← All projects

iSpaRo 2026 · Open-source simulation

MarsLab

A Martian Rover Simulator for Planetary Rover Autonomous Navigation

Hoyun Kim · Beomsu Kim · Giseop Kim

Autonomy and Perceptual Robotics Lab (APRL), DGIST

Overview

MarsLab is a customizable NVIDIA Isaac Sim 5.1 platform for planetary rover autonomy research. It combines Prebuilt Martian Environments, Rover dynamics, Cameras, LiDAR, IMU, Wheel Odometry, and independent Ground-Truth trajectories in one runtime. An ROS 2 Jazzy bridge connects these outputs to external perception, SLAM, Navigation, and Visual Place Recognition (VPR) systems.

MarsLab includes four selectable rover profiles: Perseverance, Curiosity, Opportunity, and Spirit. This guide covers their shared setup and their different geometry, cameras, sensing outputs, and tuning defaults. Perseverance is selected by default, but you can change the configuration to simulate Curiosity, Opportunity, or Spirit.

Start Here

How MarsLab is organized

Each run combines a scenario YAML with a rover profile. The runtime validates the configuration and required assets before starting Isaac Sim, loads the scene and rover, applies physics and sensor settings, and then runs control, sensing, and publication. Resources are released during shutdown.

Part Role
configs/config.yaml Scenario: scene, runtime mode, environment, rendering, and optional rover overrides
configs/robots/ Profiles for Perseverance, Curiosity, Opportunity, and Spirit
assets/ Rover USD/URDF assets, downloaded terrain, and sky textures
marslab/ Runtime, configuration, physics, control, rendering, sensors, and ROS bridge
launch/ ROS companion for the robot description and articulated TF tree

Installation

Requirements

Component Setup
Operating system Linux supported by your Isaac Sim 5.1 installation
Simulator and GPU NVIDIA Isaac Sim 5.1 and a GPU supported by its renderer
ROS ROS 2 Jazzy, required when using the ROS interfaces
Python Isaac Sim’s bundled Python for simulation; ROS 2 Jazzy’s Python for the companion
Assets Rover assets from the repository and a separately downloaded Mars scene

Install Isaac Sim first. For ROS integration, also install ROS 2 Jazzy with robot_state_publisher and tf2_ros. Run the simulator and ROS companion in separate terminals.

Clone the repository and download a scene

git clone https://github.com/kimhoyun-robotair/MarsLab.git
cd MarsLab

Download the MarsLab scene assets, and place the extracted scene directories under assets/scene/. The default configuration expects:

assets/scene/jezero_plain/jezero_plain.usdz

The repository includes rover USD, companion URDF, meshes, and textures. Keep each rover’s supporting asset files together; a USD or URDF alone is not the complete model.

Install Python dependencies

The launcher looks for Isaac Sim at $HOME/isaacsim. If it is installed elsewhere, set the path in the simulator terminal:

export ISAAC_SIM_PATH=/path/to/isaacsim

From the MarsLab repository root, install the simulator dependencies into Isaac Sim’s bundled Python:

marslab/isaac_python.sh -m pip install -r requirements.txt

The ROS companion also needs PyYAML 6+ and Pydantic 2+ available to the Python interpreter used by ros2 launch.

First run

Run the following commands from the MarsLab repository root in each terminal.

Terminal 1: simulator

marslab/isaac_python.sh marslab/main.py \
  --config configs/config.yaml

This starts the default Perseverance profile in Jezero Plain with the GUI, atmosphere controls, and ROS bridge enabled.

Terminal 2: ROS companion

source /opt/ros/jazzy/setup.bash
ros2 launch launch/rover_state_publisher.launch.py \
  config:=configs/config.yaml

The companion publishes the selected rover’s robot description and articulation transforms. Keep both rover_state_publisher.launch.py and its helper companion_urdf.py in launch/.

Both terminals must use the same scenario and rover selection. For example, --robot curiosity in Terminal 1 corresponds to robot:=curiosity in Terminal 2. Without an explicit robot argument, both read the scenario’s robot field.

