Back to projects
v1.4.0 · stableROS 2 LyricalApache-2.0Open source

Lagrangemulti-physics robot simulator

ROS 2 simulator with switchable multi-physics: robots, drones and cars in a living town with weather, water, traffic and people

13+
physics domains, each switchable
2,500+
automated tests
40 km
largest streamed world
v1.4
stable, frozen 1.x APIs

What it is

A Gazebo-style simulator where every physical domain is its own switch

Built for ROS 2 Lyrical on Ubuntu 26.04. Motion, power, heat, air, water, weather, interference and sensors run together on a multirate co-simulation master, and a deterministic CPU mode reproduces every run bit for bit.

Physics you can switch off

Every domain is a module on a unit-checked signal bus. Turn one off and an ideal, constant or replayed substitute takes its place, so a test pays only for the physics it needs.

From a warehouse to a city

Double-precision worlds stream in tiles with a floating origin that follows the robot, built from elevation models, OpenStreetMap and OpenDRIVE roads.

A native GPU viewer

A Vulkan desktop app with Dear ImGui panels: scopes at 10 kHz, the live circuit, fields and sensors, all drawn by the same Slang shaders as the camera sensors.

Physics

Thirteen domains, one signal bus

Units are checked when the run is wired: a module that writes N·m into an input that wants N stops the start instead of silently converting.

Rigid body
MuJoCo or Jolt
Electrical
battery, bus, drives at 100 µs
Thermal
lumped + conduction in parts
Aerodynamics
drag, rotor airflow, GPU LBM
Hydrodynamics
buoyancy, Fossen, waves
Weather
fog, rain, snow, dust
EMI & comms
power lines, Wi-Fi, CAN
Vehicle dynamics
MF-Tyre, suspension
Traffic & people
IDM cars, walking humans
Rotor aero
blade element, ground effect
Sensors
37 kinds, lidar, radar, camera
Scheduler
multirate, deterministic CPU mode
Airflow around a body, GPU lattice Boltzmann
Airflow around a body, GPU lattice Boltzmann
Heat conducted through a part
Heat conducted through a part
A robot's own wiring disturbing its compass
A robot's own wiring disturbing its compass
A thermal camera sees people
A thermal camera sees people

Robot classes

Mobile robots, drones, cars and arms

Import URDF, SDFormat or MJCF as they are; the samples include real models such as Autoware's Lexus RX 450h and Bitcraze's Crazyflie 2.x.

Mobile robots
Mobile robots

Nav2 and ros2_control, 2D/3D lidar on CPU or GPU, people walking the aisles.

Drones
Drones

Rotor aerodynamics, PX4 and ArduPilot SITL in lockstep, wind fields.

Cars
Cars

Magic Formula tyres, OpenDRIVE roads, OpenSCENARIO, an Autoware interface.

Robot arms
Robot arms

MoveIt 2, motor drives and heating, a battery that sags under load.

New in v1.4

Build a robot in the viewer

Press Ctrl B: start from a template or a URDF, add parts and meshes, pick how each one collides, mount sensors from a catalog, wire a battery and motors, then press Test to drive it in a world. It saves an ordinary URDF.

Links, joints and shapes, mass from density
Links, joints and shapes, mass from density
A mesh split into convex collision parts
A mesh split into convex collision parts
Battery, bus and a motor on each wheel
Battery, bus and a motor on each wheel
Test drives it on its own server
Test drives it on its own server

See it run

Every clip is a real run

Recorded from the headless viewer. Hover to play, click to watch.

Timeline

From first commit to v1.4 in five weeks

  1. v0.1Oct 14, 2026

    Viewer, protocol, mock server

  2. v0.2–0.6Oct 25, 2026

    Sim core, arms, drones, weather & EMI, cars

  3. v0.7Oct 29, 2026

    Huge worlds

  4. v1.0Oct 29, 2026

    Stable: schemas, plugin ABI, protocol frozen

  5. v1.1Nov 7, 2026

    Plugins, walking people, water, rotor airflow

  6. v1.2Nov 11, 2026

    Heat in parts, a robot's own EMI, thermal cameras

  7. v1.3Nov 14, 2026

    37 sensor kinds, 1.3–4.6× faster

  8. v1.4Nov 15, 2026

    Build a robot in the viewer

The hard part

Simulating many physical domains together without forcing every user to pay for all of them, scaling from one warehouse to city-sized worlds without losing float precision, and being honest about accuracy when almost no measured robot data is available yet.

How it is solved

Split each domain into a module on a unit-checked signal bus that can be disabled or replaced by ideal, constant or replayed data, and streamed worlds in tiles with a floating origin that follows the robot. Each model has named tests against closed forms or published references, with a validation report that states what is not met or not yet compared with real robots.

Built with

The stack

C++PythonROS 2VulkanSlangOpenUSDMuJoCoJolt PhysicsNav2MoveIt 2PX4ArduPilotAutowareOpenDRIVEOpenSCENARIONanoVDBMCAP
Share: