Google's Autonomous Robotics AI Operates Independently from the Cloud

While chatbots like Gemini and ChatGPT often get labeled as "robots," generative AI is increasingly integral to the development of actual, physical robots. Earlier this year, Google DeepMind introduced Gemini Robotics. Now, the company unveils a new on-device vision language action (VLA) model, designed to control robots autonomously, eliminating the reliance on cloud services.

Carolina Parada, head of robotics at Google DeepMind, highlights this innovation as a way to enhance robot reliability in complex scenarios. Notably, this is Google's first robotics model that allows developers to customize it for their specific needs.

AI in robotics presents unique challenges due to robots not only existing in physical spaces but also altering their environments. Predicting every possible situation a robot may face—whether it's moving blocks or tying shoelaces—is difficult. Traditional methods relying on reinforcement learning have been slow; however, generative AI offers broader generalization capabilities.

"It's leveraging Gemini's multimodal understanding to undertake entirely new tasks," says Carolina Parada. "Just like Gemini can produce text, write poetry, summarize articles, write code, and generate images, it can also generate robotic actions."

Advancing Robotics with Full Autonomy

The previous Gemini Robotics model employed a hybrid system, whereby a smaller model operated locally on the robot, while a more extensive version processed in the cloud. Although this still represents Google's pinnacle in robotics technology, it requires a balance of quick local reactions with cloud-based complex reasoning.

With the new standalone VLA model, Google DeepMind offers a robust solution for quick adaptation by unleashing the local model to function independently, underscoring its potential in advancing robotics without cloud dependency.