One motion model for any sensor placement or device

Wearable motion sensors give different readings depending on where you place them, what device you use, and how it's oriented—making it nearly impossible to train motion models that work across devices. AnyMo solves this by simulating realistic sensor data across hundreds of body placements, then pre-training a graph-based encoder on these synthetic signals paired with language. The result: a single model recognizes human activities zero-shot on 14 new datasets with 11.7% better accuracy, retrieves matching text descriptions for IMU data, and generates captions from sensor readings. This matters because it makes wearable motion sensing practical beyond controlled labs.