A new study from MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) demonstrates a significant advancement in AI-powered robotic manipulation. Researchers have developed a system that allows a robot to perform complex, dexterous tasks—such as assembling a chair or manipulating flexible cables—by learning from a relatively small number of human demonstrations. The system uses a two-stage process: first, a neural network learns a general policy from the demonstrations, and then a large language model (LLM) breaks down new, high-level instructions into a sequence of actionable steps for the robot. This approach helps the robot generalize to new tasks and environments it hasn’t explicitly been trained on, moving beyond rigid, pre-programmed routines. The research highlights progress toward more adaptable and useful robots for home and industrial settings, though challenges in reliability and handling unpredictable real-world scenarios remain. For the full details, read the complete article at https://technologyreview.com/2024/04/05/1090735/mit-robot-learns-from-demos.