From MANO to robot interaction

The original, unaligned MANO template (stage 0) is shown once at the center. Each branch has three stages around the table: final morphed hand → supplied final morphed-hand grasp → evaluated final robot grasp. Curved colored lines join identical MANO vertex indices from the central template to each final morphology. Their paths are illustrative Bézier connections, not computed optimal-transport or flow-matching trajectories. Each finger and the palm have a distinct color, assigned from native MANO skinning weights. Each connection has a light-to-rich color gradient from original MANO to the morphed hand. The region colors remain consistent across all three branches. The Correspondences control hides or shows the curves and markers. Line thickness adjusts their width, including in image and video exports.

The three source templates are verified to be identical. The original MANO, all three open morphs, and all grasp meshes use their native metric dimensions with no per-hand display scaling. Complete hand–object pairs are translated together. The three stages of each pair form a compact cluster along the circle, with wider gaps between pairs. The original MANO and all three open morphed hands present their palm surfaces toward the reference camera. An explicit facing sign selects the visible side using a shared surface patch; each open hand is turned 180° about its finger-up axis from the previous orientation. Dex3 adds a 45° turn about its finger-up axis. Open-hand and robot geometry retain native sizes. The apple grasp uses a visualization copy with a local middle/ring-finger surface merge and seam smoothing. Its surface is reconstructed at 0.8 mm resolution; the original GLB and embedded source geometry are retained. Hammer and wineglass grasp surfaces remain original. The object meshes and grasp placements are unchanged. The reference camera and object orientations are preserved, with more space allocated to Allegro and black visual materials for Dex3.

Morph states come from align_trace.zip. Integer states reproduce recorded surfaces; playback interpolates adjacent meshes, starting at initialized stage 1. Original template 0 is selectable. The stage-2 GLBs provide static final grasps. Evaluated robot poses come from benchmark_raw, without joint interpolation. Apple, wineglass, and hammer use source frames 250, 240, and 500. Independent recordings synchronize by normalized duration. Choose Objects fixed or Recorded object paths for robot playback. Drag to orbit and scroll to zoom. Save viewpoint remembers the current camera in this browser and downloads a JSON file; Restore viewpoint recalls it, and Load viewpoint opens an exported JSON. The saved camera is also restored when this viewer reopens. Reference view remains available. Export PNG captures the current figure; Export video retains the full-sequence playback option.