Check the running system

In a ROS terminal, inspect the topics and a primary image header:

ros2 topic list
ros2 topic echo /rover/rgb/image_raw \
  --qos-reliability best_effort --once --no-arr

Use the same ROS_DOMAIN_ID for the simulator, companion, and viewers. In RViz, set the Image display’s reliability to Best Effort. Use simulation time for ROS applications that consume /clock.

Send a low-speed driving command

Publish continuously while driving; the default command timeout is 0.5 simulated seconds. In a separate ROS terminal, send a forward command:

ros2 topic pub --rate 10 /rover/cmd_vel geometry_msgs/msg/Twist \
  '{linear: {x: 0.2}, angular: {z: 0.0}}'

Stop that publisher with Ctrl+C, then request a controlled stop:

ros2 topic pub --once /rover/cmd_vel geometry_msgs/msg/Twist \
  '{linear: {x: 0.0}, angular: {z: 0.0}}'

The control limits and deceleration ramp still apply. A stale command also requests a controlled stop.

Configuration

Scenario versus rover profile

Edit a scenario for experiment conditions. Edit a rover profile for model-specific geometry, joints, sensor mounts, or reusable hardware settings.

Configuration section What to change
robot Select the profile: perseverance, curiosity, opportunity, or spirit
scene Select the downloaded USDZ environment
runtime Headless mode, ROS bridge, atmosphere, and RTX motion tracking
mars_env Gravity, dust optical depth, solar position, and sun sweep
rendering Renderer, anti-aliasing, sky, light, and fog appearance
rover Override selected profile values for this experiment
wheel_odom Choose whether MarsLab publishes dynamic odometry TF

To keep the supplied scenario intact, create an experiment file beside it:

cp configs/config.yaml configs/my-run.yaml

Relative asset paths resolve from the YAML file that declares them, not the terminal’s working directory. In the supplied layout, scenario paths begin with ../assets/, while profile paths begin with ../../assets/.

Apply a small override

The examples below are partial overrides. Merge them into the existing scenario’s matching sections; do not create duplicate YAML keys or replace the complete scenario with a snippet.

For a slower test with slightly more spawn clearance:

robot: perseverance
rover:
  spawn:
    z_offset: 0.15
  control:
    max_linear_velocity: 0.2

Nested mappings merge recursively. Lists replace the whole existing list, so changing a joint or camera list requires supplying its complete intended contents. Scenario overrides remain active when --robot selects another profile; keep model-specific geometry in the profile.

Launch your experiment with:

marslab/isaac_python.sh marslab/main.py \
  --config configs/my-run.yaml

Use config:=configs/my-run.yaml for the ROS companion as well.

Validate before a full launch

From the repository root, use the same launcher to check schemas and profile composition without starting the simulator:

marslab/isaac_python.sh - <<'PY'
from marslab.config import load_config

config = load_config("configs/my-run.yaml")
print(f"Validated rover: {config.rover.name}")
PY

Unknown keys, invalid types, and incompatible settings are rejected. Full startup additionally checks required assets and runtime constraints. A successful configuration check does not test physics, rendering, or ROS delivery.

For tuning, change one subsystem at a time, restart after YAML edits, and save the scenario, profile, sensor seed, and runtime versions with each experiment. GUI atmosphere edits affect only the current run.

Scenes and rovers

Select an environment

Set scene.usdz_path in your scenario. These paths assume the file remains in configs/:

Scene USDZ path
Jezero Plain — default ../assets/scene/jezero_plain/jezero_plain.usdz
Main Crater ../assets/scene/main_crater/main_crater.usdz
Grand Canyon ../assets/scene/grand_canyon/grand_canyon.usdz
Mars Base ../assets/scene/mars_base/mars_base.usdz
Jezero Plain with the selected rover, captured in MarsLab
Jezero Plain. Open terrain and gently varying relief.
Main Crater with the selected rover, captured in MarsLab
Main Crater. Rock fields and crater-side relief.
Aerial view of the Grand Canyon terrain captured in MarsLab
Grand Canyon. Sloped terrain and irregular geometry.
Aerial view of the Mars Base habitat, antenna, solar panels, and rocket captured in MarsLab
Mars Base. Built structures and visual landmarks.

