Google DeepMind releases Gemini Robotics 1.5

Background

Industrial robots usually execute fixed, pre-programmed motions. Meet a new object, a new room or a slightly different instruction, and someone has to rewrite the program. Foundation models for robotics aim to change that by giving a robot broad, general-purpose competence that can be directed with ordinary language instead of task-specific code.

What was announced

On 25 September 2025, Google DeepMind announced two models, Gemini Robotics 1.5 and Gemini Robotics-ER 1.5.^[1] The company described them as models that bring AI agents into the physical world, letting robots perceive, plan, think, use tools and act.^[1] Google DeepMind also described Gemini Robotics 1.5 as its most capable vision-language-action model.^[1]

A vision-language-action (VLA) model is one that takes in images and a language instruction and produces actions. In general terms, such a model maps what the robot sees and what it is told directly onto what it does, instead of relying on human engineers to wire perception, planning and control together by hand.

Two models, two roles

The two releases split the work. Gemini Robotics 1.5 is the model that actually drives motion, the one Google DeepMind calls its most capable VLA model.^[1] Gemini Robotics-ER 1.5 works at a higher level: the company described it as a high-level brain that orchestrates a robot's activities.^[1] It can also natively call tools such as Google Search.^[1]

That split matters because the two jobs have different demands. Deciding what should be done next, and looking things up when information is missing, is mostly a reasoning problem; turning a decision into joint movements is mostly a control problem. Tool calling gives the planning model a way to reach beyond what it already knows, for example by querying a search engine rather than answering from memory alone.

Reasoning before acting

One detail Google DeepMind highlighted is that the VLA model reasons in natural language before it acts.^[1] Instead of jumping straight from an image to a motor command, the model produces an intermediate verbal description of what it is about to do, and then acts on that.

Such intermediate steps are a common technique in AI systems, and they have a practical benefit for robots: a plan expressed in words can be inspected, compared against the instruction, and corrected before anything moves. It also keeps the model's behaviour legible to the people working alongside it, who can read the reasoning rather than infer intent from motion alone.

Skills that carry across bodies

Google DeepMind said motions learned on one robot could transfer to robots with different embodiments.^[1] In robotics, embodiment refers to the physical form of a machine: its arms, grippers, sensors and the way those parts are arranged. Two robots with different embodiments cannot simply share a stream of joint angles, because the joints do not correspond.

Transfer of this kind is one of the harder problems in the field, since a skill demonstrated on one machine has to be expressed in a way that survives a change of body. Being able to carry motions across embodiments is what would let a capability learned once be reused across a fleet of differently built machines rather than retaught for each one.

How developers can get it

Access differed between the two models. Gemini Robotics-ER 1.5 was offered to developers in preview through the Gemini API.^[1][2] Gemini Robotics 1.5, the model that produces motion, was at that point available to select partners.^[1]

Why it matters

The framing Google DeepMind used is the useful summary: these are models meant to bring AI agents into the physical world, where a robot perceives, plans, thinks, uses tools and acts.^[1] Language models on their own produce text; a VLA model closes the loop by producing movement, and a planning model sitting above it decides what the movement should accomplish. Announcing the two together, with a preview route for developers through an API, put the pieces of that loop in front of outsiders rather than keeping them inside a single laboratory demonstration.