Humanoid robots need to estimate how they are moving, but models trained only on robot recordings see just motions the robot has already performed. Human movement recordings cover more kinds of motion, but people and robots have different bodies and sensors.
The system maps human and robot motion into matching inputs at corresponding body locations, using inertial measurement units (IMUs), sensors that measure movement and rotation. It trains one model to estimate speed and match movement signals with written descriptions, while varying sensor readings and timing to better handle real-world differences.
Without training on physical robot recordings or adjusting the model for the robot, it reaches 0.099 m/s error in estimating speed on a physical humanoid. It also retrieves descriptions for motions not seen in robot training, showing that human movement data can broaden both motion estimation and language-based interpretation.