These views were captured from the four supplied scenes. Any rover profile can be paired with any of these environments; the pairings shown here are for illustration.

The runtime reads meshes and collision data from the scene. A neighboring DEM, generation recipe, or metadata file is not required for loading the USDZ.

Select a rover

marslab/isaac_python.sh marslab/main.py \
  --config configs/my-run.yaml --robot curiosity

Select the matching rover in the ROS companion. For the example above, pass config:=configs/my-run.yaml robot:=curiosity. The simulator’s --robot option overrides the scenario’s selector for that process only; it does not update the companion automatically.

Rover Isaac argument ROS companion argument
Perseverance --robot perseverance robot:=perseverance
Curiosity --robot curiosity robot:=curiosity
Opportunity --robot opportunity robot:=opportunity
Spirit --robot spirit robot:=spirit

Alternatively, set robot in the scenario and omit both command-line robot selectors. One run loads one rover; selecting a profile does not start a multi-rover simulation.

Place the rover

Use rover.spawn.mode to choose a reference:

  • dem_center places the rover at the sampled terrain center.
  • dem_relative applies xy relative to the terrain-derived reference.
  • absolute uses the configured world-space horizontal position.

z_offset adds clearance above the sampled surface. orientation_rpy is in radians. Startup checks the rover’s collider footprint and raises the initial pose when needed for clearance; a location without supporting scene geometry fails startup. Recheck spawn and wheel contact after changing terrain.

Rover profiles

Each profile selects its own USD and companion URDF, mobility geometry, camera calibration, and sensor settings. All four share the six-wheel controller, wheel odometry, ground-truth publisher, and ROS namespace conventions.

Compare the supplied configurations

Profile Cameras Primary image Depth and LiDAR IMU sampling
Perseverance 1 Research camera, 640 × 480 Synthetic depth; 2-D and 3-D LiDAR 30 Hz
Curiosity 17 Left Navcam A, 1024 × 1024 Disabled 8 Hz
Opportunity 9 Left Navcam, 1024 × 1024 Disabled 8 Hz
Spirit 9 Left Navcam, 1024 × 1024 Disabled 8 Hz

Camera counts include the primary image and additional named views. They describe configured image-acquisition models, not every instrument or engineering sensor on a flight rover. Curiosity, Opportunity, and Spirit set both LiDAR entries to null and camera depth_output to false. Their reserved depth and LiDAR topic names do not create those sensor outputs.

The camera profiles retain NASA/JPL-derived calibration and use pinhole rendering. Instrument names identify the rendered viewpoint; they do not imply spectroscopy, detector physics, complete flight telemetry, or flight-validated dynamics.

Perseverance

Perseverance in Main Crater. A camera orbit captured in the running simulator. Compare the chassis, wheel geometry, suspension, and instrument layout from every side.

The default Perseverance-like research platform combines recognizable rover geometry with a deliberately simplified sensor suite for RGB-D, LiDAR, and inertial autonomy experiments. It uses six driven wheels, four steerable corner wheels, and passive rocker-bogie suspension.

  • Profile: configs/robots/perseverance.yaml.
  • Assets: assets/robots/perseverance/m2020_lidar.usda and m2020_lidar.urdf.
  • Primary camera: one chassis-referenced 640 × 480 research camera at the left Navcam mast reference, with 135° horizontal FOV. It does not implement a full Mastcam-Z stereo/zoom system.
  • Research sensors: synthetic depth and XYZ cloud, 2-D and 3-D LiDAR, and a 30 Hz IMU. Added depth noise is initially disabled; the separate noisy IMU stream has nonzero configured noise.
  • Odometry: configured left/right slip terms are 0.02, with wheel-rate noise sigma 0.05. These perturb the estimate rather than implementing wheel–soil interaction.
marslab/isaac_python.sh marslab/main.py \
  --config configs/config.yaml --robot perseverance

Use config:=configs/config.yaml robot:=perseverance in the ROS companion command from First run. The added LiDARs and synthetic depth are research extensions, not replicas of the flight rover’s science payload.

