Gemini Robotics 2.5 Sorts Parts on a Mercedes Assembly Line with 96% Accuracy

Picture an assembly line: metal parts are heaped in containers, and they need to be sorted into trays so that each car gets its kit in the right order. This is called kitting, and at factories this work is still done mostly by people. Google DeepMind claims that its new Gemini Robotics 2.5 model already handles it on active Mercedes-Benz lines, with a 96% success rate.

What the Model Is

Gemini Robotics 2.5 is a VLA (vision-language-action) model. Put simply, a single "brain" receives images from cameras and a text command such as "sort the parts into trays", and outputs movements of the hands and fingers. There is no need to program each grasp separately.

According to DeepMind, the model was tested on more than 200 tasks, from arranging food with tongs to working with a pipette. Compared with the company's own earlier versions, the success rate rose by 30–100% (depending on the task type), and tasks are completed 20–30% faster. The model can work with different robots: the Apptronik Apollo 2 humanoid with a five-fingered hand with 22 degrees of freedom, and the Franka Duo platform with two two-fingered grippers.

The Assembly-Line Pilot and Safety

The tests took place in Kecskemét (Hungary) and Sindelfingen (Germany). Apollo 2 picked parts from a loose pile and placed them into trays in the order required for a specific car. Two models were at work: Gemini Robotics ER oversaw the progress of the entire process, like a dispatcher, while Gemini Robotics 2.5 controlled the hands. This is a pattern familiar from agent systems: one layer plans and supervises, the other executes.

Apptronik CEO Jeff Cardenas says that after a few weeks of fine-tuning, the model improved noticeably and moved to Mercedes production with minimal loss of quality. For practitioners this is the most interesting point: the gap between the lab and the shop floor usually eats up a significant share of the metrics.

The model controls the robot in real time and can avoid physical hazards. According to DeepMind's internal measurements, safe task execution rose by 33% near obstacles and by 43% near people.

Separately, the company published an open benchmark, ASIMOV-Jailbreak, on Hugging Face. It consists of hundreds of prompts in which dangerous requests are disguised as ordinary ones across dozens of scenes: the test checks whether a robot can be talked into doing something harmful. According to Google, its Gemini 4 Argon language model shows 99.5% jailbreak resistance, while its predecessor Gemini Robotics ER 2 shows 82%.

The same Argon was used as a simulation engine. From a single real scene captured by a camera, it generates thousands of simulated runs and a digital twin. Combined with real demonstrations, this helps the robot generalize its experience to new scenes and objects faster. In addition, Argon builds hypotheses from video recordings about the causes of errors and tests fixes both in simulation and on a real robot.

Caveats and Practical Takeaways

All the figures are measurements by Google and its partners; there is no independent verification yet. 96% on a single kitting task is not 96% on any job. The comparisons are against DeepMind's own earlier versions, not against competitors. It is not disclosed what exactly counts as "success" or how many attempts were made. And a 4% failure rate on an assembly line is a lot if line downtime is expensive. Access is so far limited to partners: Agile, Apptronik, and Mercedes-Benz.

If you work in robotics or industrial automation, look at the pattern rather than a single number: a "dispatcher model plus executor model" pairing, LLM-based simulation for generating training data, and short fine-tuning for a specific site. For those building agents, ASIMOV-Jailbreak is worth adopting as a model for how to test a system where an error costs not a few extra tokens but a broken part or an injury.

For product people and founders: it is better to wait for open access and independent tests, but pilots can already be planned. Narrow, repetitive operations like kitting look like the first candidates for real-world humanoid deployment.