SimReadyBench

A benchmark for simulation-ready asset reconstruction

Chris von Csefalvay and Tamas Foldi

HCLTech

SimReadyBench measures how well models reconstruct simulation-ready USD assets from reference images, an authoring brief and nominal dimensions. Models, agent harnesses and authoring skills are independent experimental factors. Evaluation reports three independent composite scores: formal fidelity, visual fidelity and kinematic fidelity.

Reference sets

NV25 and NV100 are distinct public reference sets, separate from the withheld internal benchmark reference set. The packaged nvidia-25 manifest defines NV25, which contains 25 NVIDIA SimReady assets.

Evaluation protocol

The NV25 protocol uses four independent reconstructions per asset, giving 100 attempts for each model, harness and skill configuration. Reference observations and eligible measurements are fixed before authoring. Each attempt records the model settings and reconstruction identity.

Experimental factor	Definition
Public reference set	NV25, with the same 25 assets and their declared dependencies for every configuration
Model	The requested model with recorded generation and reasoning settings
Harness	The selected agent harness, such as Pi or OpenCode, used to author and submit the USD
Skills	The supplied authoring skill set, including an explicit empty condition
Repetitions	Four distinct reconstruction identities for each reference

During a reconstruction, the agent receives reference views, dimensions and an authoring brief. It can inspect the asset observations and request validator feedback as it authors a self-contained USD package. Final evaluation uses the submitted package, with shared orientation-aligned cameras for reference and reconstruction captures.

Evaluation factors

The three composites describe different properties of the reconstruction and are reported independently on a 0-100 scale:

Composite	Factors	Report field
Formal fidelity	Mandatory SimReady Robotics-Prop requirements and strict profile pass	formal_fidelity_percent
Visual fidelity	Reference-scale shape and appearance material correspondence	visual_fidelity_percent
Kinematic fidelity	Joint recovery, connectivity and correspondence of allowed relative poses	kinematic_fidelity_percent

Formal fidelity

SVG | Editable Mermaid source

Profile validation uses NVIDIA SimReady-Foundation Robotics-Prop 4.0.0. The 46 mandatory requirements cover the structural, physical and semantic properties of a simulation-ready prop. Each mandatory requirement contributes equally to 95 points, with five additional points when all 46 pass:

Formal fidelity = 95 x passed mandatory requirements / 46
                + 5 when all 46 mandatory requirements pass

Harbor receives this value as simready_reward. Strict acceptance requires all 46 mandatory requirements to pass. Multiple findings against one requirement count as one failed requirement. Missing or invalid candidate output scores zero against the same 46 requirements, while an incomplete evaluator run has a null score with credit bounds.

The profile's 51 optional requirements receive a separate score:

Optional score = 100 x passed optional requirements / 51

Both scores retain their passed, failed and unknown requirement counts. Partially measured profiles include coverage and lower and upper score bounds. Optional requirements do not change formal fidelity or its completion bonus.

Visual fidelity

SVG | Editable Mermaid source

Visual fidelity is the geometric mean of shape and appearance correspondence:

Visual fidelity = sqrt(shape_percent x appearance_percent)

Shape is the bidirectional surface F-score at the reference scale. Appearance compares supported colour, roughness, metallic response, opacity and texture values at corresponding reference surface samples. Each sample receives the mean of its eligible property matches, then the equal-area samples are averaged. The material-assignment flag, friction, restitution and density do not contribute to appearance credit. Assigned reference regions without measurable appearance inputs retain unknown credit, measurement coverage and lower and upper bounds. A factor with no eligible reference criteria is not applicable and is excluded from the mean.

Kinematic fidelity

SVG | Editable Mermaid source

The evaluator matches joints one to one through corresponding bodies and compares the relative poses permitted by their constraints. For fixed, revolute and prismatic joints, seventeen deterministic nonuniform states sample each motion probe. Both constraints act on the same reference-body landmarks, so reconstructed surface detail does not count twice as kinematic error.

band = 0.001 m + 0.01 x reference bounding-box diagonal
motion credit = 1 / (1 + (motion_rms_m / band)^2)
joint quality = motion credit x range credit x bound-pattern credit
group score = 100 x 2 x sum(matched joint quality)
                  / (reference joint count + candidate joint count)

Finite ranges receive continuous credit for their unwrapped travel span, using 1 / (1 + (span_error / span_band)^2) with a span band of five degrees or 5 mm plus 5% of the reference span. Bound-pattern credit is 1 / (1 + differing_endpoints), where an endpoint differs when one interval is finite and the other is infinite. Fixed constraints have unit range and bound-pattern credit.

