics 2.5, a new generation of its AI system for robots, tested in an industrial environment at Mercedes-Benz. The model, still reserved for partners, promises gains in precision and speed across more than 200 tasks, while aiming for safer use outside the lab. In practice, the challenge is simple: to have a machine perform complex gestures without turning the workshop into a slapstick scene.

A model designed to see, understand, and act

Gemini Robotics 2.5 is based on two complementary components. The first, Gemini Embodied Reasoning, analyzes instructions and live video to plan the action. The second, Gemini Robotics, translates this understanding into concrete movements, such as grasping an object, moving an arm, or adjusting a grip.

Google DeepMind presents this system as a vision-language-action platform, or VLA. This category of models does not just describe what it sees: it transforms visual and textual information into robotic commands. This enables a humanoid to attempt a delicate gesture, like tying a ribbon or handling a pipette.

Measured gains, but still far from flawless

According to Google DeepMind, Gemini Robotics 2.5 achieves a success rate 30 to 100% higher than its own previous versions, depending on the tasks, with execution 20 to 30% faster. These results come from laboratory tests conducted on more than 200 activities, including service with pincers and laboratory-type manipulations.

The most concrete test concerns the automotive industry. With Apptronik and Mercedes-Benz, Google DeepMind deployed the humanoid robot Apollo 2 on active lines in Hungary and Germany, in Kecskemét and Sindelfingen. The robot performed a kitting task, sorting loose metal parts and placing them in bins, with a reported success rate of 96%. That’s solid. Not yet enough to fully replace a human operator in this type of position.

Why robots still need to learn from the fake

To train Gemini Robotics 2.5, DeepMind used Gemini 4 Argon as a simulation engine. This model generates thousands of scenario variants from real images, creating “digital twins” designed to accelerate learning and generalization to new objects or environments.

The process complements classic physical demonstrations. In other words, the robot learns from real data but also from simulated situations, which compensates for the lack of “physical data” available for robotics compared to text or images. Argon also helps analyze failures and suggest improvements, without real-time operation.

One major limitation remains: Gemini Robotics 2.5 still runs on cloud servers. Quality gains therefore come with latency, which matters as soon as a robot must react quickly in a space shared with humans. The model cannot afford to think too long before moving.

Safety becomes a product line, not a footnote

DeepMind says it has integrated reasoning at inference time to avoid immediate dangers. If an obstacle appears, the robot can stop or wait before resuming. In its evaluations, this approach improves task success by 33% in the presence of obstacles and by 43% when people are on the path.

The company also publishes ASIMOV-Jailbreak, an open-source benchmark available on Hugging Face. This dataset contains 1,500 prompts designed to push a robot to adopt dangerous behavior, for example closing a container with a mechanism too fine for it. The goal is to measure models’ resistance to these bypass attempts.

According to Google DeepMind, adding Argon in the loop raises the safety level against these jailbreaks to 98.8%, compared to 73.5% with Gemini Robotics ER 2 alone. Another measure published by Google’s subsidiary reports a score of 99.5% for Gemini 4 Argon, versus 82% for its predecessor [to be verified], with figures varying by source. For industry, the question is no longer just whether the robot can do the job. It is also necessary to verify what it refuses to do.

Key points

  • Gemini Robotics 2.5 targets humanoid and industrial robots.
  • More than 200 tasks served as the basis for tests.
  • Apollo 2 reached 96% on Mercedes-Benz lines.
  • ASIMOV-Jailbreak contains 1,500 security test prompts.
  • Google DeepMind reports a gain of 30 to 100%.
  • The model is reserved for partners, not the general public.

By the numbers

  • 96% — reported success rate at Mercedes-Benz, in Hungary and Germany.
  • 200+ tasks — evaluation base for Gemini Robotics 2.5, according to Google DeepMind.
  • 1,500 prompts — volume of the open-source ASIMOV-Jailbreak benchmark.
  • 33% — improvement in success near obstacles, according to DeepMind.
  • 43% — increase in success when people are close to the robot.

To read