SimReadyBench

A benchmark for simulation-ready asset reconstruction

Chris von Csefalvay, Pranav Doma, Achyuthan Unni Krishnan and Tamas Foldi

HCLTech

SimReadyBench tests whether AI agents can reconstruct simulation-ready OpenUSD objects from one image, an object description and dimensions. NV100 contains 100 objects with their source assets, task inputs and evaluation data.

Formal fidelity is the primary score: it measures structural, physical and semantic compliance with the SimReady profile. Visual fidelity measures geometry and appearance together. Kinematic fidelity measures articulation separately. All three use a 0–100 scale; there is no combined overall score.

Browse the 100 tasks · Explore results · Reference objects · Run the Harbor tasks

Leaderboard

Formal, visual and kinematic fidelity are scored separately on a 0–100 scale. Formal fidelity is the primary metric. Values are point estimates from measured results.

Model	Reasoning	Harness	Formal %	Optional criteria (%)	Visual %	Kinematic %	pass@1 %	pass@3 %	pass^3 %	Tokens/asset	Inference $/asset	$/formal point	Trial min/asset
ChatGPT 6.1 Sol	high	Codex	100.00	72.84	56.86	3.68	100.00	100.00	100.00	1,990,604	$0.7145	$0.007145	14.39
ChatGPT 6.1 Terra	high	Codex	100.00	70.62–70.64	54.03	2.79	100.00	100.00	100.00	1,653,681	$0.2735	$0.002735	7.76
ChatGPT 6.1 Terra	medium	Codex + CAD/build123d	100.00	70.89	54.11	0.71	100.00	100.00	100.00	3,917,477	$0.3961	$0.003961	8.56
ChatGPT 6.1 Sol	medium	Codex + CAD/build123d	100.00	73.47	55.56	3.00	100.00	100.00	100.00	3,369,249	$0.7383	$0.007383	9.45
ChatGPT 6.1 Sol	medium	Codex	99.98	73.22–73.24	53.66	3.14	99.67	100.00	99.00	1,064,313	$0.4145	$0.004146	9.32
ChatGPT 6.1 Sol	low	Codex	97.69	73.58–73.61	51.17	2.44	71.33	96.00	39.00	478,511	$0.2072	$0.002121	5.86
Claude Opus 5.5	low	Claude Code	97.12	73.77	51.19	2.447	67.00	94.00	37.00	381,567	$0.3126	$0.003218	4.10
ChatGPT 6.1 Terra	low	Codex	91.13	69.90–70.03	47.81	0.87	2.00	6.00	0.00	359,013	$0.0969	$0.001064	4.40
DeepSeek V4.1 Flash (1017)	none	Terminus 2	86.32	71.64–73.23	35.36	0.052	2.33	6.00	0.00	2,242,278	$0.0274	$0.000317	10.98
DeepSeek V4.1 Flash (1017)	none	Terminus 2 + CAD/build123d	86.26	64.42–64.90	37.61	0.093	53.67	87.00	13.00	3,468,986	$0.0253	$0.000293	13.92
Claude Sonnet 5.5	low	Claude Code	83.60	70.02	50.06	0.66	0.00	0.00	0.00	93,819	$0.0578	$0.000691	3.28
NVIDIA Nemotron 3.5 Super VL	none	Terminus 2	76.69	68.33–69.63	21.52	0.005	0.00–1.67	0.00–5.00	0.00	2,511,512	$0.3357	$0.004377	11.18

Optional criteria show the percentage of optional checks passed; ranges retain unknown checks. They do not affect formal fidelity.

Measurement coverage varies by configuration. Missing measurements retain explicit unknowns and bounds; the tables and linked reports preserve these distinctions.

Pass means all 46 mandatory checks pass. pass@3: 1 − C(n−c, 3)/C(n, 3); pass^3: C(c, 3)/C(n, 3), averaged per object. n = 3 for Codex, Claude Code and both Terminus 2 conditions; n = 4 for the earlier Codex + CAD/build123d conditions; c is the number of mandatory passes.

