Skip to main content

Geometric Fabrics PCA

Goal​

Control a 21-DOF humanoid arm+hand using geometric fabrics — a control framework that's fast and stable without needing to solve an optimization problem every step.

The fingers have 15 joints, too many to command directly. So we use PCA to compress them to 7 numbers. PCA is trained on real human hand poses and learns the 7 directions that capture most of the variation in finger motion. You give it a 7-number command and it maps back to 15 joint angles via q_hand = W @ z + μ — just a change of coordinates. z is your 7D input, W maps it to joint space, and μ is the mean pose so that z = 0 gives you a natural resting hand.

Files​

FilePurpose
compute_pca.pyTrains PCA from the HUST hand motion dataset, saves matrix/mean/bounds
humanoid_hand_pose_fabric.pyCore controller — builds all the fabric layers (arm, hand, collision, joint limits)
run_example.pyRuns the simulation loop at 60 Hz, renders optionally
humanoid_hand_params.yamlAll the tunable params — gains, collision spheres, joint limits
pca_matrix.npy, pca_mean.npy, pca_score_*.npyTrained PCA files loaded at runtime
FABRICS/NVIDIA geometric fabrics framework (submodule)

How It Works​

Every step, four layers each compute forces and sum them together:

  1. Arm attractor — holds the arm at a fixed default config (arm doesn't move)
  2. Finger attractor — takes the 7D PCA command, maps to 15 finger joints, pulls toward them
  3. Body repulsion — 21 spheres on the robot links push away from obstacles
  4. Joint limit repulsion — pushes back when joints get near their limits

All of that sums into qdd = -M⁻¹ f, which gets integrated to get the next joint state.