Short, unpolished, dated: thinking out loud with the safety guards off.
The ACT policy I trained succeeds about 3 times out of 10. By any production standard that's a failing system. It's also a machine on my desk that learned a physical task from watching me do it 54 times, and I'm still not over it.
The robot demos you see online are highlight reels: nobody posts the seven misses. I want the real number written down while it's still 30%, so when it gets better, the climb is on the record too.
I expected imitation learning to need thousands of demonstrations. It needed 54. That's how many episodes I recorded (an episode is one complete run of the task, captured as data) before the ACT policy could pick up the sock on its own.
The real lesson isn't the count, though. It's where the failures live: almost all of them happen in spots on the table my demonstrations barely covered: far left, far right. The model doesn't need more data nearly as much as it needs more varied data. Coverage beats volume.