Step2Motion: Locomotion Reconstruction from Pressure Sensing Insoles
Eurographics 2026

Jose Luis Ponton¹, Eduardo Alvarado², Lin Geng Foo², Nuria Pelechano¹, Carlos Andujar¹, Marc Habermann²

¹ Universitat Politècnica de Catalunya (UPC)
² Max Planck Institute for Informatics

Project page | Paper (arXiv) | Video | Data

Step2Motion is a system for reconstructing full-body locomotion from multi-modal insole sensors (pressure + IMU). It enables robust motion capture in unconstrained, real-world environments, overcoming the limitations of traditional mocap suits or optical systems.

• Python: 3.9+
• PyTorch: Follow the official installation guide for your system.

1. Clone this repository.
2. Create and activate a virtual environment:

   python -m venv env
   # On Windows:
   .\env\Scripts\activate
   # On macOS/Linux:
   # source env/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

4. Install PyTorch according to your compute platform.

Unzip the provided preprocessed UnderPressure data into the data/UnderPressure directory.
Outcome: You should have a data/UnderPressure/underpressure_test.pt file.

To process raw BVH files manually (downloaded from the official repository):

python src/process_underpressure.py underpressure data/UnderPressure/

1. Download the Step2Motion dataset.
2. Unzip and place it in data/step2motion/. The structure should be:

   data/
   └── step2motion/
       ├── 00/
       │   ├── clip.bvh
       │   ├── clip.json
       │   └── clip.txt
       ├── ...
   

3. Run the processing script:

   python src/process_step2motion.py step2motion data/step2motion/ --seed 14

The dance dataset is provided in the data/dancing/ directory and is already preprocessed.

Generate a prediction for a specific clip:

python src/test.py models/UnderPressure/ skeletons/UPSkeleton_S4_AMASS.bvh --dataset data/UnderPressure/underpressure_test.pt --clip 0

Output: models/UnderPressure/predictions/underpressure_test_c0_pred.bvh. This file can be viewed in Blender or the Unity Visualizer.

Generate predictions for an entire dataset:

# UnderPressure
python src/test_model.py models/UnderPressure/ data/UnderPressure/underpressure_test.pt skeletons/UPSkeleton_S4_AMASS.bvh --only_test

# Dancing
python src/test_model.py models/dancing/ data/dancing/dance_test.pt skeletons/UPSkeleton_S1_AMASS.bvh --only_test

# Step2Motion
python src/test_model.py models/step2motion/ data/step2motion/step2motion_test.pt skeletons/step2motion.bvh --only_test

Compute metrics and generate distribution plots reported in the paper. Choose the target dataset using the argument: up (UnderPressure), dance, or step2motion.

python src/visualize_metrics.py [up|dance|step2motion]

1. Setup: Install Unity Hub and the Unity Editor (tested on version 2022.3).
2. Open Project: Open Unity/InsoleVisualization/.
3. Load Scene: Open Assets/Scenes/Visualizer.unity.
4. Configuration:
   • Select the GlobalManager_UP GameObject.
   • Locate the referenced UnderPressure ScriptableObject.
   • Update the "ModelsPath" field to the absolute path of your local models directory (e.g., C:/Users/user/Desktop/Step2Motion/models/).
5. Playback:
   • Press Play in the Editor.
   • To change clips, modify the prediction name (e.g., underpressure_test_c0) in the GlobalManager_UP script.
6. Controls:
   • SPACE: Play/Pause
   • R: Restart motion
   • G: Focus on Ground Truth
   • P: Focus on Prediction
   • S: Scene View camera
   • I: Toggle Insole visualization
   • Scroll: Zoom
   • ← / →: Previous/Next frame
   • ↑ / ↓: Increase/Decrease playback speed

To train a new model from scratch:

python src/train.py --config configs/config_underpressure.json

If you use this project in your research, please cite:

@article{2026:ponton:step2motion,
  author = {Ponton, Jose Luis and Alvarado, Eduardo and Foo, Lin Geng and Pelechano, Nuria and Andujar, Carlos and Habermann, Marc},
  title = {Step2Motion: Locomotion Reconstruction from Pressure Sensing Insoles},
  journal = {Computer Graphics Forum},
  booktitle = {Eurographics 2026},
  volume = {45},
  number = {2},
  year = {2026},
  doi = {10.1111/cgf.70405},
}

This code is released under the MIT License.