projectsim/yam-real-to-sim-runs

Real-to-Sim YAM Benchmark: run artefacts and real reference

Every simulated batch that passes the configuration gate in Measuring Asset Reconstruction Effects in Real-to-Sim Robot Evaluation (Verma, Cilio, Christodoulou, Fermelis, Ng, Chung, Liang, Kaedim, 2026; arXiv:2610.00731), the real per-trial scores they are compared against, and the committed table the paper's numbers come from. With the harness repository, Kaedim/yam-sim-harness-oss, this reproduces Tables 4 to 7 and Figures 3 to 5 of the paper on a laptop.

What's included

runs/<run_id>/results.jsonl   one line per trial: score, stage reached, final poses, gripper state
runs/<run_id>/run.log         the runner's full log; the scoring gate reads the configuration from it
runs/<run_id>/job.json        the queue job that launched it (MolmoAct2 runs; some pi0.5 runs have none)
scoring/real_reference_pertrial.csv   Robocurve's real trials, one row per trial: task, policy, run, score, max
scoring/benchmark_runs.txt            the one batch per cell and arm the paper reports
scoring/table.json                    the committed table built from those runs
configs/*.env, configs/arms.json      the per-cell configuration and the two arms as the gate knows them
runs.csv                              the table below as data

Forty-six batches: the twenty the paper reports (the MolmoAct2 bowls cell is two halves, RB_BOWLS_a and RB_BOWLS_b, pooled), the other gate-passing batches listed in the paper's Appendix A, and the six small jitter-sweep shards that Appendix A lists as pooled but not reported. In the two clear-table batches one trial was voided at its first attempt and re-run, so those files hold 21 lines; the scoring skips the voided line, leaving twenty scores per batch.

