mujoco-mojo 2.6.8

A complete MJCF lifecycle and trial orchestration suite for MuJoCo, powered by Pydantic v2.

MuJoCo Mojo bridges the gap between static XML modeling and large-scale simulation research. It provides a strongly-typed bridge for building models and a robust execution engine for running them at scale.

• Model: Build MJCFs via validated Python objects allowing for programatic generation.
• Scale: Execute multi-threaded Monte Carlo trials with built-in resume logic.
• Monitor: Track progress via a zero-dependency web dashboard and persistent logs.
• Assess: Quickly view interactive results of a trial in context of others.
• Reproduce: Automatic environment snapshotting (requirements.txt) for every job.

Installation

Install using uv (recommended):

uv add mujoco-mojo[all]

or with pip:

pip install mujoco-mojo[all]

[!NOTE]

To get the most out of MuJoCo Mojo, select additional extras from this list (or use [all] to install all of them):

• [reloaded]: the hot-reloading development loop, including its OpenGL/Viser > visualizers.
• [mesh-decomp]: concave mesh decomposition via CoACD.
• [optimize]: Bayesian optimization studies powered by Optuna.

Features

MJCF Tools

• Strongly-Typed Elements: MJCF components backed by Pydantic v2 for immediate validation.
• Semantic Validation: Early detection of structural errors and attribute mismatches before the engine starts.
• MuJoCo Alignment: Designed to mirror MuJoCo’s XML schema closely (no magic abstractions)
• Object Enumerations: Embedded MuJoCo object mappings to simplify retrieving mjOBJ IDs.
• Asset Sharing: Specialized handling of dependency by remapping assets to become shared allows for space efficient execution of complex models

Runtime Toolkit

• Forcing Functions: A library of typed Load definitions (spring-dampers, point-to-point forces, friction, and more) that apply themselves including reaction forces.
• Proximity Sensing: Built-in distance and closest-point tracking between sites and bodies, exposed as both telemetry signals and live overlays with no extra instrumentation. Works on concave geometry meshes.
• Structured Telemetry: A single post/track call records any value into a hierarchical namespace during simulations. This include prebuilt request calls for commonly desired values.
• Video Recording: Multi-camera MP4/WebM/GIF capture synced to simulation time, with playback-speed control, force/proximity overlays, on-frame labels, and single-frame snapshots.

Job Utilities

Campaign Orchestration

• Multi-Threaded Execution: Single or multi-threaded trial execution
• Environment Snapshotting: Automatically record installed Python packages to requirements.txt for job recreation (works with uv or pip)
• Resume Logic: Resume a previously started job without rerunning previous cases
• Robust Logging: Built in Rich logging for terminal and a rotating file handler for persistent logs and status files for insight on trial progress
• Global Overrides: Force specific values onto distributions via CLI or JSON overrides to test "golden" cases.

Monte Carlo

• Reproducible Sampling: Random draw tools for Monte Carlo or rerun with global variable override
• End of run summary with metric to help perform a state of health check
• Support for running jobs with SLURM for distributed compute

[!TIP]

mujoco-mojo run monte-carlo \
    --generator monte_carlo_test.Experiment.generate \
    --runtime monte_carlo_test.runtime \
    --workdir ./mc_test/ \
    --no-resume \
    --gen-arg 123 \
    --gen-kwarg 'test=1234' \
    --n-trial 10 \
    --n-proc 1

Optimization

• Bayesian Search: Intelligent design space navigation powered by Optuna integration.
• Design Variables: Continuous (DesignFloat) and discrete (DesignCategorical) parameters evolved by the solver.
• Adaptive Refinement: "Zoom" into promising neighborhoods by aggressively shrinking search bounds on resume.
• Stochastic Robustness: Multi-evaluation trials that average scores over different seeds to filter out noisy physics outliers.

[!TIP]

mujoco-mojo run optimiztion \
    -g sim.generate \
    -r sim.runtime \
    --objective sim.objective \
    --n-trial 400 \
    --n-proc 10 \
    --seed 42 \
    --storage \
    --direction minimize

Dojo Dashboard

A zero-dependency, offline-first web suite for monitoring and analyzing your simulation jobs in real-time.

Monitor: Real-Time Oversight

• Live Progress Tracking: Dynamic progress bars and color-coded status cards provide a high-level view of your Monte Carlo runs.
• Success/Failure Analytics: Automatic categorization of trials with built-in data integrity checks to identify "empty" vs. "failed" runs.
• Sensory Feedback: Optional audio cues and visual celebrations let you know exactly when a multi-hour job hits 100%.
• Deep-Linked Navigation: Jump straight from the monitor to any individual trial in the viewer with one click.

Mosaic: Advanced Telemetry Analysis

• High-Fidelity Plotting: Hardware-accelerated visualization using Plotly.js for seamless zooming and panning through millions of data points.
• Dynamic Versus Mode: Overlay current telemetry against previous trials using an intuitive range-selection slider for instant regression testing.
• Regex-Powered Filtering: Navigate high-dimensional datasets using a "folder-style" signal selector with suffix and regex support.
• State Persistence & Sharing: Every view is captured in a shareable, compressed URL by pasting a link to share your exact configuration.
• Pro-Grade Tooling: Built-in JSON configuration editor, drag-and-drop config restoration, and multi-format exports (SVG, PNG, CSV).
• Keyboard-First Design: Full hotkey support for warping between trials and managing views without leaving the home row.

Reloaded

A rapid prototyping loop that allows you to modify physics logic and model architecture on the fly without ever closing the visualizer.

• Module Hot-Reloading: Recursively reloads local Python modules and MJCF logic, allowing code changes to propagate instantly to the active simulation.
• Unified Visualizer Bridge: Synchronized visualization of custom force and torque vectors across native OpenGL, Viser web interfaces, and video recordings.
• Multiple Viewer Backends: Drive the same reload/sync pipeline through either the native MuJoCo OpenGL viewer or a Viser web viewer.
• Interactive Prototyping: A developer-centric command loop to toggle playback speeds, repeat last commands, watch for file changes and auto-reload, switch telemetry recording on or off mid-session, or trigger "generation-only" mode for rapid MJCF debugging.
• Live Session Controls: Adjust the seed and trial number on the fly with seed <N> and trial <N>, without restarting the loop.
• Asset Persistence: Automatically dumps current MJCF snapshots and model configurations to a workspace directory for post-hoc analysis or version tracking.

[!TIP]

mujoco-mojo reloaded \
    --generator monte_carlo_test.Experiment.generate \
    --runtime monte_carlo_test.runtime \

[!NOTE]

MuJoCo Mojo is an independently developed open-source toolbox. It is not affiliated with, sponsored by, or endorsed by Google DeepMind or the official MuJoCo development team. MuJoCo® is a registered trademark of Google LLC. All MJCF schemas and MuJoCo-related terminology used within this project are for compatibility and documentation purposes only.