Geometry and coordinate frames
The fingertip exterior is 32.70 × 22.41 × 19.00 mm; its rear profile is
22.40 × 19.00 mm. OBJ positions, mounting translations, and simulator
positions are stored in metres. Quaternions use (w, x, y, z).
Registration and mounting
The original layout aligns to the overmold interior by one translation of
approximately (-2.775670, -0.0000166, -1.429237) mm in the sensor frame, with
no scaling, additional rotation or individual taxel movement. Reference-CAD
comparison gave 1.37 µm RMS and 4.29 µm maximum residual. Contact poses were then
projected along the original normals through approximately 1.43 mm of overmold.
assets/sensors/dex01/reference.json records this shift and the mounting offset.
Offset quaternions use (w,x,y,z); translations and OBJ coordinates use metres.
Asset sources
The exterior OBJ and reference.json live under assets/sensors/dex01/.
The mounted sensor and Tesollo hand USD assets live under assets/robots/tesollo/.
Taxel NPZ files ship inside the Python package. Fetch them with git lfs pull.
The geometry builder and verification scripts are the reference workflow for changes; see attachment and validation.
Transform a contact sample
An NPZ pose stores its three local axes in columns 0–2 and position in column 3.
Column 2 is the outward normal. First apply the mount rotation and translation
from reference.json to obtain coordinates in the sensor rigid body. Then
apply that body’s live world pose from PhysX. For a point p:
p_body = R_mount p_pattern + t_mount
p_world = R_body p_body + t_body
Normals receive the rotations without translation. These are fixed mount transforms; the sensor’s live body pose changes as the hand articulates. The capture utility verifies world-space sample alignment against every mounted visual mesh.
Use the exterior pattern supplied by Dex01PatternCfg for contact sampling.
The reference dex01_pattern.npz retains the original layout for traceability; it is not the
projected exterior sampling surface. Both files keep the same taxel order and
firmware pixel mapping.