In a move that could redefine the future of robotics, Google DeepMind on June 24 announced a new AI model called Gemini Robotics On-Device — a powerful yet lightweight system that can run entirely offline, directly on robotic hardware.
The model builds on Google’s earlier Gemini Robotics platform, adding a crucial new capability: real-time task execution without needing an active internet connection. That’s a big deal for environments where speed, safety, and autonomy matter — think industrial floors, hospitals, or even space missions.
A smarter, self-reliant robot
At its core, Gemini Robotics On-Device is a stripped-down but still remarkably capable version of Google’s flagship AI system. Like ChatGPT, it understands natural language instructions. But unlike cloud-tethered models, it performs all reasoning and control locally — enabling faster response times and complete operational independence.
It’s designed specifically for bi-arm robots like the Franka FR3 and humanoid robots such as Apollo. And it’s already been seen executing dexterous tasks with precision — folding dresses, zipping bags, assembling belts — while dynamically adapting to unseen objects and complex new scenarios.
Efficiency without compromise
What makes Gemini On-Device stand out is its ability to deliver top-tier performance without relying on heavy computational resources. Google claims the model can match — and even surpass — other on-device alternatives in following complex, multi-step instructions. In internal testing, it also came surprisingly close to the accuracy of its cloud-based sibling.
The key lies in how the model generalizes across tasks and physical forms. Originally trained using ALOHA robots, it was later successfully adapted to different hardware platforms with as few as 50 to 100 demonstrations — a promising signal for scalable deployment across the robotics ecosystem.
Built for developers — and the real world
To help developers test and tailor the system, Google is releasing a dedicated Gemini Robotics SDK. Through this software kit and its trusted tester program, engineers can explore everything from simulation-based training to real-world applications, adapting the model to custom tasks and environments.
Because it operates entirely offline, Gemini Robotics On-Device is ideal for latency-sensitive scenarios or places with unreliable network coverage — including disaster zones, remote labs, or unstructured environments where full connectivity just isn’t an option.
What this means
This isn’t just a technical milestone — it’s a glimpse into a future where robots are no longer passive tools but adaptable, responsive agents that understand us and the world around them. Google’s bet is clear: to make intelligent autonomy accessible, reliable, and safe — even without the cloud.
And with rival AI models from NVIDIA and Hugging Face also racing to empower robots, one thing is certain: the battle for the next generation of AI-powered machines has officially gone offline.
