Researchers introduced CrowdMAPS, a web‑based system that shows robot states to online workers through a 3‑D interface. Workers add valid alternative actions without touching the robot. The method adds a “width” dimension to robot data by increasing actions per state instead of collecting more trajectories. Tests on five real‑world manipulation tasks showed that expanding from one to ten actions per state lifted average success rates by 35.6 percentage points for keypose policies and 33.0 points for trajectory policies.
Why it matters
Developers can train more reliable robots with fewer physical demonstrations, saving time and cost.
