Google DeepMind has launched a new version of its Gemini Robotics model called Gemini Robotics On Device, which allows robots to run tasks without needing an internet connection. This marks an important step in making robots more independent, especially in situations where connectivity is unreliable. Unlike traditional systems that depend heavily on cloud processing, this model enables local decision-making, giving developers the flexibility to control robotic actions using simple language instructions.
The new on-device model builds on the earlier Gemini Robotics system released in March. Although it runs locally, Google claims its performance is nearly as strong as the cloud-based version. In demonstration videos, robots using the model completed practical tasks such as unzipping bags and folding clothes. Google also highlighted the model’s adaptability, noting that while it was originally trained on ALOHA robots, it successfully transitioned to other systems like the Franka FR3 and the Apollo humanoid robot. The Franka robot reportedly managed to perform assembly tasks on an industrial belt, even when faced with unfamiliar objects.
To support developers, Google has also introduced a new software development kit, the Gemini Robotics SDK. With this, developers can train robots by simply showing them 50 to 100 demonstrations of a task using MuJoCo, Google’s physics-based simulator. This makes it easier for teams to customise robotic behaviour without requiring large amounts of specialised data or computing power. It reflects a broader trend in robotics toward smaller, smarter, and more flexible AI tools.
Google is not alone in this race. Other companies are also exploring the potential of robotics powered by foundation models. Nvidia is working on its platform to develop general-purpose models for humanoid robots. Hugging Face has started building open-source models and datasets aimed at robotic development and is also designing hardware. In South Korea, RLWRLD, backed by Mirae Asset, is working on models designed specifically for robotic control.
This growing interest from major AI players signals that robotics may be the next big space for breakthroughs in artificial intelligence. With models like Gemini Robotics On Device, robots are becoming more capable of learning, adapting, and acting independently, bringing science fiction closer to reality in everyday environments.
