Google’s DeepMind division has introduced a groundbreaking addition to its AI family — Gemini Robotics On-Device, a large language model that runs locally on robotic systems. Unveiled on June 24, the AI model is built for latency-sensitive environments and offline use, making it ideal for robots working in low or no-connectivity conditions.
Built for Dexterity and speed without the Cloud
Unlike conventional AI models that depend heavily on cloud-based computation, Gemini Robotics On-Device is engineered to operate entirely offline. This enables faster response times and enhanced privacy while executing tasks. The model has been optimised for low-compute environments but still retains powerful capabilities like understanding natural language and completing multi-step commands.
Google says that Gemini can perform highly dexterous tasks such as folding clothes, unzipping bags, and assembling parts, all through local processing. Despite its compact design, the model shows near-cloud-level performance in terms of accuracy and versatility.
Works across Robot platforms like Franka FR3 and Apollo
Initially trained on ALOHA robots, Google has extended the model’s compatibility to more advanced systems like the bi-arm Franka FR3 and Apollo humanoid robot. In tests, Gemini Robotics On-Device enabled these robots to follow voice instructions, interact with unseen objects, and even carry out industrial tasks like belt assembly.
Apollo, for instance, could generalise its motion and manipulate new objects in real-world environments — something usually only possible with cloud support.
Developers can access Gemini via the SDK
Google has released a software development kit (SDK) for Gemini Robotics On-Device, giving developers access to experiment with and deploy the model across compatible robotic systems.
While Google is advancing rapidly in the robotics space, other tech giants like NVIDIA (with its Groot N1 model) and Hugging Face are also racing to power robots with generative AI, making this an exciting era for robotic intelligence.
Gemini Robotics On-Device could mark a major shift in how robots operate — fast, flexible, and fully offline.