In-Context Learning for Robots: Methods and Applications
Project-page vision film

A robot shouldn't need a new training run for every new household task. What if teaching it were as natural as showing it what you mean?

That is the promise of in-context learning. Pretraining builds capability; demonstrations, instructions, and feedback guide it at deployment, with the model's weights fixed.

This survey, In-Context Learning for Robots, maps a path from reusable skills to robots that can be taught in the world.

The vision reaches further: carry what is taught into new situations, compose unfamiliar tasks, and ultimately improve how robots learn. Robots that learn what you need, where you need them.

Paper: https://arxiv.org/abs/2609.36012
Voices: synthetic narration using Doubao TTS 2.0 standard voice presets.