DeepMind launches Gemini Robotics 2: whole-body VLA, ER 2, on-device

Google DeepMind launched the Gemini Robotics 2 family on 30 July 2026: three models for whole-body control, multi-step embodied reasoning, and local on-robot adaptation. The VLA product keeps the Gemini Robotics 2 name; ER 2 and On-Device 2 are separate siblings, not modes of one checkpoint. Prior Gemini Robotics work centered upper-body tabletop control; this release extends motor control to full humanoids and multi-robot teamwork, per DeepMind.

From the 30 July 2026 @GoogleDeepMind post:

"One brain. For any robot. We're launching Gemini Robotics 2: our next-generation physical AI bringing full body intelligence to humanoids, advanced dexterity, multi-robot teamwork and more."

The DeepMind launch post (Carolina Parada) maps the stack. Gemini Robotics 2 is the vision-language-action model that maps vision and language to motor control for full humanoids and bi-arm platforms, including multi-finger hands and parallel grippers. Gemini Robotics ER 2 is the embodied-reasoning vision-language model: high-level agent for multi-step plans lasting several minutes, progress tracking on continuous video, tool calls, and multi-robot collaboration. Gemini Robotics On-Device 2 is the efficiency VLA meant to run locally and adapt to new bi-arm bodies in a few hours, typically with fewer than 200 examples.

Access is split. DeepMind and Google say ER 2 is available via the Gemini API, Google AI Studio, and private preview on the Gemini Enterprise Agent Platform. The VLA and On-Device models are limited to early-access partners. Launch demos include Apptronik Apollo 2 (whole-body walk and place), SharpaWave multi-finger hands, Franka Duo grippers, and multi-robot Apollo plus Franka collaboration. A developer write-up adds Live API streaming, a Boston Dynamics Spot orchestration sample, and vendor evals (57.4% progress-classification and 91.3% moment-finding accuracy on their suites). Safety notes include ASIMOV-Agentic and a Gemini Robotics 2 Safety Technical Report.

Key facts

• Social alert: @GoogleDeepMind on X, 30 July 2026
• Official primary: Gemini Robotics 2 brings whole body intelligence to robots
• Three models (official): Gemini Robotics 2 (VLA), ER 2 (embodied reasoning), On-Device 2 (local VLA)
• Access (official): ER 2 on Gemini API, AI Studio, Enterprise private preview; VLA and On-Device early-access partners only
• On-device adaptation (vendor): few hours of data, typically under 200 examples
• Safety (vendor): ASIMOV-Agentic; safety technical report PDF

How to read

This is a robotics model-family launch with gated VLA and on-device access, not open-weight day and not a general Gemini chatbot upgrade. Charts and multi-finger demos are DeepMind-reported; DeepMind still frames dexterity and movement speed as work in progress. Partner demos show lab and partner setups, not a universal deploy guarantee.

Independence

Frontier Flux is not affiliated with Google, DeepMind, Apptronik, Franka, or Boston Dynamics. Quotes and metrics come from DeepMind and Google developer primaries plus the public X alert.