OpenAI Disbands Its Robotics Research Team

OpenAI has disbanded its robotics team after years of research into machines that can learn to perform tasks like solving a Rubik's Cube. Company cofounder Wojciech Zaremba quietly revealed on a podcast hosted by startup Weights & Biases that OpenAI has shifted its focus to other domains, where data is more readily available. From a report: "So it turns out that we can make a gigantic progress whenever we have access to data. And I kept all of our machinery unsupervised, [using] reinforcement learning -- [it] work[s] extremely well. There [are] actually plenty of domains that are very, very rich with data. And ultimately that was holding us back in terms of robotics," Zaremba said. "The decision [to disband the robotics team] was quite hard for me. But I got the realization some time ago that actually, that's for the best from the perspective of the company."

In a statement, an OpenAI spokesperson told VentureBeat: "After advancing the state of the art in reinforcement learning through our Rubik's Cube project and other initiatives, last October we decided not to pursue further robotics research and instead refocus the team on other projects. Because of the rapid progress in AI and its capabilities, we've found that other approaches, such as reinforcement learning with human feedback, lead to faster progress in our reinforcement learning research."

OpenAI needs to change its name

[Junta]

The 'Open' company licensed it's most prominent offering (GPT-3) to Microsoft exclusively and is in no way Open.

They moved from being a nonprofit to be a for-profit company.

They were founded and named in the spirit of advancing AI through an Open approach, now they are just a company milking the word 'Open' in their brand without any meaning whatsoever.

Re:

[Ostracus]

Then use BERT* [wordpress.com] then.

*Near the bottom of the page.

Re:

[jmcbain]

They're "open" because they publish all their work. Here's the publication for GPT-3 [arxiv.org]. Everything you need to implement your own GPT-3 model is in there. The implementation and model weights for GPT-2 are available through HuggingFace Transformers.

In otherwords.

[jellomizer]

While this was fun and all. We are getting a lot of big offers for us to work at different companies. So we can't keep this organization running. Because we have families to feed and want more money.

I am not judging, I would probably have done the same thing. In many was it is better to disband the project, then just allow it to slowly die from lack of effort.

Re:

[Ostracus]

Maybe Boston Dynamics embarrassed them too much with their dancing robots?

Re:

[ceoyoyo]

"Open"AI's raison d'être is to find applications for deep reinforcement learning. Reinforcement learning is great for the very hardest problems, where you have no idea how to solve them. Its drawback is that it's slow as balls. And that's really, really slow if your feedback loop involves the real world. Teaching a "robot" to walk, solve Rubik's cube, or whatever in a simulated environment is pretty standard stuff. Doing it with an actual robot in the real world, well, enjoy watching it flail, crawl, t

Re:In otherwords.

[jythie]

Eh, another way I read it : 'AI is hard work! Why can't we just stick to mountains of data and lots of GPU so we can just keep rolling dice till it happens to work???'

ML is the worst way to do AI, but it just happens to have a REALLY cheap way to do it quickly. As soon as you move away from problems that can be solved by randomizing matrices over and over using lots of data, it forces them to actually start using techniques that require understanding things.

Climbing a tree to get to the moon

[dsanfte]

Indeed, the current focus on reinforcement and data-led ML is a bit like climbing a tree and claiming you've made progress in getting to the moon.

Re:

[jythie]

*nod* at some point, AI researchers are going to have to bite the bullet and start learning GOFAI again... which is probably going to take generations to gear up.

True for children, too

[clawsoon]

"other approaches, such as reinforcement learning with human feedback, lead to faster progress in our reinforcement learning research"

It's interesting that the same is true for the training of natural intelligence, e.g. human children. Leaving a group of children completely unsupervised to do unsupervised reinforcement learning will lead to them doing some learning, for sure, but you get much faster and more useful progress if a human adult is guiding them and responding to them.