Google’s new robotics AI can be executed without the cloud and still tying your shoes

Sometimes we call chatbots such as Gemini and Chatgpt “robots”, but generative AI also plays a growing role in real, physical robots. After Announce twin robotics Earlier this year, Google DeepMind has now unveiled a new VLA model (Vision Language Action) to control robots. Unlike the previous release, there is No cloud componentallowing robots to work with complete autonomy.

Carolina Parada, head of robotics at Google DeepMind, says that this approach of AI -Robotics can make robots more reliable in challenging situations. This is also the first version of the Robotics model of Google that can tailor developers to their specific use.

Robotics is a unique problem for AI because not only the robot exists in the physical world, but also changes its environment. Whether you block it to move blocks or bind your shoes, it is difficult to predict any possibility that a robot could encounter. The traditional approach to training a robot on action with reinforcement was very slow, but generative AI provides much greater generalization.

“It is drawn from the understanding of Gemini’s multimodal world to do a completely new task,” explains Carolina Parada. “What makes that possible is in the same way in which Gemini can produce text, write poetry, an article barrel, you can also write code and you can also generate images. It can also generate robot actions.”

General robots, no cloud needed

In the previous Gemini -Robotics release (still the “best” version of Google’s Robotics Tech), the platforms had a hybrid system with a small model on the robot and a larger one in the cloud. You have probably seen chatbots “think” for measurable seconds while they generate an output, but robots have to respond quickly. If you tell the robot to pick up and move an object, you do not want it to pause while every step is generated. The local model makes rapid adjustment possible, while the server -based model can help with complex reasoning tasks. Google DeepMind now unleashes the local model as an independent custard, and it is surprisingly robust.