Google DeepMind Launches Gemini Robotics 2, a Single VLA Model for Arms to Humanoids

Google DeepMind has unveiled Gemini Robotics 2, a vision-language-action (VLA) model the company describes as its most advanced yet for controlling physical robots. According to DeepMind, the model can operate systems across the full range of robot form factors, from tabletop manipulator arms to full-body humanoid robots.

VLA models fuse three capabilities into a single system: image recognition, language processing, and action control. This combination lets a robot interpret its surroundings, understand instructions given in natural language, and translate both into physical movement. DeepMind positions Gemini Robotics 2 as an "intelligence layer" that can sit on top of varied hardware rather than a model built for one specific robot body.

According to DeepMind, Gemini Robotics 2 can manage full-body movement, execute fine motor tasks, and coordinate multiple robots working together. The company has not disclosed benchmark scores, parameter counts, pricing, or a training cutoff date for the model. Developers can request early access through a waitlist; general availability details have not been announced.

Alongside the flagship model, DeepMind released Gemini Robotics ER 2, a companion model built for what the company calls "embodied reasoning." DeepMind defines this as the ability to understand the physical world and decide on a course of action based on that understanding. In practice, ER 2 functions as a higher-level planning and decision-making layer that sits above the lower-level motor control handled by Gemini Robotics 2.

ER 2 replaces Gemini Robotics ER 1.6, which DeepMind released in April 2026. Unlike the flagship model, ER 2 is already available in Google AI Studio, giving developers immediate access to test it rather than joining a waitlist.

DeepMind has not published independent benchmark results for either model, and none of the performance claims in this announcement have been verified by third parties. No pricing has been disclosed for API or licensing access to either model.

What this means

Google is betting that a single, general-purpose model can control robots with fundamentally different bodies — a two-armed tabletop manipulator and a bipedal humanoid don't share the same joints, sensors, or failure modes. If Gemini Robotics 2 delivers on that claim, it would let robotics companies skip building bespoke control software for each new hardware platform, similar to how general-purpose language models replaced task-specific NLP systems.

The split between Gemini Robotics 2 (low-level control) and Gemini Robotics ER 2 (high-level reasoning) mirrors a common architecture in robotics research, where a planner decides what to do and a separate controller decides how to physically do it. Making ER 2 available in Google AI Studio while gating the flagship model behind a waitlist suggests DeepMind wants developers experimenting with reasoning and task planning now, while it manages capacity or safety testing on the motor-control side more tightly.

Without published benchmarks, independent evaluation, or pricing, it's too early to judge how Gemini Robotics 2 compares to rival systems from companies building embodied AI, including Physical Intelligence, Figure, and NVIDIA's robotics efforts. The lack of technical specifics — no parameter count, no context window, no training data cutoff — means claims about its capabilities remain DeepMind's own until developers with waitlist access can test it against real hardware.