h1_2_pdf_hr

PDF-HR for PHUMA H1_2 (27 DoF)

Hierarchical Pose Distance Field (PDF-HR) model trained for the PHUMA H1_2 humanoid robot (27 active DoF).

This model learns a continuous distance field to the feasible pose manifold using a hierarchical joint encoder architecture inspired by Appendix B of the PDF-HR paper.

---

Model Overview

• Robot: PHUMA H1_2
• Active joints: 27 (revolute)
• Feature dimension per joint: 16
• Hierarchical encoder per joint
• Final MLP: 3 layers × 256 neurons
• Activation: ELU
• Output activation: Softplus (guarantees distance ≥ 0)
• Loss: L1 (MAE)
• Scheduler: CosineAnnealingLR
• Epochs: 50
• Batch size: 4096
• Optimizer: Adam (lr = 1e-3)

---

Architecture

Each joint has an independent local encoder:

• Root joints receive only their joint angle.
• Child joints receive their angle + parent feature.
• All joint features are concatenated.
• A final MLP predicts a scalar distance to the pose manifold.

---

Dataset

This model was trained using:

PHUMA H1_2 PDF-HR Dataset

https://huggingface.co/datasets/josabb/phuma_h1_2_pdfhr

Dataset Contents

• positive_poses_all.npy → positive poses (distance = 0)
• near_poses.npy + near_distances.npy
• interp_poses.npy + interp_distances.npy

All arrays are in NumPy .npy format.

Robot specifications:

• Model: H1_2 (PHUMA)
• Degrees of Freedom: 27 active joints
• Joint order: Matches URDF hierarchy and training script definition

---

Author

Josué Abad

Lab

NONHUMAN Robotics – Perú