Google DeepMind unveils Gemini Robotics 2 for adaptable, full-body robot control

Google DeepMind has introduced Gemini Robotics 2, a new intelligence layer designed to help robots operate beyond fixed routines and remote control. The system is built to interpret instructions, understand surroundings and coordinate movements across an entire robotic body.

Three models for embodied AI

The release includes three complementary models:

• Gemini Robotics 2 is a vision-language-action model that turns visual and spoken input into motor commands. It can control humanoid robots from their feet to their fingertips, as well as bi-arm systems.
• Gemini Robotics ER 2 is an embodied-reasoning model for communicating with people, understanding physical environments and planning tasks that unfold over several minutes. It also supports coordination between multiple robots.
• Gemini Robotics On-Device 2 is optimized for local operation and can adapt to a new robot design using only a few hours of data.

In a demonstrated scenario, an Apptronik Apollo 2 humanoid receives an instruction to move a watering can to a lower shelf. The robot walks to the object, picks it up, navigates to the shelf and places it in the requested location. DeepMind says whole-body control is important for real environments, where robots must bend, reach and balance around clutter.

The results also show the technology’s current boundaries. DeepMind reports medium-to-high success across whole-body and gripper-based tasks, while multi-finger manipulation remains difficult. Reported examples include stronger performance for some insertion and pick-and-place tasks than for tying bags or handling a dustpan.

For AI builders, the announcement points toward models that can transfer skills across different hardware rather than requiring a separate system for every robot. Gemini Robotics ER 2 is available in Google AI Studio and private preview through Gemini Enterprise Agent Platform; the VLA and on-device models are limited to early-access partners.

Source: Google DeepMind Blog