Google DeepMind reported on November 10, 2026 that robots prepared parts at Mercedes-Benz factories with a 96% success rate. The company describes a proof-of-concept, not a factory-wide productivity result.
For manufacturers weighing automation and workers wondering how jobs may change, the distinction matters. Success at one task does not establish how much useful work a system can deliver across a shift.
Preparing parts for assembly
Google says Apptronik’s Apollo 2 robots picked metal brackets from bins and placed them in vehicle-assembly order at Kecskemét, Hungary, and Sindelfingen, Germany. This is kitting: preparing the right parts in the right order.
Gemini Robotics ER tracked the workflow while Gemini Robotics 2.5 controlled the hands and arms, according to Google. RohitAI’s earlier Gemini Robotics guide explains the distinction between reasoning and movement.
This is an ongoing collaboration: Apptronik announced its Mercedes-Benz pilot agreement in March 2024. In June 2026, it described Apollo 2 as a data-collection and training platform. A workplace trial can help develop the technology without yet establishing a case for buying it at scale.
A success rate needs a definition
Google’s report does not specify the scoring unit, attempt count, trial dates or duration, or rules for retries and human help. Comparable human-performance and full-shift output figures are also absent.
The practical implication for buyers is to define success around a correctly prepared container ready when the line needs it. A successful arm movement, a finished container and an on-time container are different measures. Without the scoring rules, the percentage cannot be converted into mistakes per car or lost production hours.
Recovery also matters. A system that needs assistance may save picking time but add supervision and correction work. That is a question for a pilot to measure, not a finding that these robots required such help.
What buyers and workers need to know
A useful next-stage evaluation would compare the existing process with the robot-assisted one under the same conditions. Before expanding a pilot, ask for:
Usable output. Correct parts in the right containers, delivered within the required time. Request results by site and task, not just one combined percentage.
Human effort. Time spent replenishing bins, supervising, correcting mistakes and restarting work. Include changeovers, charging and maintenance.
Total cost. Hardware, integration, service and staffing costs per usable completed kit, compared with the current process.
Safety needs its own evidence. Apptronik describes hardware zones that pause Apollo 2 when an object enters a defined area. That is a manufacturer’s design description. A buyer still needs installation-specific safety evidence; a task-success score cannot provide it.
For workers, the partnership’s stated aim is to support skilled staff by automating repetitive, physically demanding tasks. That is an intended benefit, not a measured staffing or workload outcome from this trial.
As of November 10, Google says Gemini Robotics 2.5 is rolling out to robotics partners. Its robotics website offers a trusted-tester waitlist, not guaranteed access. The announcement gives no public price or general-access date.
The useful next milestone is evidence that correctly prepared parts arrive on time, with a known amount of human help and a defensible operating cost. Those measures would let manufacturers judge whether this particular task is ready to move beyond a pilot.