mjviser

A web-based MuJoCo viewer built on Viser.

Quick start

Run it directly with uvx (nothing to install):

uvx mjviser path/to/model.xml

Or pip install mjviser and run mjviser path/to/model.xml.

mjviser also does fuzzy path matching against the current directory:

mjviser humanoid        # finds **/humanoid*.xml
mjviser shadow_hand     # finds **/shadow_hand*.xml

Use --port to bind to a specific port (default: 8080):

mjviser model.xml --port 7070

If robot_descriptions is available, you can load any of its 57 MuJoCo models by name:

uvx --with robot_descriptions mjviser go1

Note

uvx defaults to the system Python and may not respect requires-python constraints (astral-sh/uv#8206). If your system Python is 3.14+, where MuJoCo can't build yet, pass -p 3.13 explicitly: uvx -p 3.13 mjviser path/to/model.xml.

Python API

import mujoco
from mjviser import Viewer

model = mujoco.MjModel.from_xml_path("robot.xml")
data = mujoco.MjData(model)
Viewer(model, data).run()

Open the printed URL in your browser. You get most of what the native MuJoCo viewer offers: simulation controls, joint and actuator sliders, contact and force visualization, camera tracking, keyframes, and more.

Extension points

Viewer accepts three optional callbacks:

• step_fn(model, data): called each simulation step. Defaults to mujoco.mj_step.
• render_fn(scene): called each render frame. Defaults to scene.update_from_mjdata(data).
• reset_fn(model, data): called on reset.

For full control, use ViserMujocoScene directly. The server is a standard Viser server, so you can add GUI elements, scene overlays, or anything else Viser supports.

server = viser.ViserServer()
scene = ViserMujocoScene(server, model, num_envs=1)
scene.create_visualization_gui()

with server.gui.add_folder("My Controls"):
    slider = server.gui.add_slider("Force", min=0, max=100, initial_value=0)

while True:
    mujoco.mj_step(model, data)
    scene.update_from_mjdata(data)

Examples

• active_viewer.py: simplest usage with playback controls
• active_viewer_with_controller.py: custom step_fn with random torques
• passive_viewer.py: manual simulation loop with ViserMujocoScene
• multi_env.py: 4 humanoids in parallel via mujoco-warp
• ghost_overlay.py: custom render_fn that overlays a time-delayed ghost
• motion_playback.py: recorded trajectory with timeline scrubber, speed control, and contact replay

Acknowledgments

Thanks to Matija Kecman for early feedback and suggestions.

Limitations

• No mouse interaction: clicking/dragging bodies and keyboard callbacks require upstream Viser support.