What does it actually mean to "train" a robot? @LeoKharon discussed this on @roboticsstack Most people picture reinforcement learning or world models. The reality on a factory floor today is actually behind what many people think. The dominant method is still manual programming: → A teach pendant device "allows an operator to manually jog the robot into positions" and record waypoints → The robot then "basically just repeats the path" → "This works extremely well if everything stays the same" It works beautifully until something moves. New object, new task, even a lighting change, and a human expert comes back to reprogram the whole thing from near scratch. Before any robot "learns" anything, the question is whether it can handle a world that refuses to stay still. Walmart found this out at scale. It spent years deploying Bossa Nova's shelf-scanning robots across hundreds of stores, then pulled the program in 2020. The robots worked. The stores around them kept changing faster than the systems could keep up. Source: @roboticsstack