Google DeepMind has introduced Gemini Robotics 2, a new suite of AI models built to help robots move, think, and work together more naturally. According to a blog post published by Google DeepMind, the release was announced on July 30, 2026, and brings together three separate models. Each model handles a different part of what makes a robot genuinely useful outside a lab.
Let’s take a closer look at the three models and understand what each one brings to the Gemini Robotics 2 suite.
Model 1: Gemini Robotics 2 for Whole-Body Humanoid Control
The first model, Gemini Robotics 2, is a vision-language-action model that turns what a robot sees and hears into real movement. Earlier versions mostly handled arms and hands, but this one can control an entire humanoid body. In one demo, Apptronik’s Apollo 2 robot was asked to place a watering can into a bin on a bottom shelf. It walked to the table, picked up the can, crossed the room, and set it down in exactly the right spot. The same model also improves fine hand control, letting Apollo 2 tie knots or seal a ziplock bag using a five-fingered hand with 22 degrees of freedom.
Model 2: Gemini Robotics ER 2 for Multi-Step Reasoning
The second model, Gemini Robotics ER 2, works as the system’s reasoning layer. It plans multi-step tasks, watches live videos to track how far a robot has progressed, and knows exactly when one step ends, and the next begins. It can also coordinate more than one robot at a time, splitting a task between machines that couldn’t finish it alone. In one demo, DeepMind paired this model with Boston Dynamics’ Spot robot to fetch a snack on a simple spoken request.
Model 3: Gemini Robotics On-Device 2 for Offline AI Control
The third model, Gemini Robotics On-Device 2, is designed for situations where a stable connection isn’t always available. It runs directly on the robot itself and can adapt to a completely new robot body in just a few hours, often using fewer than 200 examples.
Key Features of Gemini Robotics 2
The combined capabilities of all three models bring several major improvements to real-world robotics.
- Whole-body humanoid control — Robots can now walk, crouch, and manipulate objects using the entire body, not just the arms.
- Video-based progress tracking — Robots judge task completion in real time instead of relying on fixed timers.
- Multi-robot collaboration — Different robots can divide a task between them based on their strengths.
- Offline adaptability — New robot bodies can be supported without a live network connection.
How Gemini Robotics 2 Improves Robot Safety
DeepMind claims that Gemini Robotics ER 2 is the safest robotics model yet. It can detect nearby people better and efficiently bring a robot to a safe stop until an area clears. Moreover, the company also introduced a new benchmark, ASIMOV-Agentic, to test whether the reasoning model refuses unsafe actions and asks a human for help instead of guessing.
Gemini Robotics 2 Availability and Early Access
Gemini Robotics ER 2 is now available through the Gemini API and Google AI Studio. The other two models are rolling out to early-access partners, including Apptronik, Boston Dynamics, and Agile Robots, as Google continues building toward robots that can adapt to real, unpredictable environments on their own.
Conclusion
Gemini Robotics 2 shows how far physical AI has come in a short time. Google DeepMind doesn’t use just one model to complete all tasks. Rather, it splits the work into three specialized parts: one for movement, one for reasoning, and one for working offline. Together, allow robots to operate better in real, complex environments instead of just following repeated fixed routines. With Apptronik, Boston Dynamics, and Agile Robots already testing the system, this update looks like an early step toward robots that can adapt on their own, wherever they’re needed.
This robotics push is part of a larger pattern. Google has been extending the Gemini name across very different products, from the ChatGPT vs Gemini chatbot rivalry all the way to physical robots.
FAQs :
What is Gemini Robotics 2?
It’s a suite of three AI models built to give robots better movement, reasoning, and offline capabilities.
What does Gemini Robotics ER 2 do?
It acts as the robot’s reasoning brain, planning multi-step tasks and coordinating multiple robots together.
Can these robots work without internet access?
Yes, Gemini Robotics On-Device 2 runs locally and can adapt to new robot bodies within hours.
Which robots have been tested with Gemini Robotics 2?
Apptronik’s Apollo 2, Boston Dynamics’ Spot, and Franka Duo have all been used in demonstrations.
Is Gemini Robotics 2 available to the public yet?
Gemini Robotics ER 2 is available now, while the other models are in early access with select partners.
