# GraphHands A graph-based representation learning pipeline for dexterous hand manipulation, aligning **human hands** (EgoDex) with **robot hands** (orca tendon retargeting) in a shared latent space. ## Why Human and robot hands have different proportions, joints, and actuation, so their poses don't map one-to-one. GraphHands learns the correspondence instead of hand-coding it — turning two incompatible embodiments into one common space you can transfer skills across. ## How - **Graph construction** — DBSCAN clustering + k-NN over canonical pose sequences - **Gromov-Wasserstein alignment** — recover cross-embodiment node correspondences - **node2vec (PecanPy)** — learn embeddings on the merged human/robot graph - **Conditional autoencoder** — reconstruct poses conditioned on embodiment ID - **Posture library** — labelled exemplars (pinch, open-flat, …) per embodiment, stored as raw 63-D canonical pose vectors ## Stack - Python 3.12, PyTorch (conditional CAE) - scikit-learn (DBSCAN, k-NN), POT (optimal transport), NetworkX - Streamlit + Plotly labelling and visualization apps

tech Python, PyTorch, scikit-learn, POT, NetworkX, node2vec