A Mercedes-Benz pilot for Gemini Robotics 2.5

A parts-sorting pilot achieved 96% success, according to Google DeepMind. Testing of Gemini Robotics 2.5 also spans around 200 tasks.

At Mercedes-Benz sites in Kecskemét, Hungary, and Sindelfingen, Germany, Apollo 2 robots were used to prepare parts for production lines. According to the Gemini Robotics 2.5 announcement, the pilot achieved a 96% success rate. The task involves picking metal brackets from bins and placing them in containers in the sequence required for assembly. The post does not specify the trial’s duration or number of attempts.

The system combines two models. A Gemini Robotics ER component tracks the workflow, while Gemini Robotics 2.5 controls the hands and arms. The latter is a vision-language-action model, or VLA: visual information and instructions are converted into motor commands. Deployment is being rolled out to the lab’s robotics partners. Individual tasks show uneven results

The evaluation covers around 200 tasks, including using serving tongs, manipulating ribbons, and pipetting liquids. According to the provider, success rates are 30% to 100% higher than previous baselines, depending on the category, and execution is 20% to 30% faster. These results apply to the situations evaluated by the team.

The charts show substantial differences between tasks. Packing pastries with tongs rises from a 5% success rate with Gemini Robotics 2 to 100% with version 2.5. Untying a ribbon bow improves from 0% to 70%, while folding pants reaches 40%, compared with 0% previously. The broader repertoire therefore still comes with frequent failures on some manipulation tasks.

Several hardware platforms are used, including Apptronik’s Apollo 2 humanoid and a dual-arm Franka platform equipped with parallel grippers. The demonstrations do not all rely on the same grasping mechanism: a five-fingered hand and a two-jaw gripper offer different capabilities. Virtual training environments built from a real scene

Part of the development relies on Gemini 4 Argon. Starting from a camera view, this model can help reconstruct a simulation environment and generate thousands of training variations, according to the lab. These virtual trials are combined with real demonstrations to improve adaptation to new objects and scenes.

The work on Argon for robotics also describes analyzing videos of failures. The model helps researchers formulate explanations and test modifications, first in simulation and then on hardware. This role concerns preparing and improving training; it remains separate from the VLA’s role in executing movements. Two different safety evaluations

For physical actions, version 2.5 uses inference-time guidance to avoid environmental hazards, such as obstacles or nearby people. The lab reports improved safe task completion in these situations, although this does not constitute a general safety measurement for every industrial workstation.

Resistance to unsafe instructions is evaluated separately through ASIMOV-Jailbreak, published on Hugging Face. The test attempts to steer an agent toward unsafe actions through adversarial requests. The reported 99.5% resistance rate applies to Gemini 4 Argon, compared with 82% for the Gemini Robotics ER 2 baseline. The dataset documentation describes 845 adversarial episodes, with up to ten steps per episode, evaluated through automated checks and a model-based judge.