Curiosity

Curiosity in Jezero Plain. A camera orbit captured in the running simulator. Compare the chassis, wheel geometry, suspension, and instrument layout from every side.

The Curiosity model includes six wheel drives, four steering joints, passive rocker-bogie suspension, a two-axis mast, and a five-axis arm. Its authored structure has 22 links and 21 joints; the rocker differential couples two joints, so that count is not the number of independent degrees of freedom. The mast and arm joint envelopes are simulation settings.

  • Profile: configs/robots/curiosity.yaml.
  • Assets: assets/robots/curiosity/curiosity.usda and curiosity.urdf.
  • Primary camera: navcam_left_a, 1024 × 1024 with approximately 45.277° horizontal FOV, attached below Body_MastTilt.
  • Sensing: 17 configured cameras and an 8 Hz IMU; synthetic depth and both LiDARs are disabled.
  • Noise defaults: IMU acceleration/angular noise, wheel-odometry slip terms, and wheel-rate noise are zero. This idealizes those measurement errors; it does not eliminate physical slip or numerical effects.
Camera family Count Configuration and mounting
Navcam 4 Left/right cameras in A/B pairs; 1024 × 1024; mast tilt link
Mastcam 2 Left/right, 1648 × 1200; 34 mm and 100 mm focal lengths; mast tilt link
Hazcam 8 Front/rear, left/right, A/B pairs; 1024 × 1024; chassis
ChemCam RMI 1 chemcam_rmi, 1024 × 1024; mast tilt link
MAHLI 1 mahli, 1648 × 1200; Body_ArmTurret
MARDI 1 mardi, 1648 × 1200; chassis

The primary Navcam uses /rover/rgb/image_raw. The other 16 cameras use named topics, such as /rover/cameras/mastcam_left/image_raw, with their own CameraInfo. Read each camera’s calibration rather than assuming equal intrinsics for the left and right Mastcam views. RMI, MAHLI, and MARDI provide rendered images; the corresponding science operations and detector responses are outside this model.

marslab/isaac_python.sh marslab/main.py \
  --config configs/config.yaml --robot curiosity

Use config:=configs/config.yaml robot:=curiosity for the ROS companion. Keep camera poses relative to their configured mounting links; a mast-mounted or arm-mounted camera must not be treated as a fixed chassis camera.

Opportunity

Opportunity in Main Crater. A camera orbit captured in the running simulator. Compare the chassis, wheel geometry, suspension, and instrument layout from every side.

The Opportunity profile represents the Mars Exploration Rover MER-1 camera configuration on MarsLab’s MER mobility model. It shares the underlying visual heritage and mobility geometry with Spirit, but uses its own camera poses and calibration records.

  • Profile: configs/robots/opportunity.yaml.
  • Assets: assets/robots/opportunity/opportunity.usda and opportunity.urdf.
  • Articulation: 15 rigid links and 14 joints for six wheel drives, four steering carriers, two rockers, and two bogies. The arm, mast, and antenna remain in fixed visualization poses.
  • Primary camera: navcam_left, 1024 × 1024 with approximately 45.333° horizontal FOV.
  • Sensing: nine rendered cameras and an 8 Hz IMU, with no default LiDAR or synthetic depth. The IMU and wheel-odometry noise/slip parameters start at zero.
Camera family Count Topic name or suffix
Navcam 2 Primary left image; additional navcam_right
Pancam 2 pancam_left, pancam_right
Hazcam 4 hazcam_front_left, hazcam_front_right, hazcam_rear_left, hazcam_rear_right
Microscopic Imager 1 microscopic_imager

All nine configured images are 1024 × 1024. The other eight cameras publish below /rover/cameras/<name>/. The Microscopic Imager uses a fixed, geometry-derived stowed-arm mounting approximation; this is not an operating instrument-deployment or contact model.

marslab/isaac_python.sh marslab/main.py \
  --config configs/config.yaml --robot opportunity

Use config:=configs/config.yaml robot:=opportunity for the ROS companion. Read Opportunity’s CameraInfo even when using software developed with Spirit: shared topic names and body geometry do not imply identical calibration.

