TL;DR Gemini Robotics 2 is Google DeepMind's newest robotics model family, arriving as part of its October 2026 physical-AI push. It ships in three forms: Gemini Robotics 2 (a vision-language-action model that turns vision and language into motor control), Gemini Robotics ER 2 (an embodied-reasoning model for multi-step planning), and Gemini Robotics On-Device 2 (a lightweight version that runs locally on robot hardware).

The short version

Most robots are trained to repeat one specific task. Gemini Robotics 2 is built to generalize — to complete tasks it hasn't been trained on, adapt to new situations on the fly, and bring whole-body control to humanoid robots from feet to fingertips[1].

Rather than a single model, Google frames this as a two-part system: a vision-language-action (VLA) model that acts, and an embodied-reasoning (ER) model that plans and coordinates — with a third, smaller variant optimized to run on the robot itself[1].

Humanoid robot, representing Google DeepMind's Gemini Robotics 2 models

The three models

ModelRoleWhat it does
Gemini Robotics 2Vision-language-action (VLA)Converts vision and language input into motor control so a robot can act
Gemini Robotics ER 2Embodied reasoningReasons within physical spaces to make detailed plans and coordinate with humans and other robots
Gemini Robotics On-Device 2Efficient VLAA lightweight VLA optimized to run locally on robotic hardware

The VLA and ER models have distinct roles but are designed to operate as one system — the ER model maps out a multi-step plan while the VLA model executes it[1].

What it can do

Google highlights a handful of capabilities that separate Gemini Robotics 2 from earlier systems[1]:

  • Powers any embodiment — it can be adapted to any bi-arm robot in just a few hours, scaling from arms to full humanoid bodies.
  • Advanced dexterity — enough finesse for delicate actions like screwing in a light bulb and tying knots.
  • Whole-body control — controls an entire humanoid from feet to fingertips, enabling full-range, human-like movement.
  • Multi-robot collaboration — two robots can divide labor and coordinate on a single task.
  • Agentic reasoning — deep spatial reasoning paired with long-horizon planning, so robots can complete unfamiliar, multi-step tasks.

It's also designed to be interactive: a robot can explain its approach while acting and be redirected in plain language, which Google says is useful for instructing robots through hazardous environments[1].

Partners and availability

Google is working with robotics hardware leaders to deploy the models, naming Agile Robots, Apptronik, and Boston Dynamics among its research partners[1]. The company says it's collaborating with 100+ trusted testers — from enterprise automation teams to robotics startups — and is accepting sign-ups for a waitlist for early access[1].

The rollout sits alongside Google's broader October 2026 frontier-model push, with Gemini Robotics 2 listed as a new model on the company's models index[2].

Sources

  1. Google DeepMind — Gemini Robotics — deepmind.google
  2. Google DeepMind — Models index — deepmind.google