What changed
Google DeepMind introduced Gemini Robotics 2.5, a vision-language-action model for generalist robot control, and began rolling it out to robotics partners. DeepMind says a Mercedes-Benz production-line proof of concept with Apptronik Apollo 2 reached 96% kitting success and reports broader gains across 200 tasks.
Why it matters
The release links frontier multimodal reasoning to industrial manipulation across different robot forms. For manufacturers, the useful signal is a partner test with a concrete task and success metric—not a generally available robot model.
What to do
If eligible, request a bounded partner evaluation with a fixed task set, safety envelope, intervention log and independent acceptance test. Everyone else should wait for public access, model documentation and replicated reliability evidence.
Caveat
The model is rolling out to partners rather than a public API or open checkpoint. The 96% kitting result, 30–100% success gains, speed gains and safety scores are vendor or selected-partner evidence without independent replication.