run id	task	policy	arm	trials	status	batch mean
C1_blocks_molmo	blocks	molmoact2	authored	20	REPORTED	7.75
N20_K_blocks_molmo	blocks	molmoact2	authored	20	gate-passing, not reported	8.0
V2_K_blocks_molmo	blocks	molmoact2	authored	20	gate-passing, not reported	9.25
N20_P_blocks_molmo	blocks	molmoact2	default	20	gate-passing, not reported	10.5
S3_PS_blocks_molmo	blocks	molmoact2	default	20	REPORTED	12.75
V2_P_blocks_molmo	blocks	molmoact2	default	20	gate-passing, not reported	12.5
FINAL_K_blocks_pi05	blocks	pi05	authored	20	gate-passing, not reported	6.0
T1_blocks_pi05	blocks	pi05	authored	20	REPORTED	5.75
ZZ1_PS_blocks_pi05	blocks	pi05	default	20	REPORTED	5.5
A1_bot_molmo_plain	bottles	molmoact2	authored	20	gate-passing, not reported	6.5
K2_bot_molmo_plain	bottles	molmoact2	authored	20	REPORTED	5.95
V2_K_bottles_molmo	bottles	molmoact2	authored	20	gate-passing, not reported	3.45
S1_PS_bot_molmo	bottles	molmoact2	default	20	REPORTED	6.25
A3_bot_pi05_plain	bottles	pi05	authored	20	gate-passing, not reported	10.85
FINAL_K_bottles_pi05	bottles	pi05	authored	20	gate-passing, not reported	10.8
K1_bot_pi05_plain	bottles	pi05	authored	20	REPORTED	9.4
A0_PS_bot_pi05_rerun	bottles	pi05	default	20	gate-passing, not reported	9.65
B0_PS_bot_pi05_r2	bottles	pi05	default	20	REPORTED	8.9
S6_PS_bot_pi05	bottles	pi05	default	20	gate-passing, not reported	9.5
04_NS10_K_bowls_molmo	bowls	molmoact2	authored	5	pooled shard, listed not reported
DBG_JIT25_bowls	bowls	molmoact2	authored	8	pooled shard, listed not reported
DBG_JIT40_bowls	bowls	molmoact2	authored	1	pooled shard, listed not reported
JIT25_K_bowls_molmo	bowls	molmoact2	authored	2	pooled shard, listed not reported
JIT40_K_bowls_molmo	bowls	molmoact2	authored	3	pooled shard, listed not reported
N20_K_bowls_molmo	bowls	molmoact2	authored	20	gate-passing, not reported	20.0
PROBE_NS10	bowls	molmoact2	authored	1	pooled shard, listed not reported
RB_BOWLS_a	bowls	molmoact2	authored	10	REPORTED (pooled with the other RB_BOWLS half)	25.75
RB_BOWLS_b	bowls	molmoact2	authored	10	REPORTED (pooled with the other RB_BOWLS half)	25.75
V2_K_bowls_molmo	bowls	molmoact2	authored	20	gate-passing, not reported	23.25
S2_PS_bowls_molmo	bowls	molmoact2	default	20	REPORTED	16.75
C2_bowls_pi05	bowls	pi05	authored	20	REPORTED	30.5
FINAL_K_bowls_pi05	bowls	pi05	authored	20	gate-passing, not reported	28.5
S7_PS_bowls_pi05	bowls	pi05	default	20	REPORTED	23.0
V2_K_cleartable_molmo	clear_table	molmoact2	authored	21	REPORTED	42.5
S5_PS_ct_molmo	clear_table	molmoact2	default	20	REPORTED	38.65
V2_P_cleartable_molmo	clear_table	molmoact2	default	20	gate-passing, not reported	47.0
FINAL_K_cleartable_pi05	clear_table	pi05	authored	20	gate-passing, not reported	42.3
ZZ2_K_ct_pi05	clear_table	pi05	authored	21	REPORTED	41.4
SA_PS_ct_pi05	clear_table	pi05	default	20	REPORTED	35.7
N20_K_latte_molmo	latte	molmoact2	authored	20	gate-passing, not reported	5.0
V2_K_latte_molmo	latte	molmoact2	authored	20	REPORTED	5.0
S4_PS_latte_molmo	latte	molmoact2	default	20	REPORTED	3.75
V2_P_latte_molmo	latte	molmoact2	default	20	gate-passing, not reported	4.5
FINAL_K_latte_pi05	latte	pi05	authored	20	gate-passing, not reported	2.5
K3_latte_pi05_plain	latte	pi05	authored	20	REPORTED	3.75
S9_PS_latte_pi05	latte	pi05	default	20	REPORTED	0.25

Batch mean is in rubric points, as printed in Appendix A. Blank for the shards, which Appendix A reports only as a pooled figure.

Reproduce the paper's numbers

git clone https://github.com/Kaedim/yam-sim-harness-oss && cd yam-sim-harness-oss
# download this dataset, then with RUNS pointing at its runs/ directory:
make reproduce RUNS=/path/to/runs        # Table 5, the four measures of Table 4, the median check
make test-reproduce RUNS=/path/to/runs   # asserts the rebuilt table equals scoring/table.json
make figures RUNS=/path/to/runs          # Figures 3 to 5

The rebuilt table was checked byte-identical against scoring/table.json from exactly these files before upload. Both scoring scripts default to bootstrap seed 0.

Provenance notes

• Logs are raw. They contain the runner's per-step prints, physics settings, the object masses read from each asset, and the splat and camera configuration, which is what the gate verifies. Paths in them are the machine's local layout (/opt/isaacwork), which the harness README documents.
• The real per-trial scores were produced by Robocurve, a third-party evaluator, on its own robot cell under its own protocol, commissioned by Kaedim under a benchmarking agreement. Robocurve retained sole control of benchmark design, scoring and interpretation. The per-trial results are also published on Robocurve's results page, cited in the paper.
• Simulated trials were graded automatically by rules derived from the same rubric; see the paper's Discussion for what that means.

Licence and citation

CC BY 4.0 for the simulated run artefacts and the table. The real per-trial scores are Robocurve's results, republished here with attribution as the agreement allows.

@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}
}