Google AI for Robots Can Fold Your Clothes Without Internet

As manufacturers around the world explore innovative ways to incorporate robotics and automation into their operations, new technology to bolster robot functionality offers organisations tremendous potential for efficiency, safety and commercial gains.

In this context, Google DeepMind’s release of a new language model called Gemini Robotics On-Device offers the chance to run tasks locally on robots themselves – without needing an internet connection.

A new chapter for offline robotics

This new on-device AI model for robotics can operate without needing to be connected to the internet, which marks a major advancement in autonomous robot control and adaptability.

Building on foundations laid in March 2025, when Google DeepMind released Gemini Robotics to help solve complex problems through multimodal reasoning across text, images, audio and video, the latest version takes things a step further by developing the vision-language-action model (VLA) capabilities.

The latest iteration comes with dexterous capabilities similar to the version released in March, but Google says “it’s small and efficient enough to run directly on a robot”.

Gemini Robotics On-Device can dictate a robot’s movements, while developers can fine-tune and control the model to meet various demands using natural language prompts.

Comparative performance

Google claims that the Gemini Robotics On-Device model performs similarly to the cloud-based Gemini Robotics model in benchmarks. The company also insists it outperforms other on-device modes in general benchmarks, although it didn’t specify these competitor models.

Carolina Parada, Head of Robotics at Google DeepMind, says:

“The Gemini Robotics hybrid model is still more powerful, but we’re actually quite surprised at how strong this on-device model is. I would think about it as a starter model or as a model for applications that just have poor connectivity.”

Designed to help robots complete a wide range of physical tasks, the flagship Gemini Robotics model enables robots to perform these movements even if it hasn’t been specifically trained on them.

It also allows robots to adjust to new situations, comprehend and respond to commands, while also performing tasks that require nuanced motor skills. Demonstrations have shown robots running this local model performing tasks such as:

• Folding clothes
• Unzipping bags
• Pouring liquids
• Tying shoelaces

Simultaneous SDK release

In conjunction with this launch, Google is releasing a software development kit (SDK) so developers can fine-tune and evaluate the on-device model – a first-of-its-kind for one of Google DeepMind’s VLAs.

Google insists that developers can train robots with 50 to 100 demonstrations of tasks to train them using these models on the MuJoCo physics simulator.

The new Gemini Robotics model and its associated SDK will be made available to a group of trusted testers while Google works on minimising safety risks.

Industry context and future implications

The release of Gemini Robotics On-Device represents a major advance for AI-powered robotics. 

As the technology continuously evolves, it could have far-reaching implications for various industries around the world:

• Remote Operations: Offline functionality open up possibilities for limited connectivity robotic applications, such as space exploration, disaster response or when sites lose internet access.
• Manufacturing and Logistics: The model's ability to adapt quickly to new environments and tasks could revolutionise warehouse operations and production lines.
• Healthcare: Local processing of visual data enhances privacy, making the technology more suitable for sensitive environments like hospitals.

Explore the latest edition of Manufacturing Digital and be part of the conversation at our global conference series, Manufacturing LIVE.

Discover all our upcoming events and secure your tickets today.

Manufacturing Digital is a BizClik brand