In Robot Utility Models, we collected a large number of demonstration videos — 1000 examples in 40 environments — using a tool called “the Stick." The lead author even demonstrated it live at CoRL
on completely new furniture. This one really does 100% actually work - I’ve used it in my house.
Again, key experiment: compare only a handful of environments with lots of data, vs many many environments with only 25 examples. More environments is always better!
I'll leave my final lessons for the substack post, but in general:
- robots data DOES scale
- but we need more diverse data than you usually get in a lab setting
- we need ways to collect data in lots of different envs, more portable robots, etc
Comments
on completely new furniture. This one really does 100% actually work - I’ve used it in my house.
- robots data DOES scale
- but we need more diverse data than you usually get in a lab setting
- we need ways to collect data in lots of different envs, more portable robots, etc
https://itcanthink.substack.com/p/what-are-the-data-scaling-laws-for