---
title: AnyMo
emoji: 🏃
colorFrom: red
colorTo: indigo
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
short_description: Wearable-IMU motion captioning & activity recognition
python_version: "3.12"
startup_duration_timeout: 30m
---

# 🏃 AnyMo — motion understanding from wearable IMUs

Interactive demo for **[AnyMo](https://huggingface.co/CRUISEResearchGroup/AnyMo)**
("AnyMo: Geometry-Aware Setup-Agnostic Modeling of Human Motion in the Wild").
AnyMo consumes raw wearable-IMU time series — attached at *any* subset of 23
canonical body segments, in any order — and produces:

- **Motion captions** — free-form natural-language descriptions of the motion.
- **Zero-shot activity recognition** — the closest match among your candidate
  activity labels, via IMU-text embedding similarity.

## How to use

1. **Pick a bundled sample** — real 5-second IMU windows (Head + Left Wrist +
   Right Wrist, 60 Hz) from the AnyMo-Bench test split — **or upload your own
   `.npy`** array of shape `[T, S, 6]` (or flattened `[T, S×6]`). Channels are
   `[acc_x, acc_y, acc_z, gyro_x, gyro_y, gyro_z]` in m/s² and rad/s.
2. **Set each sensor's body location** (Head, Left Wrist, Right Wrist for the
   bundled samples).
3. Press **Analyze motion**.

The window is resampled to the model's native 60 Hz and passed through AnyMo's
geometry-aware ST-GCN encoder, product-quantized motion tokenizer, and
Qwen2.5-0.5B language backbone.

## Links

- 📄 [Paper](https://arxiv.org/abs/2605.22715)
- 💻 [Code](https://github.com/Breezelled/AnyMo)
- 🤗 [Model](https://huggingface.co/CRUISEResearchGroup/AnyMo) (Apache-2.0)
- 📊 [AnyMo-Bench](https://huggingface.co/datasets/CRUISEResearchGroup/AnyMo-Bench)

## Sample data attribution

The six bundled example windows (`samples/*.npy`) are drawn from the
**[AnyMo-Bench](https://huggingface.co/datasets/CRUISEResearchGroup/AnyMo-Bench)**
test split ([CC BY-NC 4.0](https://creativecommons.org/licenses/by-nc/4.0/)).
They are redistributed here for demonstration purposes with attribution;
non-commercial use only.