Google and University of California, Berkeley, in collaboration with 32 other robotics laboratories, embarked on the RT-X project to create general-purpose robots.

The project aimed to train a single deep neural network that can control diverse types of robots, a concept known as cross-embodiment.

The project gathered data from nearly a million robot trials across 22 types of robots, forming the largest open-source dataset of real robotic actions.

Data from different robots could be used with simple machine-learning methods, enhancing performance compared to individual control systems.

The project incorporated Internet-scale image and text data to boost reasoning capabilities.

The fusion of multirobot data and Internet-sourced knowledge improved the robot's capacity to comprehend and perform complex tasks.

The project seeks to inspire more researchers to contribute to the dataset and develop data standards, reusable models, and new techniques.

The ultimate aim is to develop a single neural network capable of controlling various robots and managing different real-world tasks.