AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild

AnyMo is a geometry-aware, setup-agnostic framework for human motion modeling with wearable IMUs.

It combines physics-based IMU simulation, graph pretraining, motion tokenization, and language alignment to learn reusable motion representations.

The method generalizes across sensor placements and datasets, delivering strong gains in recognition, retrieval, and captioning.

It also improves zero-shot performance on unseen HAR datasets, addressing a key weakness of wearable motion sensing.