Google DeepMind has unveiled Gemini Robotics On-Device, a breakthrough AI model designed to run directly on robotic hardware. This shift moves robotics AI from the cloud to the edge, enabling robots to operate independently even in environments with limited or no internet access major step forward for real-world applications.
The new Gemini model is crafted to be both compact and efficient, making it ideal for integration with a wide range of robotic platforms. It can learn new tasks quickly, often requiring just 50 to 100 demonstrations to master a skill. Google is rolling out an SDK to select developers, allowing them to fine-tune the model for specific robotic needs.
Originally trained on Google’s ALOHA robot, Gemini Robotics On-Device now supports additional hardware, including the dual-armed Franka FR3 and the Apollo humanoid robot by Apptronik. In testing, the on-device version has shown performance nearly on par with cloud-based models, even when tackling unfamiliar objects or complex assembly tasks4.
According to Carolina Parada, head of robotics at Google DeepMind, the model’s efficiency and autonomy are surprising even to its creators. “It’s small and efficient enough to run directly on a robot,” Parada noted, highlighting its potential for secure, low-latency environments and as a foundation for future robotics innovation4.
Beyond Google, companies like Nvidia and startups such as RLWRLD are also pushing the boundaries of AI-powered robotics. However, Gemini Robotics On-Device stands out for its ability to deliver advanced vision-language-action (VLA) capabilities at the edge, making robots more independent and adaptable than ever..
Early Access and Developer Opportunities
Developers interested in getting hands-on experience can join Google’s Gemini Robotics trusted tester program. Early participants will gain access to the SDK and tools needed to customize the model for a variety of robotic applications, paving the way for smarter, more self-reliant robots shortly.
Be the first to post comment!