Few-shot learning lets a model pick up a new task or object from just a handful of examples — the ability robots need to adapt on the fly to new items and situations without collecting and training on thousands of samples.
Few-shot learning is a model learning something new from just a few examples — like recognizing a new object after seeing it once or twice. Robots need this to handle new items and tasks without huge new datasets each time.
🎯 Quick challenge
Few-shot learning is valuable for robots because it…
A person shown a new tool once can usually use it again. Most machine learning needs thousands of examples to learn something new — hopeless for a robot meeting an unfamiliar object mid-task. Few-shot learning aims to close that gap: learn from just a handful of examples.
What it means
Few-shot learning enables a model to generalize to a new task or category from very few labeled examples — sometimes just one ("one-shot") or a handful. Instead of training a fresh model on a big dataset for each new object or task, the model adapts quickly using what it already knows plus a few new samples.
Learn new from a handful
Leaning on broad pretrained knowledge, the model needs only a few examples to pick up something new — rather than a full dataset and retraining.
How it's achieved
Leverage strong pretrained representations. A model pretrained on huge data (self-supervised, or a foundation model) has already learned general features, so a new category is just a small adjustment on top — a form of transfer learning. This is why large pretrained models are naturally good few-shot learners.
Meta-learning ("learning to learn"). Explicitly train the model across many tasks so it becomes good at adapting quickly to new ones from few examples (methods like MAML, prototypical networks).
In-context learning. Large models can adapt from examples given at inference time, with no weight updates — "here are two examples, now do the third."
Why robots need it
Novel objects and tasks. A home or warehouse robot constantly meets items it's never seen. Few-shot ability lets it recognize or manipulate a new object from a couple of examples instead of demanding a fresh dataset.
Fast on-site adaptation. Deploying to a new environment or a new part on a line without a lengthy data-collection-and-retraining cycle.
Learning skills from few demos. Few-shot imitation learning — teaching a robot a new task from just one or a few demonstrations — is a major goal, made increasingly possible by large pretrained policies.
The challenge
Generalizing from tiny data is fundamentally hard — with few examples it's easy to overfit or misjudge. Success depends heavily on the quality of prior knowledge: the richer and more general the pretrained representation, the better the few-shot performance. This is a big reason the field has moved toward large, broadly-pretrained models.
Why it matters
Few-shot learning is what will let robots be adaptable rather than brittle — handling the endless novelty of the real world without a data-collection project for every new object or task. Together with strong pretrained representations, it's central to the vision of general-purpose robots that learn new things quickly, like people do.
