The system must infer the demonstrated goal, important objects and action sequence from one example, then adapt that information to a new situation. A broad pretrained foundation model can supply prior knowledge that makes this limited-data learning possible.
One-shot capability does not guarantee reliable mastery. Variations in viewpoint, object shape, timing and physical contact can expose uncertainty, so robots need safe trial limits and evaluation across more than the original demonstration.
ELI5
One-shot imitation learning lets an AI try a task after watching just one example. It uses knowledge learned earlier to understand what the short demonstration is trying to accomplish.
For example, a person can show a robot one brief sequence for placing an object into a container. The robot may attempt it immediately, but people still need to test different objects and positions because one example cannot show every difficulty.
