Embodied AI Glossary中文

Gemini Robotics 2

Common

Google DeepMind's 2026 robotics model generation, with the new ability to control a humanoid's whole body.

Google DeepMind released this on July 30, 2026, as the generation following 1.5, made up of three models: the VLA model Gemini Robotics 2, the embodied-reasoning model Gemini Robotics ER 2, and an On-Device 2 version that runs locally on the robot. The main new capability is full-body control: it can drive an entire humanoid robot while it walks, crouches, and manipulates objects at the same time, and Google demonstrated multistep tasks such as retrieving items from a shelf on the Apptronik Apollo 2. It can also control the 22-degree-of-freedom, five-fingered SharpaWave dexterous hand, as well as ordinary two-finger grippers; adapting On-Device 2 to a new dual-arm embodiment takes only a few hours. The VLA and On-Device versions are currently available only to early partners, while ER 2 is accessible through Google AI Studio.

ExampleUnder Gemini Robotics 2's control, an Apptronik Apollo 2 humanoid walks to a shelf, crouches, and retrieves an item, completing a multistep task.

Related
Gemini Robotics 1.5 · Gemini Robotics-ER · Gemini Robotics On-Device · Whole-Body Control · Apptronik Apollo · SharpaWave
Sources
Gemini Robotics 2 brings whole body intelligence to robots (Google DeepMind)
Gemini Robotics 模型页 (Google DeepMind) (Chinese)
As of
2026-07

See it in the full glossary →