Overview
Gemini Robotics 2 is Google DeepMind’s vision-language-action (VLA) model for robot control. It converts visual and language inputs into motor actions and can control humanoid bodies and dual-arm robots. It was announced on July 30, 2026. (deepmind.google)
Capabilities and uses
- Whole-body control: Enables humanoids to walk, bend, reach, balance and manipulate objects, including for tasks such as putting objects away. (deepmind.google)
- Dexterous manipulation: Works with robotic hands and grippers; demonstrated tasks include tying knots, sealing a ziplock bag and tightly packing objects. (deepmind.google)
- Coordinated tasks: In a system paired with the separate Gemini Robotics ER 2 reasoning model, robots can plan multi-step work, track progress, recover from some failures and collaborate with other robots. (deepmind.google)
Availability and limitations
At announcement, the VLA model was available to early-access partners. Google reported that multi-finger manipulation remained challenging and that precision and speed still needed improvement. Physical safeguards such as emergency stops, barriers and speed or force limits remain necessary; model safety measures do not replace them. (deepmind.google)
Gemini Robotics On-Device 2 is a separate, locally optimized VLA model. Google reported adapting it to new dual-arm robot embodiments with a few hours of adaptation time, typically using fewer than 200 examples; this should not be assumed to describe the deployment requirements of Gemini Robotics 2 itself. (deepmind.google)
Pricing
As of November 16, 2026, a public USD price for Gemini Robotics 2 VLA input or output could not be verified. Google’s published Gemini Robotics ER 2 Preview token rates apply to a separate embodied-reasoning model, not to this VLA model. (aistudio.google.com)
Official sources
- Google DeepMind announcement
- Gemini Robotics 2 safety evaluations
- Gemini API pricing — checked November 16, 2026.