Google DeepMind has unveiled Gemini Robotics 2.5, a vision-language-action model that processed visual information, understood language instructions and translated them into robot movements. In a reported pilot on active Mercedes-Benz production lines, Apptronik’s Apollo 2 humanoid robot completed automotive kitting with a 96 percent success rate, according to Google DeepMind.
Gemini Robotics 2.5: A factory proof of concept
The pilot took place at production facilities in Kecskemet, Hungary, and Sindelfingen, Germany. Apollo 2 sorted metal parts from bulk bins and placed them into trays in the order required for each vehicle on the assembly line. The 96 percent result is Google’s own figure, and it has not been independently checked.
Google DeepMind, Apptronik and Mercedes-Benz collaborated on the trial. The Decoder reported that the project was a proof of concept rather than evidence of widespread commercial deployment. That distinction matters because a strong result in a defined workflow does not show that robots are ready for broad deployment across factories or other workplaces.
The factory task combined perception, sequencing and manipulation. The robot had to identify individual components in shared bins, pick the right parts and arrange them for the production order. A failure at any stage could interrupt the workflow. The reported 96 percent success rate therefore represented a more demanding test than a stationary demonstration, although it remained one application within a limited pilot.
Two models work together
The system used two Google DeepMind models in tandem. Gemini Robotics ER monitored the workflow, while Gemini Robotics 2.5 controlled the robot’s hands and arms. The Decoder described this division as embodied reasoning guiding physical action, allowing one system to track the overall job while the other handled dexterous movement.
The architecture is designed for different robot types. Apptronik’s Apollo 2 has a five-fingered hand with 22 degrees of freedom, while the Franka Duo platform uses two-finger parallel grippers. This flexibility is important because robot hardware remains fragmented. A model that depends on one specific arm or gripper would be harder to deploy than software designed to transfer across platforms.
Gemini Robotics 2.5 builds on Gemini Robotics 2, introduced earlier this year. According to The Decoder, the predecessor was designed to give robots full-body control for the first time. The new model extends Google’s effort from broader coordination toward specific manipulation tasks, with factory kitting serving as the most commercially relevant example in the reported results.
Apptronik CEO Jeff Cardenas said the models improved significantly within a few weeks of post-training and transitioned to Mercedes-Benz production with minimal performance loss. That statement points to a practical advantage of general robotics models: if performance can survive transfer to a new production site, training can be less tightly tied to every individual environment.
Broad gains, uneven hard-task results
Google DeepMind tested Gemini Robotics 2.5 on more than 200 tasks. These included serving food with tongs, tying bows on gift boxes, transferring liquids with pipettes and folding trousers. Against the company’s previous baselines, the model achieved a 30 to 100 percent higher success rate depending on task type and completed tasks 20 to 30 percent faster.
The strongest reported improvements were concentrated in particular tasks. Packing pastries with tongs rose from 5 percent to 100 percent, while closing the lid of a toolbox rose from 0 percent to 93 percent, according to Modern Robotics. Such changes show where targeted gains can be substantial, but they should not be treated as an average improvement across every activity.
More difficult tasks continued to fail often. Folding trousers rose from 0 percent to 40 percent, hanging a cardigan from 0 percent to 62 percent, and untying a ribbon bow from 0 percent to 70 percent. Those results underline the continuing difficulty of handling soft, flexible materials. The figures are Google’s own test results and have not been independently checked.
The contrast also explains the trade-off identified by Google DeepMind robotics head Carolina Parada. She described the balance between becoming very good at one application and remaining general purpose. Gemini Robotics 2.5 appears strongest when its capabilities are tested across varied tasks, but its output still depends heavily on the object, tool and conditions involved.
Safety gains and open testing
The model uses inference-time reasoning guidance to steer around physical hazards while performing a task. In Google’s own evaluations, this lifted safe task completion by 33 percent around obstacles and by 43 percent around nearby people. The Decoder also reported that the system could stop or wait when something blocked its path, an important behavior for robots operating near workers.
Google DeepMind published the open-source ASIMOV-Jailbreak benchmark on Hugging Face. It contains hundreds of prompts that disguise unsafe requests across dozens of scenes. The dataset is intended to test whether robotics models can be tricked into dangerous actions, offering researchers a common test as physical AI safety measurement develops.
Google reports that Gemini 4 Argon achieved jailbreak resistance of 99.5 percent, compared with 82 percent for the earlier Gemini Robotics ER 2 baseline. Ars Technica reported slightly different figures from the same project, at 98.8 percent against 73.5 percent. The divergence means the exact results should be treated cautiously, even though both sets show a substantial improvement over the earlier baseline.
Simulation helps reduce physical training demands
During development, Google DeepMind used Gemini 4 Argon as a simulation engine. Argon generated thousands of simulated training runs from a real camera scene and created a digital twin. Combined with real-world demonstrations, these simulations were intended to help robots generalize to unfamiliar objects and settings without requiring the same volume of physical training.
Argon can also derive hypotheses about the causes of errors from video recordings. Researchers can then test proposed improvements in simulation and on the actual robot. The Robot Daily said this approach allows a robot to practice varied scenarios from a single real camera scene, potentially reducing the amount of physical data required for each new skill.
One practical limitation remains. Ars Technica noted that the current release runs on cloud servers, so gains in accuracy and speed still come with some delay. For a robot working beside people, response time matters alongside task completion. The reported demonstrations may therefore describe software capability more strongly than the latency profile required for routine industrial deployment.
A direction of travel, not broad availability
Gemini Robotics 2.5 is being rolled out to partners including Agile, Apptronik and Mercedes-Benz rather than released to the public. Google DeepMind director Kanishka Rao called the last 4 percent of the factory result a steep road. That small gap can still matter on a production line where a human performing repetitive work is expected to get the task right almost every time.
The result is nevertheless significant for industrial robotics. It combines a general-purpose manipulation model with a workflow model, supported by simulated training and explicit safety evaluation. The next tests are commercial rather than laboratory based: whether Mercedes-Benz moves from a pilot to paid deployment, what success rate it requires, and whether external researchers adopt ASIMOV-Jailbreak and report independent findings.
For now, Gemini Robotics 2.5 is best understood as credible evidence that foundation models can improve robotic picking and sequencing on live factory lines. The reported 96 percent pilot result is meaningful, but the remaining error rate, cloud dependency and unverified performance claims leave clear work before Gemini Robotics 2.5 becomes a widely deployed industrial system.