update: re-run xhand+sharpa pipelines with warm-start IK fix on remote L40S
Chaoyi Pan
72f464b
After the corresponding spider fix (warm-start IK base 6-DOF to wrist
target), all 5 oakinkv2 right-hand tasks for both xhand and sharpa
robots are re-processed. Grasp success rate (mjwp_fast):
xhand: pick_spoon_bowl ✓, pour_tube ✓, stir_beaker ✓,
uncap_alcohol_burner ✓, wipe_board ✓ — 5/5 = 100%
sharpa: pick_spoon_bowl ✓, pour_tube ✓, stir_beaker ✗,
uncap_alcohol_burner ✓, wipe_board ✓ — 4/5 = 80%
stir_beaker mjwp optimizer cannot find a feasible grasp under
the heavy-contact stirring motion (pos_err≈0.11m, rot_err≈0.5).
The IK output is correct; this is an mjwp tuning / capability
limit, not a wrist orientation issue.
Also revert the experimental sharpa palm site quat changes. The
underlying IK now works with the original palm site definition.
fix: sharpa palm site orientation + IK cost sweep videos
Chaoyi Pan
7283a8a
The right_palm/left_palm sites in the sharpa hand were declared with
identity orientation (euler="0 0 0"), but the IK pipeline targets the
palm site's frame against the MANO wrist frame, which xhand handles by
rotating its palm site with quat="0 1 1 0" (= Rx(180°) * Rz(90°)).
Without that rotation the sharpa hand was facing the wrong way under
the IK target, so the solver hit ring/pinky PIP joint limits (clamped
at 100°) trying to compensate, producing the "stiff straight fingers"
look.
Add the same quat to sharpa right.xml + left.xml palm sites. After the
fix the achievable IK error is the same (~5mm) but ring/pinky PIP
angles fall back into the natural 30-70° grasping range instead of
saturating at 100°.
Also commits a sweep of IK cost configurations
(processed/oakinkv2/sharpa/right/pick_spoon_bowl/0/ik_sweep/) for
side-by-side inspection: combinations of wrist_pos_cost in {0.3, 2.0,
5.0, 10.0}, wrist_ori_cost in {0.3, 3.0, 5.0, 10.0, 20.0}, and
finger_pos_cost in {10.0, 30.0}.
add: sharpa hand assets (xml + meshes) for oakinkv2 retargeting
Chaoyi Pan
433516d
Allows scene.xml under processed/oakinkv2/sharpa/right/.../scene.xml to
resolve its mesh references when other users clone the dataset. Source:
github.com/jc-bao/private-robots/sharpa.
add: oakinkv2 (raw + processed) for picking-stage trajectories
Chaoyi Pan
1c839de
- raw/oakinkv2/: 7 anno_preview pickles + program (primitive segmentation
+ task targets) downloaded from kelvin34501/OakInk-v2.
- processed/oakinkv2/: oakinkv2.py outputs for 5 right-hand tasks
(pick_spoon_bowl, pour_tube, stir_beaker, uncap_alcohol_burner,
wipe_board) under both xhand and sharpa robots, including
trajectory_keypoints.npz, scene XML, IK rollout (.npz / .mp4 / .viser),
and MJWP-Fast retarget output (.npz / .mp4 / .viser).
- xhand: switch to white groundplane to match sharpa for consistent viz.
- .gitattributes: track .viser via LFS.
