On November 10, 2026, Google DeepMind announced Gemini Robotics 2.5, a new AI model that controls robot movements. In a proof of concept on live Mercedes-Benz production lines, the company says it achieved a 96% success rate on parts sorting and kitting tasks. However, that figure applies to a specific parts-preparation task, not to the success rate across the entire automotive manufacturing process.
The demonstration combined "Gemini Robotics ER," which tracks the progress of the overall task, with "Gemini Robotics 2.5," which controls the movements of the robot's hands and arms. In addition to improved task performance, DeepMind published evaluation results on hazard avoidance during operation and resistance to dangerous instructions. The robot used in the demonstration also has a safety mechanism that detects surrounding objects and stops. These address different kinds of risk, so the AI model's judgment and the hardware's safety measures need to be evaluated separately.
Automating parts preparation on live Mercedes-Benz lines
According to Google DeepMind's official announcement, the proof of concept was carried out on Mercedes-Benz's live production lines in Kecskemét, Hungary, and Sindelfingen, Germany. The robot used was "Apollo 2," a humanoid developed by Apptronik. It was tasked with picking up metal brackets tossed loosely into supply bins and arranging them in a different bin in the order required for each vehicle.
In this task, Gemini Robotics ER tracks how the work is progressing, while Gemini Robotics 2.5 controls hand and arm movements. DeepMind also describes a function that detects missing parts and corrects for them. The aim is to let robots handle this kind of parts preparation so that human workers can focus on more complex assembly steps.
The 96% success rate DeepMind reported is strictly the result of a proof of concept for this parts sorting and kitting task. The relevant section of the official announcement does not provide details such as the number of trials, the evaluation period, or how often humans intervened during the work. Performing a specific task on an operating production line is meaningful, but this number alone cannot be used to judge long-term stable operation or applicability across an entire factory.
Improved dexterity and task speed across roughly 200 tasks
DeepMind evaluated Gemini Robotics 2.5 on roughly 200 tasks. Examples include handling food with tongs and tying a bow with ribbon on a gift box. These tasks require not only accurate tool use but also fine movements for handling materials that easily change shape. The evaluation emphasizes tasks that demand dexterous hand manipulation, going beyond simply moving objects.
According to DeepMind, success rates improved by 30 to 100% over previous baselines, depending on the type of task. Task completion speed also rose by 20 to 30%. However, these are relative improvements over earlier models; they do not mean that every task reached a 100% success rate or that success rates rose by up to 100 percentage points. Actual success rates and the size of the improvement vary by task.
The supported robots are not limited to a single machine, either. Apollo 2, used in the factory demonstration, has a SharpaWave hand with five fingers and 22 degrees of freedom, while the Franka Duo uses two-finger parallel grippers. On its Gemini Robotics model page, DeepMind says the model also supports whole-body control that coordinates the entire body, not just the hands. Movements such as walking while adjusting the center of gravity, squatting, and bending over are envisioned, but these are capabilities the company describes, not a guarantee that the model works on every machine or in every environment.
"Gemini 4 Argon" was also used in training the model. DeepMind describes a method that generates thousands of simulated practice scenarios from footage of real environments captured by cameras, building digital twins that replicate real-world settings. By combining these simulations with demonstration data from real robots, the company aims to improve the model's ability to handle diverse situations. The system also helps form hypotheses about the causes of failures from footage of failed attempts and test improvements. However, no specific figures have been given for how much this approach shortens training time.
Three different mechanisms underpin robot safety
The safety measures DeepMind described fall broadly into three categories: the robot's ability to avoid obstacles and people while operating, resistance to instructions that encourage dangerous behavior, and a physical stopping mechanism built into the hardware. Because the hazards addressed and the evaluation methods differ, the individual performance metrics cannot be combined into a single safety indicator.
