Why we re-generate a single recorded action into many worlds — and label every one.
Actions1 min read
One demonstration is not enough
A robot that learns to pick up a mug in one kitchen has learned one kitchen. Move the light, swap the counter, add clutter, and the policy falls apart.
Collecting every variation by hand doesn't scale. So we record the action once and multiply the world around it:
- new rooms and materials
- light, weather and time of day
- clutter and edge cases
- new embodiments — from human hands to humanoid robots
Same motion, new world
Every variation keeps the original hand and timing, so labels line up frame by frame across all of them. Segmentation, depth, hand pose and action timelines transfer for free.
One capture in. Endless realities out.
This is a starter post — edit or delete it from the editor.