Google DeepMind has unveiled a Gemini Robotics on-device Vision Language Action (VLA) model that can run locally on robotic devices. The AI model is built for general purpose tasks and can run without the internet.
This is Google’s first VLA model that has made available for fine-tuning. Developers can sign up for the tester programme and access the software kits.
This new model comes two months after the search giant released its Gemini Robotics model based on Gemini 2.0’s multimodal reasoning and real-world understanding of the physical world.
The flagship model can run both on-device and on the cloud, and is built for bi-pedal robots.
The model can also be customised for different robotic form factors. “While we trained our model only for ALOHA robots, we were able to further adapt it to bi-arm Franka FR3 robot and the Apollo humanoid robot by Apptronik,” the company said in a blog post.
With the bi-arm Franks, the VLA model can perform other tasks like folding clothes or work on industrial belt assembly tasks too.
