Google DeepMind: Gemini Robotics 2.5 Hits 96% at Mercedes

Key Takeaways

1Driver at Mercedes-Benz: Apollo 2 achieves kitting with a reported 96% success rate.

2Safety: +33% near obstacles and +43% near people according to Google DeepMind.

3ASIMOV-Jailbreak open source: Argon reaches 99.5% resistance to jailbreaks, compared to 82% for ER 2.

Why it matters : The model is being tested in production and comes with announced safety tools and metrics, as well as massively simulated training.

Google DeepMind highlights Gemini Robotics 2.5, its vision-language-action model, tested on active lines at Mercedes-Benz for kitting tasks. The company claims performance gains, improved safety around obstacles, and an open-source benchmark to test the robustness of agents.

Safety: 33% Near Obstacles, 43% Near People

Google DeepMind describes real-time control of Gemini Robotics 2.5 aimed at avoiding physical dangers while performing a task. In its internal evaluations, safe completion improves by 33% near obstacles and 43% near people. To test the robustness of the agents, the company is releasing ASIMOV-Jailbreak as open source, available on Hugging Face. This test suite gathers hundreds of instructions that conceal dangerous intentions in dozens of scenes, to verify whether models can be tricked into risky actions. According to Google DeepMind, Gemini 4 Argon achieves 99.5% resistance to jailbreak attempts, compared to 82% for Gemini Robotics ER 2.

Pilot in Production at Mercedes: 96% Success Rate

With Apptronik and Mercedes-Benz, trials have been conducted on active lines in Kecskemét, Hungary, and Sindelfingen, Germany. The humanoid robot Apollo 2 performed kitting by sorting metal parts from bins and placing them into trays according to the required order for each vehicle. Two models operated in concert: Gemini Robotics ER to monitor the flow and Gemini Robotics 2.5 to control hands and arms. Google DeepMind reports a success rate of 96% on this pilot. Jeff Cardenas, CEO of Apptronik, states that performance has significantly improved over a few weeks of training, with a transition to production at Mercedes and minimal loss.

Le brief IA que lisent les pros

Le brief IA que les pros lisent chaque soir

Les 7 actus IA du jour, décryptées en 5 min. Gratuit.

Inclus dès l'inscription : notre sélection des meilleurs guides & comparatifs IA.

Choisissez votre rythme

Gratuit · Pas de spam · Désabonnement en 1 clic

Gemini 4 Argon Trains Gemini Robotics 2.5 from Filmed Scenes

Gemini 4 Argon serves as a simulation engine to train Gemini Robotics 2.5. It can generate thousands of training sessions from scenes filmed by a camera and build a digital twin. Combined with real-world demonstrations, this input aims, according to Google DeepMind, to accelerate generalization to new scenes and objects. Argon can also hypothesize about the source of errors from videos and test improvements in simulation and then on the robot. Google DeepMind reports having evaluated the model on over 200 tasks, including food service with tongs and pipetting operations.

Robotic Compatibility and Deployment with Partners

Gemini Robotics 2.5 is designed to control various systems, including Apptronik's humanoid Apollo 2 equipped with a five-fingered hand and 22 degrees of freedom, as well as the Franka Duo platform with two-finger parallel grippers. Google DeepMind claims success rates higher by 30 to 100% depending on the tasks, and execution times 20 to 30% shorter than its previous models. The model processes visual signals, interprets natural language instructions, and converts them into actions, continuing the work of a predecessor introduced earlier this year that aimed for complete body control. According to Google DeepMind, deployment is underway with partners such as Agile, Apptronik, and Mercedes-Benz.