Open X-Embodiment

2023: Project Launch

On October 3, 2023, Google DeepMind, together with partners from 33 research laboratories, presented the Open X-Embodiment^[1] project]. It is the largest open database for robot training.

It is noted that training robotic systems is a complex and resource-intensive process that must take into account many factors: from the tasks performed to the environment of the machine. In this case, often changing just one variable requires starting the procedure from zero. The goal of the Open X-Embodiment project is to create an extensive data repository for training robots of different types.

Various examples of robots from the database

The generated database contains more than 500 skills and approximately 150 thousand tasks that 22 different types of machines are capable of performing. As of early October 2023, it is said to be the most comprehensive robotics dataset of its kind. The base will help in creating a universal model capable of controlling many different robots, following a variety of instructions and performing basic reasoning about complex tasks.

The project participants presented a RT-1-X - a model of a robotic transformer (RT), trained on a new data set. It has been tested in five different research laboratories on five different types of widely used robots. Compared to specialized models developed for each machine individually, the unified model showed on average 50% higher success rates for tasks such as opening doors and interacting with objects. In other words, RT-1-X can achieve higher performance thanks to the much more diverse data on which the model is trained. The study demonstrates the potential to develop next-generation helper robots by training using a variety of information and improved unified models.^[2]

Notes