Simulation & digital-asset dossier · 01

MuJoCo Menagerie: treating every simulation model as a versioned evidence package

Connect MJCF, meshes, scenes, model-specific licences, minimum MuJoCo versions and changelogs without mistaking a curated simulation asset for manufacturing geometry or validated physical truth.

Evidence statusCurated simulation collection · model-specific provenance and licences · physical fidelity must be measured per useSources checked 4 August 2026

Identify the exact asset first

A catalogue name cannot replace the exact format, upstream commit, generation arguments, runtime and licence.

MuJoCo Menagerie is a Google DeepMind-curated collection of high-quality MuJoCo models contributed from many upstream projects. Each model directory is its own package of XML, meshes, scene, README, licence and often a model-level changelog; some models also provide MJX-compatible variants.

Digital asset map

Put formats, geometry, dynamics, scenes, loaders, versions and provenance onto one dependency map.

01

Model definition

Publicly available

A primary MJCF XML and scene XML, with MJX variants for selected models.

Verify before adoption

Freeze model directory, repository commit, minimum MuJoCo version, variant, includes, compiler defaults and runtime options.

02

Geometry & materials

Publicly available

Visual and collision meshes, textures and scene resources supplied with each model.

Verify before adoption

Separate visual from collision geometry; inspect scale, orientation, mesh repair, simplification, material provenance and missing dependencies.

03

Dynamics & actuation

Publicly available

Mass, inertia, joints, limits, contacts, friction, tendons and actuator parameters encoded in MJCF.

Verify before adoption

Compare against measured hardware data and document every inferred, fitted, default or deliberately simplified parameter.

04

Provenance & change record

Publicly available

Per-model README, licence and changelog plus repository-level citation and global changelog.

Verify before adoption

Retain upstream source, conversion steps, exact asset licence, local modifications and the behaviour affected by each update.

Layered licence reading

Loaders, model XML, meshes, textures, brands and generated outputs may belong to entirely different rights layers.

Asset layerKnown statusAdoption decision
Individual model XML & assets

The repository explicitly states that each model directory can carry different licence and copyright terms.

Read and archive the model directory's own LICENSE before use; never inherit the top-level licence automatically.

Repository-level tooling & content

Content outside individual model assets is identified as Apache-2.0 at repository level.

Preserve notices and identify whether each copied file is repository tooling or model-specific material.

Upstream brands & physical designs

A model licence does not necessarily grant trademarks, product documentation, patents or manufacturing rights.

Keep simulation reuse separate from claims about affiliation, physical reproduction or commercial hardware rights.

This table is a source-navigation and engineering-decision aid, not legal advice. Use original licence texts and qualified advice for an actual decision.

Model & asset validation gates

From identity freeze and structural checks to behavioural calibration and experiment replay, prove what the model can support one gate at a time.

  1. 01

    Asset identity freeze

    Evidence to preserve

    Repository commit, model path, README, licence, changelog, MuJoCo version, variant and checksums for every mesh and XML.

    Stop condition

    Stop if licence, source model or compatible runtime cannot be resolved unambiguously.

  2. 02

    Structural validation

    Evidence to preserve

    Compiler warnings, joint tree, limits, inertials, collision pairs, contacts, actuator mapping and static pose checks.

    Stop condition

    Do not train or compare controllers on unresolved compiler, topology or unit errors.

  3. 03

    Behavioural calibration

    Evidence to preserve

    Physical measurements for free fall, joint response, contact, friction, payload, saturation and latency under declared conditions.

    Stop condition

    Do not claim sim-to-real relevance for behaviours that have no physical calibration evidence.

  4. 04

    Experiment reproducibility

    Evidence to preserve

    Frozen model and scene, solver, timestep, seed, controller, initial state, hardware/software environment, outputs and failure runs.

    Stop condition

    A video or average score without configuration and failure records is not reproducibility evidence.

FUURAA analysis

A digital asset becomes valuable when a team can explain where it came from, how it changed and which behaviours have been proven.

01

Best use

Rapidly discovering well-structured MJCF references and comparing modelling choices across arms, quadrupeds, humanoids and other form factors.

02

Main risk

Visual quality and successful loading can conceal unvalidated dynamics, model-specific licences and changes introduced during upstream conversion.

03

FUURAA judgement

Use Menagerie as a model-evidence starting point. A trustworthy experiment adds an immutable asset manifest and a calibration card for every behaviour the result depends on.

Applicability boundary

Simulation assets can accelerate research, training and integration; they do not automatically become manufacturing drawings, a physical digital twin or a safety case.

Can support

  • MJCF model discovery and comparison
  • Versioned simulation experiments
  • Model-calibration and provenance training

This page cannot replace

  • Manufacturing drawings or physical BOM
  • Measured dynamics and hardware safety evidence
  • Per-model legal and commercial rights review
Continue into the robot product development pathFrom digital model and verification to controlled prototype →