Few-shot

Learning or performing a new task correctly from just a handful to a few dozen examples.

Few-shot describes handling a new task with very few examples. In large language models, it usually means writing a handful of examples directly into the prompt and getting the model to follow them without updating any parameters; the 2020 GPT-3 paper, “Language Models are Few-Shot Learners,” made this usage widely known, and the capability is also called in-context learning. In robotics, few-shot more often means collecting just a handful to a few dozen demonstrations — recordings of a human teleoperating the robot through a task — and fine-tuning a pretrained model to learn the new task from them. Real-robot data collection is slow and expensive, so how many demonstrations a method needs to learn a new task is an important measure of how practical it is. It contrasts with zero-shot, where no examples at all are given.

ExampleRVT-2 learns manipulation tasks that require high precision using only 10 demonstrations on a real robot.

Also called

Few-shot Learning

Related

Zero-shot · In-Context Learning · Fine-tuning · Demonstration Data · Sample Efficiency · One-shot Imitation Learning

Sources

Language Models are Few-Shot Learners (GPT-3, arXiv 2005.14165)
RVT-2: Learning Precise Manipulation from Few Demonstrations (arXiv 2406.08545)