Rubik's Cube Regrets: OpenAI's Robotics Lead Speaks

Cool interview with OpenAI’s robotics lead @npew:

Highlights:

They chose the hardest task (manipulate Rubic’s cube) to convince the community but in retrospect they believe they should have started simple and added complexity. 1/n https://open.spotify.com/episode/2fZiMqT47SQyjKtsxMmY5j?si=bee8b1e74d9c4bcf
They thought they can accomplish this in ~6 months. They managed to do it in simulation in just 2-3 weeks. But transferring this to robotic hand took 2 years!
When doing sim2real, it turned out that learned model had no respect for delicateness of robotic hand!
Robotic hand will break all the time, thumb and fingers dangling down. Shadow folks said many labs returned their hardware back in pristine condition because no one wanted to take risk.
The sweet spot for the team size to accomplish this sort of significant research milestone is between 10 to 20. Less than 10 people probably won’t be enough for robotics projects and more than 20 has too much of an overhead.
Everyone in the team was fixing “bugs” all the time. There are of course different levels of “bugs”. They had huge backlog of things to do. When people get bored, they jump from one task to another. Generalists are important for such teams!