projectsim/yam-real-to-sim-assets-default

Real-to-Sim YAM Benchmark: default-recipe assets

The sixteen objects of the default reconstruction in Measuring Asset Reconstruction Effects in Real-to-Sim Robot Evaluation (Verma, Cilio, Christodoulou, Fermelis, Ng, Chung, Liang, Kaedim, 2026; arXiv:2610.00731). These are the comparison arm, built by us following a PolaRiS-inspired recipe from the same single photograph per object as the authored arm (yam-real-to-sim-assets):

• mesh and texture generated by TRELLIS in single-image mode;
• metric scale typed by hand as the object's largest dimension, judged from the photograph and everyday familiarity, with no measurement (scales_by_eye.json is the sheet);
• collider: convex decomposition on every object;
• no physics authored: no mass, no material. PhysX therefore applies its defaults at load, density 1000 kg/m3 over the collider volume, friction 0.5, restitution 0.

They are the exact files loaded by the ten default-arm runs the paper reports. The recipe and every departure from it are documented in asset-arms/default-path/ of the harness repository, Kaedim/yam-sim-harness-oss, together with the generation scripts.

What's included

object	task	size (cm)	typed size (m)	collider	mesh faces
Heineken bottle (sm_bottle_glassheineken_8d6)	bottles in bin	6.3 x 6.2 x 23.0	0.23	convex decomposition	4872
Sapporo bottle (sm_beerbottle_glasssapporo_d06)	bottles in bin	6.1 x 6.2 x 23.0	0.23	convex decomposition	4926
Pellegrino bottle (sm_bottle_glasspellegrino_d5d)	bottles in bin	8.0 x 8.1 x 28.0	0.28	convex decomposition	5256
blue storage bin (sm_storagebin_plasticopen_f43)	bottles in bin, clear table	25.5 x 30.0 x 14.9	0.30	convex decomposition	36495
red wooden block (sm_block_woodreda_e0a)	stack blocks	3.8 x 3.7 x 4.0	0.04	convex decomposition	34584
blue wooden block (sm_block_woodblue_d8b)	stack blocks	3.8 x 3.7 x 4.0	0.04	convex decomposition	30150
ribbed plastic bowl (sm_bowl_plasticribbed_07f)	stack bowls (x3)	15.0 x 15.0 x 7.5	0.15	convex decomposition	20582
steel latte art cup (sm_pitcher_metalbrushed_046)	move latte cup	12.0 x 8.9 x 9.2	0.12	convex decomposition	35678
bamboo cutting board (sm_cuttingboard_bamboofunctional_c14)	move latte cup	35.0 x 25.7 x 1.0	0.35	convex decomposition	10550
drinks can (sm_cokecan_aluminumcylindrical_5ba)	move latte cup	6.5 x 6.5 x 12.0	0.12	convex decomposition	12282
spray bottle (sm_cleaningbottle_plasticlabel_3fc)	move latte cup	14.2 x 8.4 x 25.0	0.25	convex decomposition	6880
puzzle cube (sm_rubikscube_plasticpuzzle_64e)	clear table	6.0 x 6.0 x 6.0	0.06	convex decomposition	28518
baseball (sm_baseball_leatherstitched_713)	clear table	7.5 x 7.5 x 7.5	0.07	convex decomposition	26833
plastic grapes (sm_grapes_thinplastichollow_a7d)	clear table	15.0 x 8.6 x 7.0	0.15	convex decomposition	14033
red mug (sm_mug_plasticred_02c)	clear table	11.7 x 9.3 x 12.0	0.12	convex decomposition	9814
small plate (sm_plate_plasticspeckled_3b0)	clear table	20.0 x 20.0 x 1.5	0.20	convex decomposition	13586

The three bowls in stack bowls are three instances of the same asset. Sizes are the axis-aligned bounding box after the typed scale is applied, Z up, metres per unit 1.0. Compare with the authored arm's table to see where a by-eye scale lands: the plate is typed at 20 cm and measures 14.9 cm on the ChArUco-scaled asset, the cutting board is a 1 cm slab here and 2.4 cm there.

Total: 16 assets, ~40 MB (TRELLIS bakes a single 1K base-colour texture).

Per-asset contents

assets/<asset_id>/
├── <asset_id>.usd        the binary OpenUSD file the benchmark runs loaded (unchanged)
├── <asset_id>.usda       the same stage exported to text USD
├── <asset_id>.json       generation record: source GLB, typed size, applied uniform scale,
│                         bounding box, face count, and the physics that is NOT authored
└── textures/<asset_id>_basecolor.png

objects.csv is the table above as data. scales_by_eye.json is the size sheet used for every object (including four that no task uses), and photo_to_asset_mapping.json says which capture photograph each asset was generated from.

Provenance and fidelity notes

• Same photograph per object as the authored arm; only the reconstruction method differs.
• Because no mass is authored, PhysX derives it from the collider volume: a hollow convex decomposition of a bowl or bin is treated as solid, so these objects are far heavier than the real ones. That is part of what the paper measures, not a bug to fix here.

Licence and citation

CC BY 4.0. TRELLIS is MIT licensed. The real-robot trials these objects were evaluated against were run by Robocurve under a benchmarking agreement with Kaedim; see the paper.

@misc{verma2026realtosim,
  title  = {Measuring Asset and Scene Reconstruction Effects in Real-to-Sim Robot Evaluation},
  author = {Verma, Sanya and Cilio, Luca and Christodoulou, Velissarios and Fermelis, Tyler and Ng, Qi Ting and Chung, Rachel and Liang, Sofia},
  year   = {2026},
  eprint = {2610.00731},
  archivePrefix = {arXiv},
  note   = {Code: https://github.com/Kaedim/yam-sim-harness-oss}
}