Traditionally a robot was programmed narrowly for one task. Now large AI models trained on broad data about language, images and actions are being used to give robots more general understanding, so they can interpret instructions and handle objects and situations they were not explicitly programmed for.
This could make robots far more flexible, adapting to new tasks without bespoke programming for each. The physical world is unforgiving, and a mistake has real consequences, so these systems are tested carefully and often kept under supervision. The approach is early but promising, bringing the adaptability of modern AI into machines that act in the real world.