The table below summarizes each measure and its published evaluation results, based on DeepMind's official announcement of November 10, 2026, and Apptronik's Apollo 2 product page. It compiles evaluations and design descriptions obtained under different conditions and is not a comparison of safety performance under identical conditions.
| Safety measure | Target and mechanism | Published figures / design | Caveats |
|---|---|---|---|
| Hazard avoidance during operation | Inference-time control selects safe movements around obstacles and people | Ability to complete tasks safely improved 33% around obstacles and 43% around people | A relative improvement; absolute values before and after cannot be confirmed from the announcement text. Not an accident rate |
| Resistance to dangerous instructions | Resistance to instructions that disguise dangerous requests, evaluated on the public "ASIMOV-Jailbreak" benchmark | Gemini 4 Argon scored 99.5%; Gemini Robotics ER 2, used for comparison, scored 82% | An evaluation of Gemini 4 Argon; does not indicate the safety performance of Robotics 2.5 on its own or physical accident rates |
| Hardware stopping mechanism | Apollo 2 stops when it detects an object within a configured range | Has a zone that triggers an immediate stop on detecting an object, and a surrounding zone with configurable behavior as objects approach | A hardware design feature; the specific settings used in this factory demonstration cannot be confirmed |
On hazard avoidance during operation, DeepMind says the ability to complete tasks safely improved by 33% around obstacles and 43% around people. This evaluates the ability to finish tasks while avoiding dangerous movements through inference-time control, and does not mean physical accident rates fell by the same proportion. The announcement does not give absolute success rates before and after, so there is not enough information to judge what level of safety was reached.
For resistance to dangerous instructions, the evaluation used the public benchmark "ASIMOV-Jailbreak." The test measures resistance to hundreds of prompts that disguise dangerous requests across dozens of scenarios. Gemini 4 Argon scored 99.5%, ahead of the 82% recorded by Gemini Robotics ER 2, the comparison model. However, this figure is an evaluation result for Gemini 4 Argon and is not a direct measurement of the operational safety of Gemini Robotics 2.5 itself. Moreover, being able to refuse dangerous instructions is a separate matter from a real robot avoiding physical accidents.
On the hardware side, Apollo 2 has an object detection and stopping mechanism. According to Apptronik, Apollo 2 has a zone that triggers an immediate stop when an object is detected, and a surrounding zone in which behavior can be configured when people, equipment, or obstacles approach. This mechanism provides hardware-side safety that is independent of whether the AI model recognizes and refuses a dangerous instruction. However, how each zone was configured in the Mercedes-Benz factory demonstration cannot be confirmed from the published materials.
None of these results guarantees that the robot will not cause an accident. The AI model that chooses safe movements, the mechanism that refuses dangerous instructions, and the hardware stopping mechanism each play a different role. In real workplaces, these must be combined and then verified to confirm that contact with equipment and people can be prevented.
Full factory deployment will require validation of recovery from failures and sustained operation
DeepMind is rolling out Gemini Robotics 2.5 to partners in the robotics field, and the model page also directs interested parties to an early-access waitlist. However, availability to partners is different from being freely available to general developers and companies. Anyone considering adoption needs to check who is actually eligible and what the terms of use are.
On the operational side, the robot's own management systems come into play alongside the AI model. According to Apptronik's product description, "Artemis" coordinates functions such as perception, planning, and control, while "Fleet Connect" handles robot status monitoring, task coordination, and deployment management. However, detailed specifications on exactly how these systems were used in this factory demonstration, or how they connect to Google's model, have not been confirmed.
A 96% success rate on live Mercedes-Benz lines represents progress toward practical use of humanoid robots for parts preparation. But completing a specific task with high probability is not the same as operating stably in a factory over a long period. Going forward, the key questions will be whether the robot can recover on its own when a task fails, how much human intervention is needed, and how performance and work efficiency change with continuous operation. If such operational data becomes available, it will be easier to connect the results of this proof of concept to deployment decisions on actual manufacturing floors.