So much brilliant work is happening on robot training.
A lot of the cost in robot learning sits in labeling video, step by step and hand by hand. If that cost can be reduced by about 8x at nearly the same quality, much more of the footage teams already have can be used.
And now we also have proof that specialized model harnesses are closing in on general frontier models for narrow video tasks, at a fraction of the cost.
Upload a first-person video of someone doing chores, and we have an API returns each step with timestamps, what each hand is doing.
Perceptron AI's research shows, For labeling robot training video, paying 8x more for the biggest AI model buys you almost no extra accuracy.
A smaller system built only for this job gets the same result for the price of a sandwich per hour of video.
https://x.com/perceptroninc/status/2118018939708182843/video/1