Spirit

Spirit in Mars Base. A camera orbit captured in the running simulator. Compare the chassis, wheel geometry, suspension, and instrument layout from every side.

The Spirit profile uses the MER-2 camera records with the same MarsLab MER mobility geometry as Opportunity. Both models retain all six wheel drives and the shared rocker-bogie controller.

  • Profile: configs/robots/spirit.yaml.
  • Assets: assets/robots/spirit/spirit.usda and spirit.urdf.
  • Articulation: 15 rigid links and 14 joints. The mast, arm, antenna, and payload keep their fixed visualization poses.
  • Primary camera: navcam_left, 1024 × 1024 with approximately 45.834° horizontal FOV. Its pose and calibration differ from Opportunity’s.
  • Sensing: nine rendered cameras and an 8 Hz IMU; no default LiDAR or synthetic depth, with zero injected IMU and wheel-odometry noise/slip terms.
Camera family Count Topic name or suffix
Navcam 2 Primary left image; additional navcam_right
Pancam 2 pancam_left, pancam_right
Hazcam 4 hazcam_front_left, hazcam_front_right, hazcam_rear_left, hazcam_rear_right
Microscopic Imager 1 microscopic_imager

All configured images are 1024 × 1024. The camera names match Opportunity’s so the same subscriber layout can be reused, while the selected profile supplies the appropriate intrinsics and transforms. The fixed Microscopic Imager is a rendered viewpoint, and the model does not reproduce historical wheel failures or mission degradation.

marslab/isaac_python.sh marslab/main.py \
  --config configs/config.yaml --robot spirit

Use config:=configs/config.yaml robot:=spirit for the ROS companion. The current YAML values, rather than historical mission speed or an older report, determine command limits and sensor settings.

Environment and rendering

Dust, sun, and runtime switches

Parameter Supplied value Tuning purpose
mars_env.gravity 3.72 m/s² Mars gravity; the schema accepts 3.6–3.85 m/s²
mars_env.dust_optical_depth 0.05 Optical depth τ; YAML accepts 0 < τ ≤ 6, GUI covers 0.05–6
mars_env.sun_azimuth_deg 90.07° Solar direction around the scene
mars_env.sun_elevation_deg 3.47° Solar height; allowed range 0–90°
mars_env.dynamic_atmosphere.enabled false Start with a fixed manual sun or an automatic sweep
runtime.headless false Run without the viewer and atmosphere GUI
runtime.ros2_enabled true Enable ROS topics and command input
runtime.atmosphere_enabled true Enable the Mars sky, sun, fog, and atmosphere controls

For a controlled dust comparison, keep the sun and camera pose fixed, then vary only τ. For example:

mars_env:
  dust_optical_depth: 0.5
  sun_azimuth_deg: 180.0
  sun_elevation_deg: 45.0
  dynamic_atmosphere:
    enabled: false
Perseverance and Main Crater at dust optical depth 0.05
τ = 0.05. Strong directional light and distinct shadows.
Perseverance and Main Crater at dust optical depth 0.5
τ = 0.5. Reduced direct light and a changed sky contribution.
Perseverance and Main Crater at dust optical depth 2.0
τ = 2.0. A more diffuse appearance as direct sunlight weakens.
Perseverance and Main Crater at dust optical depth 6.0
τ = 6.0. The high-dust endpoint, with much weaker direct illumination.

Actual renders with a fixed camera, fixed exposure, and the same sun position (135° azimuth, 35° elevation). Only optical depth changes. This demonstrates the renderer’s response, not a calibrated visibility measurement.

