cfahlgren1/simready-halloween-candy

SimReady Halloween Candy and Porch

Fun-size Halloween candy, Halloween props and a front-porch digital twin as simulation-ready OpenUSD assets, laid out like nvidia/simready-assets. Made for teaching an SO-101 robot arm to take candy out of a bowl and drop it on top of a sheet ghost for trick-or-treaters, in simulation first.

The wrappers are photorealistic look-alikes of the candy in a typical bulk Halloween bag, with the same colors, layout and plastic finish but no brand names or logos. They were made with GPT-6 image generation. Each candy has a wrapper variant set, so one USD file can wear any of its wrappers.

Assets

Asset	Type	Size (cm)	Mass (g)	Mass source	Colliders	Wrappers
food_candy_bubbleGumTwist_a01	Twist-wrapped bubble gum	4.2 × 1.9 × 1.9	4.5	about 4.5 g per piece	capsule_x	2
food_candy_fruitChew_a01	Wrapped fruit chew	2.1 × 2.1 × 1.3	5.0	about 5 g per piece	box	3
food_candy_fruitSnackPouch_a01	Fruit snack pouch	9.0 × 7.0 × 1.5	22.7	0.8 oz (22.7 g) pouch	box	3
food_candy_funSizeChocolate_a01	Fun-size chocolate bar	7.0 × 3.0 × 1.1	17.0	0.6 oz (17 g) fun-size bar	box	2
food_candy_funSizePouch_a01	Fun-size candy pouch	7.5 × 5.5 × 1.4	15.6	0.55 oz (15.6 g) fun-size bag	box	4
food_candy_licoriceTwist_a01	Licorice twist	11.0 × 1.3 × 1.3	11.3	about 11 g per twist (4 twists = 45 g)	capsule_x	2
food_candy_lollipop_a01	Lollipop	7.7 × 2.5 × 2.5	6.0	about 6 g per lollipop	sphere, capsule_x	3
food_candy_powderPacket_a01	Sour powder packet	8.0 × 6.0 × 0.5	10.0	NET WT 0.35 oz (10 g), from the packet label	box	3
food_candy_rootBeerBarrel_a01	Root beer barrel hard candy	4.0 × 2.2 × 2.1	6.0	about 6 g per piece (3 pieces = 18 g)	capsule_x	2
food_candy_sourStrawsPack_a01	Sour candy straws pack	15.0 × 3.5 × 1.0	15.0	about 15 g fun-size pack	box	2
food_candy_tabletRoll_a01	Candy tablet roll	5.0 × 1.8 × 1.8	7.1	0.25 oz (7.1 g) roll	capsule_x	2
food_candy_taffyBar_a01	Chewy taffy bar	9.0 × 2.4 × 1.0	15.6	0.55 oz (15.6 g) bar	box	3
food_candy_taffySquare_a01	Mini taffy square	4.5 × 4.0 × 0.9	9.8	0.35 oz (9.8 g) mini bar	box	3
food_candy_twistWrap_a01	Twist-wrapped hard candy	4.8 × 3.0 × 0.9	5.7	3 pieces per 17 g serving	box	3

Masses are typical nutrition-label net weights per piece.

Props and porch

Asset	Description	Size (cm)	Mass (kg)	Colliders
halloween_bowl_skeletonOrange_a01	Orange dancing-skeleton candy bowl	30.0 × 30.0 × 12.0	0.35	mesh
halloween_decor_ghostSheet_a01	Porch sheet ghost with candy dish top	123.9 × 100.5 × 77.1	1.5	cylinder_z, mesh
halloween_env_porchCorner_a01	Front porch corner with cobwebbed door	320.0 × 266.0 × 262.0	500.0	box, box, box
plant_mum_orange_a01	Orange garden mums in a white planter	41.2 × 38.6 × 52.0	4.0	cylinder_z, sphere
plant_houseplant_pot_a01	Small houseplant in a pale blue planter	20.9 × 20.0 × 31.7	1.5	cylinder_z
plant_houseplant_pot_a02	Small houseplant in a pale green planter	20.0 × 20.0 × 31.7	1.5	cylinder_z
furniture_table_side_a01	Wooden side table for the SO-101 arm	26.0 × 26.0 × 81.0	4.0	box
furniture_stool_round_a01	Round metal stool for the candy bowl	28.0 × 28.0 × 67.0	3.0	cylinder_z

All props are static. They were modelled on photos of a real porch. The ghost is a bed sheet draped over a round frame with MuJoCo cloth simulation (tools/bake_ghost_cloth.py), so its folds, pooled hem and sagging top (the 2.5 cm dish the candy lands in) come from physics. The porch, bowl and ghost textures were made from the photos with GPT-6 and tile at real-world scale. Sizes are estimates from the photos.

Scenes and robot

scenes/porch_scene.usda            the porch layout in OpenUSD: every asset referenced and placed,
                                   20 candies at their settled physics poses with wrapper variants
scenes/porch_mujoco/porch.xml      the same scene for MuJoCo (meshes and textures included)
robots/so101/so101.usdc            SO-101 visual geometry at its rest pose

The SO-101 model comes from TheRobotStudio/SO-ARM100 (Apache-2.0, license included next to the robot files), with fingertip collision pads and a wrist camera added. robots/so101/so101.usdc is visual only; the arm's physics (joints, servos, contacts) is in scenes/porch_mujoco/porch.xml.

