IllustrationNo official image on file. Showing a schematic category illustration.
DeepMind's most advanced VLA, extending control from tabletop arms to whole-body humanoids with advanced dexterity and multi-robot teamwork.
Key specs
| Product type | VLA |
|---|---|
| Embodiments | Humanoids (feet to fingertips), bi-arm robots, 22-DoF hands and grippers |
| Availability | Early-access partners |
Technical details
Gemini Robotics 2 is Google DeepMind's most advanced VLA, extending control from tabletop arms to whole-body humanoid control (walking, crouching, manipulation) and advanced dexterity. A single checkpoint controls Apptronik Apollo 2 with SharpaWave or Inspire hands and a Franka Duo with Robotiq grippers; DeepMind reports medium-to-high success on whole-body and gripper dexterous tasks while multi-finger dexterity remains challenging. Long-horizon tasks are orchestrated by Gemini Robotics ER 2, which also enables multi-robot collaboration. Access is limited to early-access partners.
| Architecture | Vision-language-action (VLA) model built on Gemini, converting vision and language input into motor control; paired with Gemini Robotics ER 2 as high-level planner/orchestrator. |
|---|---|
| Inputs outputs | Inputs: vision + natural-language instructions; outputs: motor control for whole-body humanoids, bi-arm robots, dexterous hands and grippers. |
| Embodiments supported | Apptronik Apollo 2 humanoid (with SharpaWave 22-DoF hands or Inspire hands) and Franka Duo with Robotiq gripper, using the same checkpoint; full humanoids and other bi-arm robots. |
| License weights | Proprietary; available to early-access partners. |
| How to try | deepmind.google/…/vla |
| Sources |