Auto Sweep starts from the latest sun angles. time_scale controls its speed, and sun_sweep.max_elevation_deg controls the intended peak. This is an interactive illumination sweep, not an astronomical day/night model. Headless mode keeps the configured simulation behavior but does not create the GUI panel.

Watch the shadows move. A controlled sweep of solar azimuth (85–275°) and elevation (12–48°), with τ = 0.25 and fixed exposure. This short loop visualizes parameter changes; it is not a real-time Martian day.

Renderer and image quality

rendering.mode accepts ray_tracing for interactive work or path_tracing for more expensive visual studies. This choice also applies when the atmosphere is disabled.

rendering:
  mode: ray_tracing
  ray_tracing:
    antialiasing_op: 4  # 4 = DLAA; 3 = DLSS
    dlss_exec_mode: 1   # DLSS only: Balanced

DLAA is the supplied default. It uses native-resolution anti-aliasing and was selected to address observed DLSS viewport-edge flicker. DLSS can reduce rendering cost, with results depending on the scene and GPU. dlss_exec_mode changes the DLSS quality mode; it does not select DLAA. For path tracing, tune spp, total_spp, and max_bounces together with runtime cost.

What the atmosphere model represents

Direct normal irradiance follows I0exp⁡(−τ/cos⁡z)I_0\exp(-\tau/\cos z) above the horizon, where zz is the solar zenith angle. It is zero at and below the horizon. The secant approximation is supported up to about 80° zenith angle; the runtime extends it toward the horizon for continuous rendering without claiming the same accuracy there.

Diffuse light uses a τ-indexed COMIMART reduction at fixed reference conditions. The sky-dome gain, HDRI textures, sun gain, and fog remain renderer controls. Simple fog approximates appearance rather than multiple scattering, and rendered light intensities are not calibrated measurements in W/m².

The dust controls affect scene radiance and appearance. They do not model aerosol-induced attenuation of LiDAR returns. Hold exposure, renderer, textures, and light gains constant when comparing dust conditions. The repository’s assets/atmosphere/README.md records the model assumptions and reference comparisons.

Rover tuning

From command to motion. Perseverance travels approximately 5 m during a 12-second simulation at a commanded 0.6 m/s. The footage follows the actual physics and MarsLab controller, including acceleration from rest; playback uses simulation time.

Geometry and profile-dependent settings

The controller and wheel odometry use the selected profile’s geometry. These are simulation model dimensions, with derived values rounded for readability:

Profile Wheel radius Corner wheelbase Maximum steer angle
Perseverance 0.2667 m 2.2600 m 0.9 rad
Curiosity 0.2500 m 2.2600 m 1.2 rad
Opportunity 0.1300 m 1.2120 m 1.2 rad
Spirit 0.1300 m 1.2120 m 1.2 rad

All four current profiles set max_linear_velocity: 4.0, max_angular_velocity: 0.7, and command_timeout: 0.5. These are simulation command limits. Perseverance uses negate_steer: true, while the other three use false to match their joint conventions. Opportunity and Spirit use steer_stiffness: 200000.0; Perseverance and Curiosity use 50000.0. Keep these asset-dependent settings in the profile when switching rovers.

Physics and contact

The rover USD provides geometry, collision shapes, mass, inertia, and center of mass. YAML controls damping, wheel contact materials, suspension damping, and joint drives. Mass and inertia override keys are not supported; the authored mass properties include geometry-based approximations rather than measured flight-rover distributions.

Tune in this order: confirm links and joints, check the asset’s mass properties, adjust wheel contact, then suspension damping, then drive and steering response.

