AlphaNIST lets a robot learn 20 assembly tasks with just a binary reward
The NIST Assembly Task Board #1 packs twenty tasks with parts as small as 4 mm, where a tilted nut can jam or cross-thread. Instead of hand-scripting every maneuver, AlphaNIST trains one policy with a binary success reward across diverse starting states—loose parts, in-hand grasps, and cached partial assemblies. Success-guided sampling focuses practice on states solved about half the time, yielding emergent assembly skills.
Perhaps the answer is to stop telling the robot how we think the task should be done, and let it learn how to get there.