assets: add the stacking-cup tower bundle (10 cups + duck, MJCF + USD)

#42
by Everloom - opened
SomaStacks Inc org

Adds the stacking-cup tower bundle that backs the two new hand_track tasks
(stack_10f_tower, stack_6f_tower) in WUJI_SimBench PR #137.

What is here

assets/local_stacking_tower/ β€” eleven pieces from Meshy multi-image
reconstructions of the real graduated stacking-cup toy, rescaled onto the YCB
065 cups ladder:

  • cup01 (56.5 mm) .. cup10 (102.6 mm) β€” the ten graduated cups
  • duck (49 mm) β€” the topper

Each piece ships a visual OBJ + 4K texture, a cavity-true procedural collision
decomposition (49 convex parts per cup: azimuthal wall wedges on the measured
outer taper, the moulded shoulder ledge the cups seat on, and a cap hull), a
real-mass MJCF, and a textured USD sibling for IsaacSim.

One bundle backs both tasks β€” the six-floor tower uses the upper five cups plus
the duck β€” so there is only ever one asset set to keep in sync.

Validated

  • MuJoCo: cups assemble into a standing tower; per-pair seat depths measured
    across a yaw sweep and under load with measure_stacking_tower_seats.py
    (shipped in this bundle), which is what the tasks' checker bands are
    calibrated from.
  • IsaacSim RTX Real-Time: all pieces render with correct textures, colours
    and size ladder from the included USDs.

Two known limitations, recorded rather than hidden

  • The assembled ten-cup tower is only transiently stable: most towers stand
    a few seconds after the last placement and most have toppled within thirty.
    Deeper seats would be an asset change.
  • The duck perches rather than sockets. Its base has no recess matched to a
    cup top, so it balances on the cup's domed closed end: when it engages it does
    so at a very repeatable depth (0.0612-0.0613 m on cup01, 0.1 mm spread), but
    across a yaw sweep it engages in only about half of orientations. A recess
    sized to cup01's 47.4 mm top would fix it.

Note for whoever regenerates this

The MJCF header comments must not contain a literal -- (an XML comment cannot).
The generator's earlier --src <meshy_dir> regeneration hint made every MJCF
invalid XML: MuJoCo's parser tolerates it, but metasim loads MJCF through lxml,
which refuses, so the tasks could not boot at all. The comments here are
rewritten to avoid it.

Everloom changed pull request status to open
Everloom changed pull request status to merged

Sign up or log in to comment