Google DeepMind brings Gemini Robotics 2 to robots with whole-body control and collaboration

Google DeepMind introduced Gemini Robotics 2, a new generation of robotics models that seeks to move physical systems beyond repetitive or teleoperated tasks. The company describes the launch as an intelligence layer for robots capable of reasoning about whole-body movements, manipulating objects with greater dexterity, and coordinating with other robots in shared workflows.

The announcement was published on July 30 by Google DeepMind and accompanied by a demonstration on YouTube in which robots identified as Apollo and Duo carry out organizing and collaboration tasks. The video transcript does not serve as an extensive technical explanation; it shows commands and responses from the robots during the test. For that reason, the main factual basis for this article is DeepMind’s official announcement, Google’s blog about Gemini Robotics ER 2, and independent technical coverage.

Three models within the launch

DeepMind frames Gemini Robotics 2 as a family of three models. Gemini Robotics 2 is the vision-language-action model that converts visual and language instructions into motor control for robots. According to the company, it can operate humanoid robots and two-arm systems, including tasks that combine locomotion, posture, and object manipulation.

Gemini Robotics ER 2 plays a different role: it acts as a high-level reasoning layer. Google describes it as a “brain” for robotics applications, capable of understanding real-time video, planning multistep tasks, calling tools such as Google Search or developer-defined functions, and handing execution over to lower-level action models. Google says ER 2 is available through Gemini API, Google AI Studio, and, in private preview, Gemini Enterprise Agent Platform.

The third component is Gemini Robotics On-Device 2, designed to run locally on the robot. That part is important because robotics cannot always rely on cloud latency: many physical decisions require an immediate response, continuity, and adaptation to the environment.

What Google showed and what it does not yet prove

The central promise is that a robot can observe a scene, interpret a human instruction, break it into steps, and adjust the plan if something goes wrong. In its examples, DeepMind talks about robots that clean up a messy room, handle tools, or work together to complete a task that a single body would not accomplish as quickly.

MarkTechPost’s technical coverage agrees that the launch combines whole-body control, dexterity, and multi-robot collaboration, and highlights that the advance is organized into models with different levels of access. Even so, these demonstrations do not amount to general-purpose robots ready to operate without supervision in homes, hospitals, or factories. DeepMind acknowledges that manipulation with multiple fingers remains challenging, even as it reports better results in whole-body tasks and those involving pincers or grippers.

The news within the race for “physical AI”

The launch is significant because it shows where AI competition is heading: from chatbots and software agents toward systems that make decisions in physical spaces. In that domain, the model must not only produce correct text; it must coordinate perception, planning, tools, motor control, and safety.

It also signals a platform strategy. By offering ER 2 through Gemini API and AI Studio, Google is trying to get developers and businesses to experiment with robotics using pieces of the Gemini ecosystem, not just closed models inside laboratories.

The most cautious editorial interpretation is that Gemini Robotics 2 represents an important technical step in robotics based on multimodal models, but does not yet demonstrate mass commercial adoption or fully resolved autonomy. What is confirmed is the launch, the capabilities Google reports, the announced access routes, and the published demos. What remains to be seen is how this architecture behaves outside controlled environments, with real costs, safety standards, and less predictable physical tasks.