Cost per formal point is inference cost per asset divided by mean formal fidelity. Cost, tokens and trial time include retries.

Token breakdown per asset

Model	Reasoning	Harness	Total	Input	Output	Cached input
ChatGPT 6.1 Sol	high	Codex	1,990,603.84	1,959,382.44	31,221.40	1,873,388.42
ChatGPT 6.1 Terra	high	Codex	1,653,681.07	1,633,905.19	19,775.88	1,556,437.42
ChatGPT 6.1 Terra	medium	Codex + CAD/build123d	3,917,476.69	3,901,989.54	15,487.15	3,798,976.06
ChatGPT 6.1 Sol	medium	Codex + CAD/build123d	3,369,249.31	3,351,437.94	17,811.38	3,257,605.53
ChatGPT 6.1 Sol	medium	Codex	1,064,313.18	1,045,908.96	18,404.22	989,904.26
ChatGPT 6.1 Sol	low	Codex	478,511.38	470,783.68	7,727.70	436,238.10
ChatGPT 6.1 Terra	low	Codex	359,012.89	353,438.26	5,574.64	310,615.98
DeepSeek V4.1 Flash (1017)	none	Terminus 2	2,242,278.06	2,210,364.42	31,913.65	2,012,997.93
DeepSeek V4.1 Flash (1017)	none	Terminus 2 + CAD/build123d	3,468,985.56	3,434,797.42	34,188.14	3,300,037.55
Claude Opus 5.5	low	Claude Code	381,567.07	374,666.49	6,900.59	353,221.31
Claude Sonnet 5.5	low	Claude Code	93,819.46	91,460.50	2,358.97	80,386.77
NVIDIA Nemotron 3.5 Super VL	none	Terminus 2	2,511,512.17	2,464,376.29	47,135.88	2,031,610.29

Total = input + output. Cached input is already included in input.

Benchmark JSON · Sortable leaderboard · CSV · JSON

Quality and efficiency

The frontier shows configurations that cannot be improved on one axis without losing on the other. Scores use the same measured object/repetition pairs across configurations. Marker colours and initials identify model families. An outer ring identifies a custom agent. Frontier markers are filled with white initials; dominated markers are outlined with coloured initials.

Inference cost, tokens and trial time include retries. Hosted compute is separate.

Interactive charts · Chart data

Scores

Formal fidelity

The primary compliance score uses the 46 mandatory requirements of NVIDIA SimReady Robotics-Prop 4.0.0:

Formal = 95 × passed mandatory requirements / 46
         + 5 if all 46 requirements pass

The requirements cover hierarchy, naming, units, mesh properties, materials, semantics, collision and rigid-body configuration, and thumbnail metadata. Each requirement counts once, regardless of how many findings it produces. A score of 100 requires all 46 to pass. Optional requirements are reported separately and do not alter formal fidelity.

Visual fidelity

Visual fidelity combines geometry similarity G with rendered appearance similarity A:

Visual = sqrt(G × A)
Appearance = mean(four foreground RGB SSIM scores)

Each appearance score is structural similarity (SSIM) over the union of reference and candidate foreground pixels, clipped to 0–1 and expressed as a percentage. The four fixed comparison views use a shared whole-object alignment. Appearance is a component of visual fidelity, not an alternative name for it.

Geometry measures bidirectional surface agreement at the reference scale, without scale fitting. Evaluator 2.2.0 uses 65,536 area-weighted surface samples per object: directional credit is mean(1 / (1 + (distance / band)^2)), with a band of 2% of the reference cloud's 1st–99th-percentile bounding-box diagonal. G is 100 times the harmonic mean of the two directional credits.

Kinematic fidelity

Kinematic fidelity measures recovered joints and the relative motion their constraints permit. Matched joint quality combines motion agreement, travel-range agreement and limit-pattern agreement. Missing joints lose credit; extra joints increase the denominator.

