Google DeepMind unites researchers in bid to create an ImageNet of robot actions
Google's DeepMind robotics division has joined forces with 33 research institutes to create a massive shared database known as Open X-Embodiment. This initiative is likened to ImageNet, a database of over 14 million images established in 2009. The aim is to advance the field of robotics by providing a comprehensive dataset for training robots in various tasks and scenarios.
In the realm of robotics, the ultimate goal is to achieve machine learning that can equip robots with general-purpose capabilities. Currently, most robots are designed to excel at specific tasks. To transition from single-purpose to general-purpose robotics, learning is key. Researchers in the field, as well as startups and corporations, are actively working on solutions to make robot programming more accessible.
While various approaches exist, it's becoming evident that there's no single solution to this complex challenge. Instead, building more sophisticated and versatile robotic systems will likely require a combination of methods. A critical component in most of these approaches is the availability of a substantial shared dataset.
DeepMind's robotics team has taken a significant step by collaborating with research institutions to develop Open X-Embodiment, which is akin to the influential ImageNet in computer vision research. ImageNet consists of a vast collection of labeled images, and its impact on the field has been profound.
Open X-Embodiment goes beyond images and focuses on robot actions. It encompasses over 500 different skills and 150,000 tasks drawn from 22 diverse types of robots. As the name suggests, this dataset is open and accessible to the broader research community, fostering collaboration and innovation.
The DeepMind researchers behind this project emphasize that the scope of this endeavor is too vast for any single laboratory to undertake. They believe that sharing this data and providing safe yet limited models will lower barriers and accelerate research in the field of robotics.
In the words of the DeepMind team, "The future of robotics relies on enabling robots to learn from each other, and most importantly, allowing researchers to learn from one another." This collaborative effort aims to bring us closer to the era of general-purpose robotics, where robots can adapt and perform a wide range of tasks effectively.