University of Maryland researchers have developed μ₀ (Mu Zero), an AI world model that learns reusable motion patterns from human and robot videos.

Given an image and an instruction, it predicts 3D movement traces for hands, tools and objects, which can then be translated into commands for a specific robot. In physical robot-arm tests across three tasks, the approach reached an average success rate of 91.7%, pointing to a way to reduce reliance on costly hardware-specific training data.