Google reveals Gemini Robotics 2.0, promising improved dexterity and safety
Gemini Robotics 2 includes three models, but only one is publicly available right now.
Google has rolled out Gemini Robotics 2.0, a three-model package aimed at pushing its robots from scripted stunts toward something resembling general-purpose behavior. The headline pitch is whole-body control, finer dexterity, and the ability for machines to reason about a changing room rather than replay a fixed routine.
The structural story is the model split. Only one of the three sub-models is open to developers today. The others remain internal, which is the usual pattern: ship the safe, useful slice to the ecosystem while keeping the higher-leverage reasoning layers in-house. That choice tells you where the real value is being captured.
The "physical AGI" framing is the part worth watching. Google DeepMind is openly positioning these robots as instruction-followers rather than pre-programmed performers. If that claim holds up in production settings, the labor implications are not subtle: tasks currently done by humans in warehouses, kitchens, and care environments become candidates for substitution on a much shorter timeline than the old robotics roadmap suggested.
For now, the release is a signal, not a product. Developers get an API, a set of demos, and a vocabulary shift from narrow automation to generalist embodiment. The market will read it as Google drawing a line in the sand against rivals chasing the same humanoid-robot prize.