Zero-shot

A model handling a new task or environment directly, with no training examples for it at all.

Zero-shot originated as a machine-learning concept: at test time, a model must recognize a category it never saw during training, transferring learned knowledge with the help of side information such as attributes or text descriptions; a 2009 NeurIPS paper by Mark Palatucci and colleagues introduced the term “zero-shot learning.” In the era of large models and embodied AI, the meaning has broadened to “no new data collection and no fine-tuning for a new task, object, or scene — the model is used as is.” A generalist policy opening a drawer in a kitchen it has never seen, for example, is called zero-shot generalization. It contrasts with few-shot, where a handful of examples are given, and is commonly used to measure a robot foundation model's generalization ability. When reading papers, note that some claims of “zero-shot” only mean the scene was unseen, while the type of task was actually seen during training.

ExampleRobot Utility Models trains one policy per task for five task types, such as opening cabinets and opening drawers, then deploys them directly to unseen new environments with no further data collection or fine-tuning, reporting an average success rate of about 90%.

Also called

Zero-shot Learning, Zero-shot Generalization

Related

Few-shot · Generalization · Out-of-Distribution · Open-vocabulary · Fine-tuning · Generalist Policy

Sources

Zero-shot learning - Wikipedia
Robot Utility Models: General Policies for Zero-Shot Deployment in New Environments