AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild
Researchers have developed AnyMo, a novel framework designed to overcome the setup-dependency challenges in modeling human motion from wearable inertial measurement units (IMUs). The system utilizes physics-grounded simulation to generate synthetic data, enabling a graph encoder to learn representations that are agnostic to sensor placement and device variations. AnyMo tokenizes multi-position IMU data and aligns it with a large language model for enhanced motion understanding, demonstrating significant improvements in zero-shot activity recognition, cross-modal retrieval, and motion captioning. AI
IMPACT Enables more robust and transferable human motion analysis from wearable sensors, potentially improving applications in healthcare, sports, and robotics.