one-shot imitation learning

One-shot imitation learning aims for the kind of flexibility people take for granted: watch a single demonstration of a brand-new task and immediately do it yourself. Show the system one example of stacking these particular blocks in this particular order, and it should perform that task on the spot — no retraining, no second demonstration.

The trick is to train, in advance, not a policy for one task but a policy that reads a demonstration as part of its input. Across a large distribution of related tasks, the network learns to attend to the demonstration, extract the intent, and condition its actions on it. At test time a fresh demonstration of an unseen task simply becomes new context, much as a language model follows a new instruction it was never explicitly trained on.

It trades the heavy per-task data of ordinary imitation for heavy up-front meta-training and a strong assumption that the new task resembles the training distribution; truly novel tasks still break it.