Known unbounded revolute joints use a canonical probe spanning one full revolution and prismatic joints use one reference diagonal. A half-bounded probe is anchored at its finite endpoint, while a fully unbounded probe uses a canonical coordinate origin. Authored bounds remain unchanged in the measurement record. Missing, blocked or contradictory bound observations remain unknown.

Type, connectivity and the relevant enabled/kinematic states must agree for motion credit. Missing joints receive no credit and additional joints increase the denominator. Fixed attachments and moving joints form separate groups. Their geometric mean gives the kinematic composite when both groups apply, so fixed attachments cannot compensate for missing movable joints. Equivalent axis signs, reversed limits, exchanged joint endpoints and pivots shifted along the same hinge line retain the same score. Unsupported reference constraints and ambiguous body correspondence remain unknown. A reference with no joints is not applicable when the candidate also has none, while additional candidate joints score zero.

Detailed correspondence measurements

Units and up-axis are normalised to metres/Z-up, then one whole-object translation and rotation about the vertical axis align the candidate with the reference. The same transform is applied to geometry, bodies, joint frames and material samples, preserving object scale and relative part transforms. The evaluator performs no scale fitting or independent part fitting. Reference and candidate captures use this alignment and shared cameras.

Factor	Measurements
Shape	Bidirectional surface correspondence, surface F-score, precision, recall and distance statistics
Rigid parts	Corresponding body geometry, body count, relative arrangement and enabled/kinematic states
Joints	Joint count, type, connectivity, enabled state, anchors, relative frames, axes and limits
Materials	Supported surface properties, UV textures and authored friction, restitution, density and combine modes

The structured composite_scores readout retains each composite's factors, earned credit, eligible denominators, measured counts, coverage and score bounds. Unknown measurements remain null and criteria absent from the reference are not applicable. Candidate omissions of eligible properties receive no credit. The evaluation guide defines the comparison bands and component calculations.

Orientation-aligned comparison

Reference observations

Eyeglasses	Open cardboard box

Reference observations cover geometry, physical structure and material appearance across the asset set. Reconstruction readouts contain reference and candidate captures from matching isometric, front and side cameras, together with the generated USD package, its NVIDIA validation findings and the Harbor conversation trace.

Readouts and leaderboard

Reconstruction readouts group completed reconstructions by evaluation configuration. Each readout contains a browsable report, a score table and the generated assets and evaluation files:

runs/<run-id>/
  report.html
  results.json
  scores.csv
  assets/<asset-id>/reconstruction-<n>/candidate/
  evidence/<asset-id>/reconstruction-<n>/

The dataset viewer displays the score tables. The leaderboard summarises complete NV25 evaluations with four reconstructions per asset, while individual readouts retain the measurements for every completed reconstruction.

Evaluate a model

Use Python 3.12 in a Linux or WSL shell. Set DAYTONA_API_KEY and the credentials required by the model provider, then install the package and import the published runtime:

git clone https://github.com/hcltech-robotics/simreadybench.git
cd simreadybench
python3.12 -m venv .venv
. .venv/bin/activate
python -m pip install '.[daytona,publish]'
simready-bench setup daytona --output ../simreadybench-runtime.json

Run the benchmark with the saved runtime configuration:

simready-bench benchmark nvidia-25 --repetitions 4 \
  --model MODEL_ID --harness pi --deploy daytona \
  --runtime-images ../simreadybench-runtime.json \
  --authoring-turns 64 --workers 4 --supervise \
  --output-dir ../simreadybench-results

Replace MODEL_ID with the model identifier and add --skills PATH for an authoring skill condition. Add --max-spend-usd AMOUNT to set a separate monetary ceiling. For local execution, follow the deployment guide to configure a Linux Docker host, then use --deploy local with its local-runtime-images.json. Add --hf-repo your-org/your-results --hf-run your-run-id to publish completed reconstruction readouts to a Hugging Face dataset.

The quickstart covers model evaluation and result inspection, and the deployment guide describes the local and Daytona runtimes.

Citation

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

NV25 leaderboard

No complete NV25 sets have been published.