Researchers including Zipeng Fu and Chelsea Finn have built ALOHA Lightning, a system that teaches a robot fast and precise manipulation, according to a paper on arXiv. Robots that learn from human demonstrations usually work much more slowly than people.
Operators guide a two-armed robot by hand to record demonstrations at close to human speed. The learning pipeline smooths the robot's actions while it works and masks the camera images so that what the robot sees at work matches its training data.
With 50 demonstrations per task, ALOHA Lightning tossed and caught balls, folded shorts, inserted batteries and cleared tables on its own with success rates above 80%, close to human speed.