Group score = 100 × 2 × sum(matched joint quality)
              / (reference joint count + candidate joint count)

Moving joints and fixed constraints are scored separately. Their geometric mean gives kinematic fidelity when both groups apply; a single applicable group supplies the score directly. Motion is compared at 17 deterministic probe states on common reference-body landmarks.

Run NV100 with Harbor

For local execution, use Linux x86_64, Python 3.12, Docker Engine and Docker Compose. Tasks request four CPUs and 8 GiB for the agent environment, with a separate four-CPU, 16 GiB verifier. Choose concurrency to fit the host.

python3.12 -m venv .venv
. .venv/bin/activate
python -m pip install harbor==0.24.0
harbor auth login

Set the credentials required by your model provider, such as OPENAI_API_KEY for OpenAI. Select the agent and model, then run all 100 tasks and upload the results:

AGENT=codex
MODEL=openai/gpt-6.1-terra

harbor run \
  -d hcltech-robotics/simreadybench \
  -a "$AGENT" -m "$MODEL" \
  --n-attempts 1 --n-concurrent 1 \
  --max-retries 3 --upload

Change AGENT and MODEL for another supported agent. --n-attempts sets scored attempts per task; --max-retries limits exception retries. Increase --n-concurrent for parallel execution. --upload sends trial results, trajectories and artefacts to Harbor Hub; uploads are private by default.

Use the same command with --dry-run to check the configuration without running trials. Unversioned dataset names select the latest release, currently evaluator 2.2.0. For hosted execution and provider-secret configuration, see Harbor's run guide.

Each task supplies one observation image and its brief. The agent authors a self-contained package with asset.usd at its root and can request up to five formal-validation feedback calls. Final visual evaluation uses four separate comparison views. Complete trajectories are required; retrieving or importing external USD assets is a protocol failure. Locally authored USD and task-provided inputs are allowed.

Use the data

Browse tasks · Browse results · Reference objects · Published runs and rollouts

The default tasks/nv100 view contains one row per object, with its observation image, instruction, dimensions, reference comparison images and Harbor task link. The observation and task brief are agent inputs; reference images and source assets support evaluation.

Use the configuration selector to browse final reconstruction results or executions including retries. Final-result views are named results or <run>_results; execution views are named result_attempts or <run>_attempts. The example below loads the tasks and references and combines the final-result views across current benchmark runs.

from datasets import concatenate_datasets, get_dataset_config_names, load_dataset

repo = "hcltech-robotics/simreadybench"
tasks = load_dataset(repo, "tasks", split="nv100")
references = load_dataset(repo, "references", split="nv100")

result_configs = [
    name for name in get_dataset_config_names(repo)
    if name == "results" or (name.endswith("_results") and not name.startswith("legacy_"))
]
results = concatenate_datasets([
    load_dataset(repo, name, split="test") for name in result_configs
])

by_id = {row["id"]: row for row in references}
reference = by_id[results[0]["reference_id"]]

Join results to tasks or references using reference_id. Filter by requested_model, condition_id or harbor_job_id to select a model, configuration or run.

Each result links to its available scores, generated assets and rollouts. Trial folders follow results/harbor/<harbor_job_id>/<harbor_trial_id>/. Use the trajectory and recording links to inspect an agent's reconstruction process. Download USD dependencies together with the entry point.

Benchmark code is MIT-licensed. NVIDIA source assets retain their CC BY 4.0 licence and attribution.

Citation

@software{simreadybench2026,
  title = {SimReadyBench: Evaluating models, harnesses and skills
           for simulation-ready asset generation},
  author = {von Csefalvay, Chris and Doma, Pranav and
            Unni Krishnan, Achyuthan and Foldi, Tamas},
  organization = {HCLTech},
  year = {2026},
  version = {1.0.0},
  url = {https://github.com/hcltech-robotics/simreadybench}
}