projectsim/yam-real-to-sim-captures

Real-to-Sim YAM Benchmark: capture inputs

The raw captures behind Measuring Asset and Scene Reconstruction Effects in Real-to-Sim Robot Evaluation (Verma, Cilio, Christodoulou, Fermelis, Ng, Chung, Liang, Kaedim, 2026; arXiv:2610.00731): the photographs every object asset was built from, and the handheld video both scene splats were trained from. With these, anyone can run their own reconstruction pipeline on the same inputs and score the result against the real robot with the released harness, Kaedim/yam-sim-harness-oss.

All captures were taken on 31 August 2026 at the evaluator's robot cell, on an iPhone, with the objects on the cell's table.

What's included

objects/<asset_id>/
├── Photo_*.HEIC / .jpg              the single photograph both reconstructions were built from (no board in frame)
├── Front_*.HEIC / .jpg              front view with the printed ChArUco board, for metric scale
└── Side1_*.HEIC, Side2_*.HEIC / .jpg  two side views with the board
scene/Background_Option3.MOV         the 0.5x wide-angle walk-around of the cell both scene splats were trained from (517 MB)
objects.csv                          which photo is the pipeline input for which asset

HEIC originals are kept as shot; the .jpg beside each is a 92% quality conversion for tools that do not read HEIC. The asset ids match the two asset datasets, yam-real-to-sim-assets and yam-real-to-sim-assets-default, and the two scenes in yam-real-to-sim-scenes were trained from the video here.

object	asset id	pipeline input photo
baseball	sm_baseball_leatherstitched_713	Photo_NoCharucoBoard_Baseball.HEIC
Sapporo bottle	sm_beerbottle_glasssapporo_d06	Photo_Sapporo.HEIC
blue wooden block	sm_block_woodblue_d8b	Photo_NoCharucoBoard_BlueUBlock.HEIC
red wooden block	sm_block_woodreda_e0a	Photo_NoCharucoBoard_RedMBlock.HEIC
Heineken bottle	sm_bottle_glassheineken_8d6	Photo_Heineken.HEIC
Pellegrino bottle	sm_bottle_glasspellegrino_d5d	Photo_Pellegrino.HEIC
ribbed plastic bowl	sm_bowl_plasticribbed_07f	Photo_CreamBowl.HEIC
spray bottle	sm_cleaningbottle_plasticlabel_3fc	Photo_NoCharucoBoard_SprayBottle.HEIC
drinks can	sm_cokecan_aluminumcylindrical_5ba	Photo_NoCharucoBoard_CokeCan.HEIC
bamboo cutting board	sm_cuttingboard_bamboofunctional_c14	Photo_NoCharucoBoard_CuttingBoard.HEIC
plastic grapes	sm_grapes_thinplastichollow_a7d	Photo_RedGrapes.HEIC
red mug	sm_mug_plasticred_02c	Photo_RedMug.HEIC
steel latte art cup	sm_pitcher_metalbrushed_046	Photo_NoCharucoBoard_Pitcher.HEIC
small plate	sm_plate_plasticspeckled_3b0	Photo_PurplePlate.HEIC
puzzle cube	sm_rubikscube_plasticpuzzle_64e	Photo_RubiksCube.HEIC
blue storage bin	sm_storagebin_plasticopen_f43	Photo_BluePlasticBin.HEIC

Both reconstructions in the paper used only the Photo_* image per object. The authored arm also used the board views to estimate metric scale; the default arm typed scale by hand and did not use them. The board is a printed ChArUco pattern of known square size, placed next to the object in the front and side views.

Licence and citation

CC BY 4.0. The cell belongs to Robocurve, the third-party evaluator that ran the real trials, and the captures are released with their agreement.

@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},
  primaryClass  = {cs.RO},
  url           = {https://arxiv.org/abs/2610.00731}
}