Layout: the candy bowl sits on a stool against the brick wall, the arm on a side table hidden behind the ghost, and the ghost in front where kids take candy from its top. Cameras: porch_photo (where the reference photo was taken), top (overhead workspace) and wrist.

Layout

<asset_id>/simready_usd/sm_<asset_id>_01.usd     # the asset
<asset_id>/simready_usd/Textures/*.png           # textures (candy: one per wrapper variant)
<asset_id>/simready_usd/.thumbs/256x256/*.png    # thumbnails (dark + _light)
<asset_id>/com.nvidia.simready.packaging.json     # SimReady package marker
<asset_id>/.metadata/                             # root USD + bill of materials (sha256)
assets.json                                       # metadata for every asset
tools/                                            # generator, MuJoCo importer, validator

What's inside each USD

• metersPerUnit = 1, kilogramsPerUnit = 1, Z up, defaultPrim = /RootNode (kind = component) with SemanticsLabelsAPI:wikidata_class labels.
• /RootNode/Geometry/candy_body has PhysicsRigidBodyAPI.
  • visual_* are textured render meshes (watertight, with UVs). They have no collision.
  • collider_* are Cube / Sphere / Capsule prims (purpose = guide) with PhysicsCollisionAPI, PhysicsMassAPI (net weight split by volume) and a physics material binding.
• Physics materials: hard_candy (static 0.45, dynamic 0.35, restitution 0.3); licorice (static 0.75, dynamic 0.65, restitution 0.05); paper_stick (static 0.5, dynamic 0.4, restitution 0.1); wrapper_plastic (static 0.55, dynamic 0.45, restitution 0.05).
• MuJoCo contact hints on each collider: mjc:condim = 6, mjc:solref = [0.004, 1], mjc:solimp = [0.95, 0.99, 0.001, 0.5, 2], mjc:torsionalFriction = 0.005, mjc:rollingFriction = 0.002. MuJoCo's default contact softness lets 5-17 g objects sag 4-5 mm into the table; these keep resting penetration below 1 mm.

The colliders are primitives, not the rounded wrapper mesh. Real wrappers are soft and lie flat, while a rigid rounded mesh keeps rocking in simulation.

Physics checks

Every candy and prop passes NVIDIA's omni_asset_validate (all default rules, every wrapper variant). The scene and robot files pass with three mesh-quality rules skipped (ManifoldChecker, ZeroAreaFaceChecker, WeldChecker) because of the SO-101's source CAD meshes. tools/validate_simready.py checks the conventions above. A MuJoCo test drops each candy 5 times from 10 cm at random orientations, settles 40 mixed candies in a bowl, and has an SO-101 grasp each type off a table. Grasp success is informational; flat packets lying flat are hard for the real gripper too.

candy                      mass  drop x5: resting / impact penetration    SO-101 grasp
OK   bubbleGumTwist_a01       4.5 g  rest 0.06 mm, impact 5.3 mm              2/3
OK   fruitChew_a01            5.0 g  rest 0.06 mm, impact 5.3 mm              1/3
OK   fruitSnackPouch_a01     22.7 g  rest 0.03 mm, impact 7.4 mm              0/3
OK   funSizeChocolate_a01    17.0 g  rest 0.03 mm, impact 7.0 mm              3/3
OK   funSizePouch_a01        15.6 g  rest 0.03 mm, impact 7.6 mm              0/3
OK   licoriceTwist_a01       11.3 g  rest 0.06 mm, impact 7.7 mm              2/3
OK   lollipop_a01             6.0 g  rest 0.89 mm, impact 4.9 mm              3/3
OK   powderPacket_a01        10.0 g  rest 0.03 mm, impact 7.8 mm              0/3
OK   rootBeerBarrel_a01       6.0 g  rest 0.06 mm, impact 5.4 mm              3/3
OK   sourStrawsPack_a01      15.0 g  rest 0.03 mm, impact 7.7 mm              2/3
OK   tabletRoll_a01           7.1 g  rest 0.06 mm, impact 5.2 mm              2/3
OK   taffyBar_a01            15.6 g  rest 0.03 mm, impact 7.0 mm              2/3
OK   taffySquare_a01          9.8 g  rest 0.03 mm, impact 8.6 mm              0/3
OK   twistWrap_a01            5.7 g  rest 0.03 mm, impact 7.5 mm              0/3
OK   pile in SimReady bowl: 40 candies, max speed 2.0 mm/s, 31 spilled over the rim

Loading in MuJoCo

import mujoco
from tools.simready_import import add_simready_asset

spec = mujoco.MjSpec()
spec.worldbody.add_geom(type=mujoco.mjtGeom.mjGEOM_PLANE, size=[1, 1, 0.01])
body = spec.worldbody.add_body(pos=[0, 0, 0.05])
body.add_freejoint()
add_simready_asset(spec, body, "food_candy_taffyBar_a01/simready_usd/sm_food_candy_taffyBar_a01_01.usd",
                   "taffy", collision="hull", variants={"wrapper": "bar_02"})
model = spec.compile()

Isaac Sim / any USD tool: open the .usd directly and pick a wrapper with the wrapper variant set.