Gemini Robotics 2.5 teaches Apollo 2 humanoid to kit Mercedes parts with a 96% success rate

A robot that works flawlessly in a laboratory will often stumble at the first slippery part or slightly shifted crate on a factory floor. DeepMind is trying to close precisely this gap between demo footage and the real shop floor with its new Gemini Robotics 2.5 model. The most interesting aspects of the release are the Mercedes-Benz assembly-line pilot and how the model was trained, rather than benchmark records.

What the model can do

Gemini Robotics 2.5 is a VLA model (vision-language-action). It takes camera images and a text instruction as input and produces movements of the arms and fingers. DeepMind tested it on more than 200 different tasks: serving food with tongs, tying a bow on a gift box, and drawing liquid into a pipette.

These tasks demand finesse rather than strength. They involve tools, soft and flexible objects such as fabric and ribbon, and fragile items. According to DeepMind, successful attempts increased by 30–100%, depending on the task type, while execution was 20–30% faster than with previous versions.

The model is not tied to a single hardware platform. It ran on the Apptronik Apollo 2 humanoid with a five-fingered SharpaWave hand (22 degrees of freedom), and on the Franka Duo dual-arm platform with simple parallel grippers.

From laboratory to factory

The main demonstration is a pilot on Mercedes-Benz production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Apollo 2 performed kitting: taking parts out of large crates and arranging them in trays in the order they will be needed for a specific car on the line. The work is repetitive and physically demanding, and a sequencing error creates problems for assembly workers downstream.

The architecture has two layers. The Gemini Robotics ER model monitors the overall process and decides what to do next. Gemini Robotics 2.5 controls the arms: reaching into a crate, picking up metal brackets, placing them in a tray, and correcting its actions if something goes wrong. The result is a 96% success rate.

That figure needs to be kept in perspective. This is a proof of concept, not routine production deployment, and a 4% failure rate is still high on an assembly line. No comparison with human operating speed is provided. However, Apptronik's CEO says the model improved noticeably over a few weeks of fine-tuning and transferred to the Mercedes factory with only a small drop in performance. For the industry, that matters more than the percentage itself: transferring learned skills is cheaper than configuring a system separately for every shop floor.

Safety and the simulator's role

For physical agents, safety is an engineering challenge in its own right. Gemini Robotics 2.5 includes inference-time reasoning guidance that helps the robot navigate hazardous areas. According to DeepMind's measurements, safe task completion improved by 33% near obstacles and by 43% near people.

The team also tested resistance to jailbreaks, in which a dangerous request is disguised as a harmless one. For this, DeepMind released the ASIMOV-Jailbreak benchmark: hundreds of prompts across dozens of scenes. Gemini 4 Argon blocks 99.5% of these attempts, compared with 82% for the previous Gemini Robotics ER 2 baseline.

Gemini 4 Argon also played a significant role in development itself. It served as a simulation engine: using a single real camera scene to build a digital twin and generate thousands of variations for training runs. It also used video to infer why a robot had failed a task and helped test fixes in simulation and on real robots. Real-world demonstrations and simulation together help the model generalize more quickly to new scenes and objects.

What to do with this

The model is currently available only to partners, so you cannot try it directly. Still, there are several practical takeaways.

• If you build robotic systems, consider separating the roles: a planning model and an execution model. This makes individual components easier to debug and replace.
• The approach of capturing a real scene and generating thousands of simulations is worth adopting wherever real-world data is scarce.
• ASIMOV-Jailbreak is publicly available and can be run against your own agents as an additional test.

All these figures share one limitation: they come from DeepMind itself, there is no independent verification, and the baseline versions used for comparison are only sparsely described. But the direction is clear: humanoids are moving out of laboratories and into specific, narrowly defined operations. The central question is now how cheaply and safely they can scale.