Gemini Robotics 2.5 Reaches Factory Trials, Where 96% Is a Starting Point

Chinese

Google DeepMind introduced Gemini Robotics 2.5 on November 10 and reported a 96% success rate in automotive kitting proof-of-concepts with Apptronik and Mercedes-Benz. The model is rolling out to robotics partners; the results do not establish general availability or production-wide reliability.

AI-generated concept illustration: abstract grippers and a parts tray symbolize dexterous manipulation

Sorting the right automotive parts into the right tray is less theatrical than a robot tying a bow, but it gives a manipulation model a concrete job to finish. On November 10, Google DeepMind introduced Gemini Robotics 2.5 and reported a 96% success rate in automotive kitting proof-of-concepts with Apptronik and Mercedes-Benz. [R1]

The work involved Apollo 2 humanoids on active production lines in Kecskemét, Hungary, and Sindelfingen, Germany. Google describes an embodied-reasoning model tracking the workflow while Robotics 2.5 controls hands and arms. The announcement identifies this as a proof-of-concept; it does not disclose the trials’ exact dates or establish production-wide performance. [R1]

The missing four percent cannot be interpreted without knowing what counts as one attempt. A failed grasp, a wrong part and a delayed handoff impose different costs. An industrial evaluation therefore needs the task denominator, cycle-time distribution, intervention frequency and recovery procedure, alongside the headline success rate. None can be inferred from a short demonstration.

DeepMind says it tested the model across more than 200 tasks, including tool use and manipulation of flexible or delicate objects. It reports success-rate gains of 30% to 100% and task completion 20% to 30% faster than previous baselines, depending on the task. These are developer evaluations, not a uniform improvement applicable to every robot or factory. [R1]

Breadth and reliability answer different questions. A model that can attempt many operations may be easier to adapt to a new workstation. A production process still needs assurance that its particular operation will finish within tolerances, and that failures leave equipment and people in a controlled state. A demonstration of generality should not replace acceptance testing for the actual station.

The release is rolling out to robotics partners. Ars Technica also describes partner access rather than general availability, and industry coverage from the Association for Advancing Automation corroborates the announcement and testing scope. This report has not operated the model or independently replicated the factory result. [R1][R2][R3]

Sources:automate.org

For a manufacturer, a productive pilot would start with an operation whose inputs, layout and exception cases can be documented. It should count retries and human rescues as well as successful cycles. The resulting record would show whether a more flexible system reduces integration work without shifting an unreasonable supervision burden onto technicians.

The reported factory work is a useful milestone because it connects a research model to a recognizable production task. Its next test is sustained, auditable operation: how performance changes across shifts, different parts, wear and interruptions. That is the distance between a promising proof-of-concept and a dependable production tool.

Sources:

[R1] Google DeepMind model announcement|2026-11-10
[R2] Ars Technica coverage|2026-11-10
[R3] Association for Advancing Automation|2026-11-10