infosys/OpenHumnoidActuatedFaceData

🤖 Open Humanoid Actuated Face Dataset

Dataset Summary

The Open Humanoid Actuated Face Dataset is designed for researchers working on
facial‑actuation control, robotics, reinforcement learning, and human–computer interaction.

• Origin – collected during a reinforcement‑learning (RL) training loop whose objective was to reproduce human facial expressions.
• Platform – a modified i2Head InMoov humanoid head with a silicone skin.
• Control – 16 actuators driving facial features and eyeballs.
• Pairing – each example contains the raw RGB image and the exact actuator angles that produced it.

---

Dataset Structure

Field	Type	Description
image	Image	RGB capture of the humanoid face (resolution 1280x720).
actuated_angle	struct	16 integer key - values ("0", "1" .. so on)

Actuator Index Reference

Idx	Actuator	Idx	Actuator
00	Cheek – Left	08	Eyelid Upper – Right
01	Cheek – Right	09	Eyelid Lower – Right
02	Eyeball Sideways – Left	10	Forehead – Right
03	Eyeball Up/Down – Left	11	Forehead – Left
04	Eyelid Upper – Left	12	Upper Nose
05	Eyelid Lower – Left	13	Eyebrow – Right
06	Eyeball Up/Down – Right	14	Jaw
07	Eyeball Sideways – Right	15	Eyebrow – Left

---

Actuator Mapping Images

Full‑Face Map	Eye‑Only Map

---

Dataset Statistics

Split	Samples	Size
Train (full)	[135k]	≈ 153 GB
Train (small)	20k	≈ 22 GB

---

Usage Example

from datasets import load_dataset, Image

# load the small subset
ds = load_dataset("infosys/OpenHumnoidDataset", name="small", split="train")
ds = ds.cast_column("image", Image())   # decode image bytes ➜ PIL.Image

img = ds[0]["image"]
angles = ds[0]["actuated_angle"]        # {'0': 90, '1': 20, ...}
img.show()
print(angles)

Tip  For the full corpus use name="full" (may require streaming=True once the dataset grows).

---

Data Collection & RL Setup

A detailed description of the RL pipeline, reward design, and actuator hardware will appear in our upcoming paper (in preparation, 2025). Briefly:

1. Vision module extracts target expression keypoints from live human video.
2. Policy network predicts 16 actuator set‑points.
3. Real‑time reward computes expression similarity + smoothness penalties.
4. Images & angle vectors are logged every N steps, forming this dataset.

---

License

Released under the MIT License – free for commercial and non‑commercial use.

---

Citation

@misc{amirul2025openhumanoidface,
  title   = {Open Humanoid Actuated Face Dataset},
  author  = {Amirul et al.},
  year    = {2025},
  url     = {https://huggingface.co/datasets/infosys/OpenHumnoidActuatedFaceData}
}

---

Contribution

1. Amirul Islam ([email protected])
2. Anant Pande ([email protected])
3. Allahbaksh Asadullah ([email protected]) - Mentor

Acknowledgements

Big thanks to the Infosys and Mohammed Rafee Tarafdar (CTO, Infosys) for providing us resources to work on this research project and thanks to all the community and everyone who helped us making this possible.