Google Releases Gemini Robotics 2.5: Behind the 96% Factory Success Rate, the Real Challenge of the Remaining 4%

On November 10, 2026, Google DeepMind released Gemini Robotics 2.5 and published factory validation data the same day: Apptronik's Apollo 2 humanoid robot, equipped with the model, completed sorting of automotive parts kits on Mercedes-Benz's active production lines in Kecskemét, Hungary, and Sindelfingen, Germany, achieving a 96% success rate. The test was conducted on live production lines.

In an industrial context, a 96% success rate means 4 failures out of every 100 grasps. On an assembly line with tight takt times, the cost of a single line stoppage caused by one failure far exceeds the cost of a retry. Google DeepMind characterized the trial as a "proof of concept" in its official blog.

Dual-Model Architecture: Dividing Reasoning and Action

Gemini Robotics 2.5 technically adopts a dual-model architecture. When Apollo 2 operates on the Mercedes-Benz production line, Gemini Robotics ER is responsible for tracking the overall workflow—understanding the task sequence, judging the current step, and perceiving nearby people and obstacles; the action model of Gemini Robotics 2.5 is responsible for translating reasoning results into precise motion commands for the arms and fingers.

High-level reasoning has a higher tolerance for latency, while low-level motion control has stricter real-time requirements. Separating the two allows the reasoning model to run more complex chains of thought in the cloud, while the action model responds to changes in the physical world locally with lower latency. During model development, Gemini 4 Argon was used as the simulation engine, and part of the training data came from synthetic environments, significantly reducing data collection costs.

From 0% to 62%: Tasks the Previous Generation Could Not Do, Now Possible

A more valuable metric for evaluating a model is its performance on tasks where the previous generation failed completely. Across more than 200 dexterity task tests, Gemini Robotics 2.5 improved overall success rates by 30% to 100% over Gemini Robotics 2, and increased task completion speed by 20% to 30%. Specific task data is as follows:

• Handling pastries with food tongs: from 5% to 100%
• Untying gift box ribbons: from 0% to 70%
• Hanging a cardigan sweater: from 0% to 62%

The previous model simply could not perform these tasks. Deformable objects, unstructured grasping, and dynamic targets have long been the three major challenges in robotic manipulation. Judging from the figures above, Gemini Robotics 2.5 has in some dimensions crossed the threshold between "unusable" and "usable."

Safety: Where the 99.5% Jailbreak Resistance Comes From

Unlike pure software AI, safety failures in physical robots carry real injury costs. Google DeepMind released the ASIMOV-Jailbreak benchmark (now open on Hugging Face); Gemini Robotics 2.5 achieves a 99.5% jailbreak resistance success rate on this benchmark, compared with a previous baseline of 82%.

The safety rate for completing tasks near obstacles improved by 33%, and in environments with people by 43%. These two metrics map directly to the core requirement of factory collaboration scenarios: robots must be able to operate on production lines where humans and machines work side by side.

Who Can Use It? The Closed Logic of the Partner Program

Gemini Robotics 2.5 is currently being rolled out in batches to Google DeepMind's robotics partners. Confirmed partners include Apptronik, Agile Robots, and Boston Dynamics. The announcement did not mention a general API access path, a waitlist, or pricing information.

This "partner-first" distribution strategy has particular reasons in the field of robot AI. The performance of VLA models depends heavily on the kinematic parameters, sensor configurations, and control interfaces of specific robot hardware. Apptronik's Apollo 2 is equipped with 22-DOF SharpaWave hands, while Franka Duo uses a two-finger gripper—the same model requires different adaptation on the two hardware platforms. The closed partnership model essentially shifts hardware adaptation costs onto partners.

Apptronik said in a statement: "This is exactly why we built Apollo: to work alongside people in real industrial settings. With Gemini Robotics, we saw the model make significant gains in just a few weeks of post-training, and transfer to the Mercedes-Benz production line with minimal performance loss."

Apptronik's wording mentions "post-training" and "minimal performance loss," indicating that there is performance degradation from the lab to the production line, just within an acceptable range. This is a common problem in deploying robot AI—models trained in simulation environments often experience performance declines when they encounter noise, lighting changes, and part variations in the real physical world.

Competitive Landscape: The Window for Physical AI Is Narrowing

Google DeepMind chose to debut Gemini Robotics 2.5 in a factory scenario rather than a home or medical scenario. This choice reflects the reality of current competition in physical AI: industrial scenarios have clearly defined tasks, quantifiable success criteria, and relatively controllable fault-tolerance costs, making them the easiest entry point for VLA models to form verifiable business cases.

From the perspective of the industry chain, robot hardware manufacturers are the direct beneficiaries of this release—with the publicly quotable factory figure of 96%, Apptronik now has negotiating leverage it never had before when selling Apollo 2 to industrial customers. For automakers like Mercedes-Benz, the pilot provides a low-risk entry point for evaluating AI suppliers: even if this collaboration ultimately does not scale, it has accumulated first-hand data on the reliability boundaries of humanoid robots.

For robot developers not on the partner list, what they can do in the short term is to keep iterating on in-house models or open-source VLA frameworks, while watching for when Google opens broader access channels. If Google DeepMind announces a Gemini Robotics API access plan at its 2027 developer conference, that would be a clear signal of physical AI transitioning from a closed partnership model to a platform model. Another signal is whether the 96% success rate can be reproduced in more factory scenarios—only if a second and third automaker beyond Mercedes-Benz disclose similar data would it prove that this is not a one-off result deeply optimized for a specific production line.

At the current stage, the most accurate positioning of Gemini Robotics 2.5 is this: a prototype that has crossed the threshold of industrial usability, but is still several mass-production cycles away from being a "plug-in-and-use" platform product. The remaining 4% is the real battlefield that will determine the pace of commercialization.