Robot learns to fold towels by learning what progress looks like, not what the end goal is

Researchers built a system that teaches robots long-term tasks by having humans label whether each step moved forward, backward, or stayed stuck — instead of requiring humans to label every single correct action. This cuts the human annotation work dramatically while letting robots learn from messy, incomplete demonstrations.

For years, teaching robots to do complex multi-step tasks required either massive amounts of human labeling (expensive and fragile) or perfect demonstrations (rare in practice). This approach sidesteps both problems by asking humans to judge only the direction of progress, which is cognitively easy and consistent across different people. The practical shift: robots can now learn from real, imperfect data people actually have, not pristine synthetic data someone has to build.