Google DeepMind Introduces Gemini Robotics 2 for Whole-Body Control
The three-model release adds whole-body motion, longer task planning, multi-robot coordination and an on-device path for adapting to new hardware.
Google DeepMind introduced Gemini Robotics 2 on July 30 as a three-model family for controlling and coordinating robots. The official release covers a vision-language-action model for whole-body movement, an embodied-reasoning model for longer tasks and teamwork, and an on-device model designed to adapt to new robot hardware with limited training data.
What changed
Gemini Robotics 2 converts visual and language inputs into motor control for full humanoids and other two-arm robots. Google says the same model checkpoint can work across several robot bodies, extending its earlier robotics work from tabletop manipulation to actions that require walking, crouching, reaching and handling objects.
Gemini Robotics ER 2 acts as the higher-level planner. It interprets instructions, observes the scene, breaks work into steps, communicates with people and coordinates with the motor-control model. Google says it can manage sequences lasting several minutes and involving hundreds of decisions, recover when a step fails, and coordinate multiple robots on one workflow.
The third model, Gemini Robotics On-Device 2, runs locally where network latency or connectivity is unsuitable. DeepMind says it can adapt to a new two-arm robot with a few hours of data and typically fewer than 200 examples. Those figures are vendor-reported and have not been independently verified.
Availability and limits
Access differs across the family. The ER 2 reasoning model is available through Google AI Studio and in private preview on the Gemini Enterprise Agent Platform. The vision-language-action and on-device models are limited to early-access partners. That means the announcement is a real model release, but not every component is generally available.
DeepMind also acknowledges unfinished work. Its published examples show that multi-finger manipulation remains difficult, and the company says robot movement speed still needs improvement. The demonstrations and evaluations come from Google, so real-world reliability across unfamiliar hardware, workplaces and safety conditions remains an open question.
Why it matters
The release joins perception, planning and physical control in one product family while separating cloud reasoning from local execution. For robotics teams, the practical signal is not just a stronger demonstration: it is a clearer deployment stack, with public access to the reasoning layer and controlled access to the models that move hardware.
Status
Confirmed. Google DeepMind announced the models and access routes; capability and safety claims remain vendor-reported, so internal confidence is medium.
Sources
Update note: Last reviewed 2026-08-02. We will revise this post as access broadens or independent evaluations become available.
Sources
- Google DeepMind - Gemini Robotics 2 — official
Drafted with AI assistance from source briefs; reviewed for citation completeness and label accuracy.