Google DeepMind has announced a new breakthrough in robotics with the release of Gemini Robotics On-Device, a language model capable of running locally on robots without requiring an internet connection.

The model, unveiled on Tuesday, builds on the original Gemini Robotics model launched in March. Unlike its predecessor, this version allows robots to perform complex tasks such as folding clothes or unzipping bags entirely offline. Developers can control and customize the robot’s behavior using simple natural language commands.

According to Google, the on-device model delivers performance close to its cloud-based counterpart in internal benchmarks. While the company claims it outperforms other on-device models, it did not specify which models it was compared against.

In demonstrations, Google showed the model operating across different platforms, including the ALOHA robot for which it was initially trained, the bi-arm Franka FR3, and the Apollo humanoid robot developed by Apptronik. Impressively, the Franka FR3 was able to complete tasks it had not encountered before, such as assembling components on an industrial conveyor belt.

To support further development, DeepMind is also releasing a Gemini Robotics SDK, enabling engineers to train robots using 50 to 100 task demonstrations within the MuJoCo physics simulation environment.

The move places Google DeepMind in a growing field of companies advancing robotics through AI. Nvidia is developing a platform to power foundation models for humanoid robots, while Hugging Face is building both open-source models and physical robots. In South Korea, RLWRLD—a startup backed by Mirae Asset—is also creating foundational AI systems for robotics.

With Gemini Robotics On-Device, Google DeepMind aims to make robotic intelligence more accessible, adaptive, and independent of cloud infrastructure—a potential leap forward for both industrial automation and personal robotics.