原帖引用 · 原文
Rohan Paul@rohanpaul_ai
Human hands are becoming one of the most useful training signals in robot manipulation.
Maxinsights records people doing real tasks, like work in a home kitchen, and labels every clip with hand pose and language.
Maxinsights collects in 8 countries using recording hardware they build themselves, inside one pipeline that runs from task design to delivered clips.
Owning the hardware is what makes that spread useful. If the same capture setup goes to every location, the recording conditions stay fixed while the environment changes. Environment is the variation you want a model to see.
What we do all day, roughly: take a team's question ("what should the robot learn?") and turn it into clips ready to train on. Task design, our own recording hardware, processing, hand pose and language labels, quality checks. Same setup for all 50+ teams in our pipeline.
Come grill us on any step at CoRL, Mon Nov 9: https://t.co/Yu09S5gkVF
在 X 查看这条帖子 ↗
· UTC+8