Features¶

Core Architecture¶

Full MuJoCo C API accessible (mjModel*, mjData*) -- users can call any MuJoCo function directly
Async physics thread with configurable timestep
Component-based design matching MJCF element hierarchy
Singleton MjManager orchestrates compilation, stepping, and state

MJCF Import¶

Drag-and-drop XML import into Content Browser. Auto-generates Unreal Blueprints with full component hierarchy, mesh conversion, material import, and default class inheritance.

See MuJoCo Import for details.

Quick Convert¶

Convert any Unreal Static Mesh to a MuJoCo physics body with one click
CoACD convex decomposition with adjustable threshold (0.01-1.0)
Visual/collision mesh separation (group 2 visual, group 3 collision)
Static, dynamic, or mocap-driven modes
Friction, solver parameters configurable per component

Heightfield Terrain¶

Convert Unreal Landscapes to MuJoCo heightfield collision
Configurable sample resolution (2-512)
Complex trace mode for accurate mesh sampling
Actor whitelist filtering
Binary cache for fast recompilation
Editor grid preview visualization

Physics Simulation¶

Solvers: PGS, CG, Newton
Integrators: Euler, RK4, Implicit, ImplicitFast
Sim speed control (5-100%)
Pause, resume, reset, single-step
Snapshot capture and restore (full state)
Keyframe reset and hold
Sleep support with per-body policies
Multi-CCD collision detection
Gravity, wind, density, viscosity

Actuators (8 types)¶

Motor, Position, Velocity, IntVelocity, Damper, Cylinder, Muscle, Adhesion

All transmission types: joint, tendon, site, body, slider-crank
Runtime control via SetControl() (BlueprintCallable)
Gain/bias parameter arrays
Control/force/activation range limits

Sensors (40+ types)¶

Site-attached: Touch, Accelerometer, Velocimeter, Gyro, Force, Torque, Magnetometer, RangeFinder, CamProjection

Joint/tendon/actuator: Pos, Vel, Frc for each

Ball joint: Quat, AngVel

Frame: Pos, Quat, XYZ axes, LinVel, AngVel, LinAcc, AngAcc

Subtree: COM, LinVel, AngMom

Geometric: InsideSite, GeomDist, GeomNormal, GeomFromTo

Global: Contact, EPotential, EKinetic, Clock

All readable via GetSensorReading() / GetSensorScalar() on the articulation (BlueprintCallable)

Controllers¶

PD, passthrough, keyframe, and twist controllers for articulations. Keyframe controller includes built-in presets for Franka, UR5e, Spot, ANYmal, and Go2.

See Controller Framework for details.

Debug Visualization¶

Hotkey-driven toggles during PIE for contact forces, collision wireframes, joint axes + limit arcs, visual-mesh visibility, constraint-island colouring, instance + semantic segmentation shaders with sleep modulation, and smooth muscle/tendon tubes with activation-driven colour and radius.

See Debug Visualization for the full hotkey table and per-mode details.

Interactive Perturbation¶

Mouse-driven body manipulation during PIE. Double-click LMB to select, Ctrl+RMB drag to translate, Ctrl+LMB drag to rotate — the same gesture map as MuJoCo's simulate viewer. Driven by the upstream mjv_applyPerturbForce / mjv_applyPerturbPose primitives, so the spring feel is identical. Works while paused for manual pose authoring (pairs well with snapshot keyframe capture).

See Interactive Perturbation for gesture mapping, gizmo reference, and paused-mode workflow.

Camera System¶

MjCamera: scene capture attached to MuJoCo sites; per-camera CaptureMode for photoreal RGB, depth, semantic segmentation, or instance segmentation; streams via ZMQ; respects post-process volumes
MjOrbitCamera: auto-orbiting cinematic camera with target detection, height oscillation, configurable focal length
MjKeyframeCamera: waypoint-based camera path with smooth interpolation, O key play/pause

See Camera Capture Modes for per-mode behaviour and the simulate-widget preview, and Blueprint Reference for camera API details.

Networking (ZMQ)¶

Binary pub/sub protocol broadcasting joint state, sensors, and camera images. Receives actuator targets and PD gains. Prefix-based filtering for multi-articulation scenes.

See ZMQ Networking for details.

Python Integration (urlab_bridge)¶

Policy GUI: DearPyGui dashboard for policy selection, articulation targeting, force twist override, motion/gait preset selection
UnrealEnv: RoboJuDo-compatible environment over ZMQ with auto-detection, forward kinematics, born-place alignment
Pre-integrated policies: Unitree G1 locomotion (12/29 DOF), BeyondMimic (dance/violin/waltz/jump), Walk-These-Ways Go2, AMO, H2H, Smooth, ASAP
ROS 2 bridge: republishes ZMQ streams to standard ROS 2 topics

Replay System¶

Record and play back full simulation state. Supports CSV trajectory import, snapshot capture/restore, and multi-session management.

See Blueprint Reference for recording/replay API.

Blueprint API¶

40+ BlueprintCallable functions for simulation control, state queries, actuator commands, snapshots, and replay.

See Scripting with Blueprints for the full API reference.

In-editor dashboard replicating MuJoCo's simulate UI
Physics options: timestep, solver, integrator, iterations, sim speed
Per-actuator control sliders
Joint state readouts
Sensor value display
Camera feed thumbnails
Articulation selector dropdown
Record/replay/snapshot controls
Possess button for spring-arm camera follow

Forces¶

MjImpulseLauncher: velocity-based object launching (matching MuJoCo's object_launcher_plugin approach)
Direct mode (arrow direction + speed) or targeted mode (ballistic arc toward target actor)
Reset-and-fire for iterative testing during PIE
F key fires all launchers in scene

Simulation Hotkeys¶

See Hotkeys for the complete list of keyboard shortcuts.

Mesh Preparation Tool¶

Scripts/clean_meshes.py: XML-aware mesh converter
Parses MJCF, detects GLB stem conflicts, renames automatically
Converts OBJ/STL to GLB with UV preservation
Strips embedded textures to prevent Interchange import conflicts
Outputs updated _ue.xml ready for drag-and-drop import