fix: sharpa scenes hard-reset from upstream private-robotsβ
Chaoyi Pan
9e0451a
The sharpa right.xml/left.xml on the dataset side had drifted from the
upstream jc-bao/private-robots: it still carried experimental
``quat="0 1 1 0"`` on the right_palm/left_palm sites from earlier
debugging, even though I had already updated the dataset-side xml
back to ``euler="0 0 0"`` once. The drift survived because:
1. ``processed/oakink/assets/robots/sharpa/`` had become a real
directory with its own copy of the xml (overlaying the symlink
used during initial bring-up).
2. ``generate_xml.py``'s mj_spec.compile()/to_xml() faithfully
wrote the rotated palm site into every produced scene.xml, which
then stuck around even after I re-pointed the source.
Hard reset: ``rsync --delete`` the upstream private-robots/sharpa
contents over the dataset path, wipe ``processed/oakinkv2/sharpa/``,
re-run the full pipeline (decompose β generate_xml β ik_fast β
mjwp_fast) with the clean source.
After this reset the compiled scene.xml correctly emits the palm site
with nβ¦
fix: sharpa scenes regenerated from original palm site (no stale quat)β
Chaoyi Pan
db32d7a
Earlier commits (7283a8a, 72f464b) accidentally baked a non-identity
``right_palm`` quat (`quat="0 0.707107 0.707107 0"`) into the compiled
sharpa scene.xml from a temporary experimental palm-site change. The
source ``right.xml`` was reverted but the dataset still shipped the
old compiled scene, causing the IK to align the hand mesh ~90Β° off the
correct grasp orientation (palm facing wrong way for pour_tube,
stir_beaker, etc.).
Delete and regenerate every sharpa scene.xml/scene_eq.xml from the
clean upstream private-robots/sharpa source (where right_palm has
``euler="0 0 0"`` only, no quat). Re-run IK + MJWP-Fast for all 5
right-hand tasks.
Grasp success (mjwp_fast):
sharpa pick_spoon_bowl β, pour_tube β, uncap_alcohol_burner β,
wipe_board β β 4/5 = 100% (stir_beaker still hits mjwp
optimizer rotation tracking limits, unrelated to palm quat).
xhand: unchanged, 5/5 = 100%.
update: re-run xhand+sharpa pipelines with warm-start IK fix on remote L40Sβ
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 β, pour_tube β, stir_beaker β,
uncap_alcohol_burner β, wipe_board β β 5/5 = 100%
sharpa: pick_spoon_bowl β, pour_tube β, stir_beaker β,
uncap_alcohol_burner β, wipe_board β β 4/5 = 80%
stir_beaker mjwp optimizer cannot find a feasible grasp under
the heavy-contact stirring motion (pos_errβ0.11m, rot_errβ0.5).
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.
add: sharpa retarget output for oakinkv2 (5 right-hand tasks)β
Chaoyi Pan
8f34a8d
The earlier oakinkv2 commit (1c839de) was supposed to include the sharpa
results but the .gitignore rule ``sharpa/`` (intended for ignoring the
private-robots symlink) silently excluded everything under
processed/oakinkv2/sharpa/. Add an explicit gitignore exception and
push the missing scene XMLs, trajectory NPZs, IK + MJWP-Fast videos and
viser exports for: pick_spoon_bowl, pour_tube, stir_beaker,
uncap_alcohol_burner, wipe_board.