Google DeepMind Unveils Local AI Model for Robots

Google DeepMind has launched a new language model, Gemini Robotics On-Device, designed to enable robots to operate locally without an internet connection.

We’re bringing powerful AI directly onto robots with Gemini Robotics On-Device. ?

It’s our first vision-language-action model to help make robots faster, highly efficient, and adaptable to new tasks and environments — without needing a constant internet connection. ? pic.twitter.com/1Y21D3cF5t

— Google DeepMind (@GoogleDeepMind) June 24, 2025

“This is our first Vision-Language-Action (VLA) model, which will help make robots faster, highly efficient, and adaptable to new tasks and environments — without the need for a constant internet connection,” highlighted the AI division of Google.

Sergey Lonshakov, architect of the “Robonomics” project, commented to ForkLog that VLA is an advanced solution in the field of humanoid robotics.

“Figure is currently demonstrating its collaborative production scenarios using the same type of models,” he noted.

In February, Figure introduced its own artificial intelligence, Helix, for integration with robots. According to the creators, the model is capable of “reasoning like a human.” AI-equipped humanoids can handle “virtually any household items without any code or pre-training.”

Lonshakov described the local deployment of the model on the device as a correct decision that aligns with current trends.

“[…] It’s called a high-level task planner — it used to be only in the cloud, when it was exclusively RL used for training a simple model to act in a simulation, and then the ready sequence of operations was uploaded to the robot. Now roboticists are trying to create seamless models, where at the planning stage there are no processes stopping the robot’s operations when switching activities. If the bot stopped tightening bolts on the conveyor and went for new parts, now no one wants to wait for the change of ‘equipment’ in its head,” added the expert.

Key features of Gemini Robotics On-Device:

• versatility and agility of Gemini Robotics with the ability to operate locally on the device;
• execution of a wide range of complex tasks using both hands;
• acquisition of new skills after 50-100 demonstrations.

An additional SDK has been launched to assist developers in customizing the model for their own applications, including testing it in the MuJoCo physical simulator. They can use natural language prompts.

From humanoids to industrial bi-arm robots, the model supports multiple embodiments, even though it was pre-trained on ALOHA — while following instructions from humans. ?

These tasks may seem easy for us but require fine motor skills, precise manipulation and more. ↓ pic.twitter.com/GhBkCj4juZ

— Google DeepMind (@GoogleDeepMind) June 24, 2025

Many companies continue to actively develop the sector. In June, it was revealed that Amazon is working on AI software to enable humanoid robots to deliver packages in Rivian electric vans.

In March, Nvidia introduced a motion simulation engine for robots.

Earlier, 21 humanoid robots participated in the Beijing half marathon.