Google DeepMind announces Gemini Robotics 2.5; Apollo 2 factory proof of concept reportedly achieves 96% success rate

Google DeepMind announced “Gemini Robotics 2.5,” a vision-language-action (VLA) model for robots, on Nov. 9, 2026. roboticsandautomationnews.com reported on Nov. 15 that the model was evaluated on more than 200 tasks and demonstrated on a Mercedes-Benz production line. According to the article, Apptronik’s Apollo 2 performed automotive-parts kitting and recorded a 96% success rate in a proof of concept.

Google DeepMind describes Gemini Robotics 2.5 as a VLA model that improves the range of skills and learning speed. The reported 96% on the Mercedes-Benz production line was for an Apollo 2 proof of concept involving parts kitting. Other metrics, such as operating time and human intervention, are needed to assess continuous factory operation or production-scale deployment.

The key point in the model announcement is that it pairs the company’s claims of broader skills and faster learning with a factory proof of concept using Apollo 2. The former involves evaluations across multiple tasks; the latter assesses a specific parts-kitting task. The 96% success rate cannot be treated as an indicator of continuous factory uptime or the performance of an entire production line. Improvements across more than 200 tasks are also task-specific results; they do not show that the same benefits apply to every operation. We covered the announcement and the same proof of concept in Google DeepMind announces Gemini Robotics 2.5; 96% success rate in automotive-parts picking. The latest report describes the types of evaluation tasks and comparison figures against the previous version, but provides limited information to assess the conditions under which the 96% result was achieved. Benchmark improvements and the ability to operate reliably and repeatedly in a real-world setting therefore need to be evaluated separately. For the use of robots in real-world work, an important point in the report is the combination of Apollo 2’s arm and hand manipulation with task planning by Gemini Robotics ER, rather than the model alone. However, the report describes a proof of concept and does not specify the types of parts, task cycle, operating time, exception handling or frequency of human intervention. The work differs in both tasks and status from the Apptronik Apollo logistics-site trial and deployment plans covered in the article about GXO, so the two should not be grouped as the same deployment record. GXO and Apptronik expand Apollo warehouse trials, plan deployment of up to 1,000 units The next details to verify are the evaluation period for the factory demonstration, how task success was defined, real-world operating metrics that account for retries and human intervention, and whether the work will expand to other processes. Even with published model improvements, actual operating costs, safety and stability across multiple shifts remain unclear. See the original source for details.