This article seems to be vague about what 'one shot' means. The closest it gets is 'one video', but one video might contain many examples. In the first case - put the peach into the red bowl - one example is not enough to distinguish this from, for example, put it in the most distant container. If the demonstration contained only one example, then it is unclear to me how strongly, and how specifically, the meta-training needed to pre-condition the system in order for it to pick the 'correct' interpretation of the actions in the demonstration.
yes, it's true there can be overfitting going on here. Yes it's true the article isn't clear about what exactly is expected to be learned to imitate. But i believe it's very similar to how kids learn. Sometimes you get baffled to how kids interpret your actions and what you ask them.
For example, i remember an adult asked a kid to recite the alphabet backwards, and surprisingly to all, he turned his body away from us and recited the alphabet from A to Z instead of what we were expecting (ie Z to A)!!
Anyway, we cannot deny the fact that these experiments are steps toward progress. Great job!
It was a real cringe here watching him drop the peach.
Peaches should never be dropped. They are easily bruised. We shouldn't teach our robots to drop them! Bad robot trainer!
The point is -- while I realize the peach thing may seem trivial -- if our AI trainers can't even get this right, how do we expect them to instill AIs with decent behaviors?
Imitation learning is both imitation and learning. The difference between it and reinforcement learning is that the latter learns a policy for decisions based on rewards and the former learns to approximate observations as ground truth.
Just wondering, what is a simpler (e.g. non-physical) experiment to study one-shot imitation? What type of experiment has been typically used in research? And how far is research in this area?
