Gemini Robotics 1.5

What it is

Gemini Robotics 1.5 is a vision-language-action model: it takes in camera images and a spoken or written instruction, and outputs the actions a robot should carry out. It was announced by Google DeepMind on 25 September 2025 together with a second model, Gemini Robotics-ER 1.5, which is built for embodied reasoning — reasoning about objects, space and physical outcomes rather than about text alone.^[1] DeepMind described Gemini Robotics 1.5 as its most capable vision-language-action model, and Gemini Robotics-ER 1.5 as its most capable vision-language model for reasoning about the physical world.^[1]

The two are meant to be used as a pair. Rather than cramming planning and motor control into one network, the release splits the job: one model decides what should happen, the other turns that decision into movement.

Reasoning before acting

The distinguishing feature of Gemini Robotics 1.5 is that it can think before taking action, generating internal reasoning in natural language before it moves.^[1] In a conventional action model, an instruction goes in and a motion comes out. Here, an intermediate step appears first: the model writes out, in ordinary language, what it takes the task to be and how it intends to proceed, and then produces the action.

That intermediate step matters for two reasons. It lets the model break a long task into parts and keep track of where it is, which is hard to do if every decision has to be made in a single jump from image to motor command. And because the reasoning is in natural language rather than in an opaque internal code, it is something a person can read and check — useful when a robot is about to do something expensive or unsafe.

A planner that can use tools

Gemini Robotics-ER 1.5 handles the other half. It acts as a high-level planner and orchestrates a robot's activities like a high-level brain, deciding what the sequence of steps should be and handing the pieces down.^[1] It can also natively call tools such as Google Search.^[1]

Tool calling changes what a robot planner can know. A model's weights are fixed at training time, so anything that happened afterwards, or any detail too specific to have been learned, is out of reach. Letting the planner call a search tool means it can look that information up during a task instead of guessing. The planner is therefore not only deciding how to move, but also gathering whatever background it needs in order to decide.

One set of skills, different bodies

Robot learning has a long-standing problem: a policy trained on one machine tends to fail on another, because bodies differ in size, joint layout, reach and sensors. Tasks trained only on the ALOHA 2 robot also worked on Apptronik's Apollo humanoid and on a bi-arm Franka robot, and the reverse also held.^[1]

The two-way transfer is what matters here. What carries over is something about tasks and objects, which survives a change of body, rather than the memorised motor habits of a single machine. A lab can teach a skill on one robot and expect that skill to run on hardware it does not have.

Who could use it, and what came next

Access was split. Gemini Robotics-ER 1.5 was made available to developers through the Gemini API in Google AI Studio, where it appeared as a preview model with a release date of 25 September 2025, while Gemini Robotics 1.5 was available only to select partners.^[1][2]

Further previews followed in the Gemini API: gemini-robotics-er-1.6-preview on 14 April 2026, and gemini-robotics-er-2-preview on 30 July 2026.^[2] The 1.5 preview was listed with a shutdown date of 30 April 2026.^[2]

Where it sits

The release is a statement about how robotics models are being assembled: a reasoning model that plans and looks things up, plus an action model that turns plans into motion, with the two communicating in natural language. It also pushes the idea of embodiment-independent skills — the same trained task running on a tabletop arm pair and on a humanoid — from a research curiosity toward something developers are expected to rely on.