Gemini Robotics On-Device: Google’s AI model for offline robotic tasks
Gemini Robotics On-Device brings advanced dexterity and natural language understanding to offline robot systems.
On Tuesday, 24 June, Google’s DeepMind division announced the release of a new large language model named Gemini Robotics On-Device, designed to operate locally on robotic systems.
In a blog post, the company stated that the AI model has been optimised to function efficiently on-device and demonstrates strong general-purpose dexterity and task generalisation capabilities.
The offline model is an advancement of the earlier Gemini Robotics system introduced in March this year. Unlike cloud-based models, this version can operate offline, making it suitable for limited connectivity or critical latency.
Engineered for robots with dual arms, Gemini Robotics On-Device is designed to require minimal computational resources.
It can execute fine motor tasks such as folding garments and unzipping bags. According to Google, the model responds to natural language prompts, enabling more intuitive human-robot interaction.
The company claims the model outperforms comparable on-device alternatives, especially when completing complex, multi-step instructions or handling unfamiliar tasks. Benchmark results indicate that its performance closely approaches that of Google’s cloud-based AI solutions.
Initially developed for ALOHA robots, the on-device model has since been adapted for other systems, including the bi-arm Franka FR3 robot and the Apollo humanoid.
On the Franka FR3, the model followed diverse instructions and managed unfamiliar objects and environments, including industrial tasks like belt assembly. The system demonstrated general object manipulation in previously unseen contexts on the Apollo humanoid.
Developers interested in trialling Gemini Robotics On-Device can access it via the provided software development kit (SDK).
Google joins other major players exploring AI for robotics. At GTC 2025, NVIDIA introduced Groot N1, an AI system for humanoid robots, while Hugging Face is currently developing its own open-source, AI-powered robotic platform.
Would you like to learn more about AI, tech, and digital diplomacy? If so, ask our Diplo chatbot!
