projectsim/yam-real-to-sim-scenes

Real-to-Sim YAM Benchmark: scene splats

The two Gaussian-splat scenes of Measuring Asset Reconstruction Effects in Real-to-Sim Robot Evaluation (Verma, Cilio, Christodoulou, Fermelis, Ng, Chung, Liang, Kaedim, 2026; arXiv:2610.00731). The scene is one of the two things that differ between the paper's authored and default reconstructions (the other is the objects). Both were built from the same handheld phone video of the evaluator's robot cell, both are aligned to the same simulated table and cut the same way, and both render through Isaac Sim's RTX splat renderer. Physics never comes from the splat: the table and floor are authored colliders hidden from every camera.

What's included

scenes/authored/splat_env2_k080.usd          the authored-arm scene, as loaded by the reported runs (168 MB)
scenes/authored/option3_splat.ply            its source: the 3DGS training output, 796,079 gaussians, before alignment and cut (197 MB)
scenes/default/polaris_env_final.usd         the default-arm scene, as loaded by the reported runs (110 MB)
scenes/default/polaris_scene_2dgs_raw.ply    its source: the 2DGS training output, 604,023 gaussians, before alignment and cut
scenes/default/fit.json                      the registration of that splat onto the aligned authored scene
scenes/default/splat_xform.txt               the resulting 4x4 (16 floats, row-major USD layout), read by the runner as SPLAT_XFORM_FILE
scenes/default/colmap_cam_centers.json       COLMAP camera centres in the splat frame; fixes which way is up
scenes/default/fit_2dgs.json, fit_3dgs.json  the same fit computed on two intermediate clouds, kept as a record

	authored scene	default scene
splat	3D Gaussian Splatting of the room, 796,079 gaussians as trained (training tool not recorded)	COLMAP poses, then 2D Gaussian Splatting (hbb1), 30k iterations, following a PolaRiS-inspired scene recipe
alignment	table-plane fit with the plateau height set by hand, then a hand correction of the translation, baked into the USD	registered onto the aligned authored scene: scale from table height (0.1249 m per unit), yaw 203 + 3 degrees, shift from the table band; applied at load through splat_xform.txt
cut	tabletop contents and streaking gaussians removed, 796,079 to 165,704 gaussians, then converted to USD	the same filters with the 2DGS switches (CUT_NEEDLE=0, CUT_KEEP_FLAT=1), then converted
runner settings	SPLAT_DZ=-0.005	SPLAT_DZ=-0.0105, SPLAT_XFORM_FILE=splat_xform.txt

The scripts for every step are in scene/ of the harness repository, Kaedim/yam-sim-harness-oss, with the build commands in its README.

Use with the harness

Place splat_env2_k080.usd at $HARNESS_ROOT/ and the default folder at $HARNESS_ROOT/polaris_splat/scene/ with splat_xform.txt under fit/, which is where the runner looks when SPLAT_USD is unset (it follows the asset arm). Or point SPLAT_USD, SPLAT_XFORM_FILE and SPLAT_DZ at any of these files directly; the scoring gate records which splat a run used.

Provenance notes

• The USDs are the exact files the reported runs referenced; the gate checks the splat file name and SPLAT_DZ in every run log.
• Both source clouds are included, so either scene can be re-aligned and re-cut with the scripts in scene/. The authored cloud is in standard 3DGS PLY layout (62 float properties per gaussian); the default cloud is 2DGS output and needs pad_scale2.py before Isaac's converter will read it.
• The capture video both were trained from is released with the capture data.
• The room is the evaluator's (Robocurve's) premises, captured with their permission for this study.

Licence and citation

CC BY 4.0. 2D Gaussian Splatting (hbb1) and COLMAP are used under their own licences for the default scene.

@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}
}