Google DeepMind's Gemini Robotics 2 moves robot control from table-top to whole body
Google DeepMind released three models: a vision-language-action model that walks and manipulates on Apptronik's Apollo 2, including its 22-degree-of-freedom five-fingered hand, an embodied-reasoning model that plans multi-minute tasks and coordinates multiple robots, and an on-device model that adapts to a new robot body in a few hours. The jump from upper-body table-top tasks to whole-body motion is the real advance, and adapting to a new embodiment on fewer than 200 examples makes hardware access, not demonstration data, the scarce input in robotics. Apptronik, Boston Dynamics, and Agile Robots are the named partners; ER 2 is in Google AI Studio and private preview, while the action models remain early-access by application.
Source: deepmind.google ↗
For the first time, our model can now control entire humanoid robots, translating intent into intelligent whole-body control.
Why this matters
- → Robots can now operate full bodies—walking, reaching, manipulating—not just tabletop arms
- → Adapts to new robot designs in hours with minimal data; hardware access becomes the constraint, not training d
- → Multi-step reasoning + multi-robot coordination unlock complex workflows no single robot could solo