Google DeepMind's 'Open X-Embodiment': A Groundbreaking Step in Advancing Robot Learning

Quincy Cahilig, Tech Times 05 October 2023, 01:10 pm

Google's DeepMind team took a major step in advancing robotics by partnering with 33 research institutes to build a massive shared database known as "Open X-Embodiment." This endeavor is comparable to ImageNet, a well-known repository of over 14 million photos launched in 2009.

Quan Vuong and Pannag Sanketi of DeepMind noted, as per a TechCrunch report, "Just as ImageNet propelled computer vision research, we believe Open X-Embodiment can do the same to advance robotics," in a comparison of the Open X-Embodiment project's potential impact to that of ImageNet. Building a dataset of multiple robot demos is essential to training a model that can manage different robots, read different commands, reason about complicated tasks, and generalize.

The project's founders recognize the scale of the challenge and underline the value of teamwork, emphasizing that it is more than what can be accomplished in a single laboratory. Over 500 skills and 150,000 jobs taken from 22 different kinds of robots are available in Open X-Embodiment. The "Open" label on this priceless dataset indicates that its creators are committed to making it accessible to the larger research community.

Developing Intelligent Robots

The Google DeepMind team hopes open-sourcing data and providing safe but restricted models will eliminate obstacles and boost research. They genuinely think that allowing robots to learn from one another and, more crucially, supporting researchers' collaborative learning, are essential for the advancement of robotics.

AIM reported that DeepMind researchers have successfully trained two new generalist models using the Open X-Embodiment dataset. The first, the RT-1-X, is a transformer model made for robot control. It outperforms models made specifically for these kinds of jobs by having an average success rate of 50% higher.

The second model, RT-2-X, also known as a vision-language-action model, can understand visual and aural input while also absorbing training data from the internet. Despite having the same basic design as their predecessors, RT-1 and RT-2, these programs are notable for outperforming them. The early models were developed using smaller datasets.

Surprisingly, the Open X-Embodiment-trained robots showed the ability to carry out activities for which they had never received formal instruction. The information encoded through a variety of experiences learned from different robot types contributed to the development of these emergent talents. Notably, this phenomenon was more important in tasks requiring strong spatial reasoning.

Introducing 𝗥𝗧-𝗫: a generalist AI model to help advance how robots can learn new skills. 🤖

To train it, we partnered with 33 academic labs across the world to build a new dataset with experiences gained from 22 different robot types.

Find out more: https://t.co/k6tE62gQGP pic.twitter.com/IXTy2g4Lty
— Google DeepMind (@GoogleDeepMind) October 3, 2023

Science Fiction Stuff Turning into Reality

This most recent development comes in the wake of news that a former employee of Google's DeepMind artificial intelligence research division has founded a new venture fund called Mythos Ventures.

Former DeepMind researcher Vishal Maini established Mythos earlier this year and has so far raised $14 million from more than forty investors.

According to a Bloomberg report, Maini joins a growing group of former Google employees who are well-equipped with resources and knowledge and who share a common desire to have an impact on the development of AI technology as business owners and investors.

Maini highlighted the revolutionary potential of AI in fields including space exploration, climate change, longevity, and healthcare in his first interview about the fund. He said that humans will continue to be in charge of this revolutionary technology, referring to AI as the "sci-fi future" humanity has long imagined.