Parameter under rover Perseverance value Effect
linear_damping / angular_damping 0.5 / 12.0 Chassis motion damping
wheels.friction_static / wheels.friction_dynamic 0.6 / 0.5 Wheel contact friction; static must be at least dynamic
wheels.restitution 0.0 Wheel contact bounce
suspension.rocker_damping / suspension.bogie_damping 75.0 / 50.0 Passive suspension response
control.drive_type acceleration PhysX drive response; switching to force requires retuning
control.max_linear_velocity 4.0 m/s Simulation command cap, not a claimed flight-rover speed
control.max_angular_velocity 0.7 rad/s Angular command limit
control.max_steer_angle 0.9 rad Steering travel, checked against geometry and USD limits
control.command_timeout 0.5 simulated s Controlled stop after stale commands
control.brake_stiffness 50000.0 Parking hold after wheel-rate ramps reach zero; 0 disables it

The passive rocker differential couples the two differential joints. Drive and steering lists must match the USD exactly: six drive joints and four steering joints in the documented order. Wheel radius, wheelbase, steering-axle offset, and track geometry are shared by control and wheel odometry.

Driving and steering

The controller supports curved driving and pivot turns. A zero linear command with nonzero yaw rate selects a pivot turn. For a substantial steering change, the controller brakes, waits for low measured wheel motion, aligns steering, and then resumes driving. Tune max_wheel_accel_rate, steer_ramp_rate, decel_multiplier, and steering_alignment_tolerance after the geometry and joint limits are correct.

A zero Twist requests braking while retaining steering orientation. Pause preserves control and simulation state; Stop resets the episode, including command state, odometry, and sensor timing. ROS consumers must handle simulation time restarting when a new episode begins.

Sensor tuning

Actual 640 by 480 image from Perseverance’s primary simulated camera, including foreground rover geometry
RGB. The native primary camera’s 135° field of view, captured at 640 × 480.
The same camera’s simulated optical-axis depth, shown in a purple-to-yellow distance colormap
Synthetic depth. Optical-axis distance from the same settled camera view. Added depth noise is disabled.
Logarithmic depth color scale: 0.3, 1, 5, 20, 100, and 200 meters, from dark purple to light yellow
Depth spans 0.3–200 m on a logarithmic color scale; pixels without a valid depth are dark.
Three-dimensional scatter plot of 10,035 actual simulated LiDAR returns, with X, Y, Z axes in meters and distance coloringRTX LiDAR · one acquired frame
Geometry beyond the image. A visualization of 10,035 points read from the RTX LiDAR annotator. Colors show distance from the raw point-coordinate origin; the cloud is one acquired frame, not an accumulated SLAM map. Curiosity, Opportunity, and Spirit do not enable this sensor by default.

Acquisition settings by rover

Profile Primary camera IMU Range sensing by default
Perseverance 640 × 480; 135° horizontal FOV; 0.3–200 m clip 30 Hz; nonzero noise on the noisy stream Synthetic depth, 2-D LiDAR, 3-D LiDAR
Curiosity 1024 × 1024; 45.277° horizontal FOV; 0.03–200 m clip 8 Hz; zero injected noise Disabled
Opportunity 1024 × 1024; 45.333° horizontal FOV; 0.06–200 m clip 8 Hz; zero injected noise Disabled
Spirit 1024 × 1024; 45.834° horizontal FOV; 0.06–200 m clip 8 Hz; zero injected noise Disabled

Additional camera settings differ within each mission profile; see the rover profiles for camera families and mounts. The three 8 Hz settings describe the simulator’s configured sampling, not a guarantee of observed ROS delivery or flight telemetry cadence.

Perseverance’s 2-D LiDAR covers a forward 135° sector at nominal 10 Hz, over 0.2–200 m. It acquires a full turn internally and crops the published scan. Its 3-D LiDAR is configured for 220° horizontal / 45° vertical FOV at 30 Hz, over 3.5–200 m. The 3.5 m near limit avoids the rover’s own geometry. Tune range_min, range_max, rotation_rate_hz, FOV, and pose together; these are configured limits rather than guaranteed usable detection ranges.

Cameras and depth

The primary rover.sensors.camera supplies RGB and CameraInfo; depth and XYZ clouds are available only when depth_output is enabled. Additional named cameras live in rover.sensors.cameras. An explicit horizontal_fov_deg overrides the focal-length setting. Depth represents distance along the optical axis rather than radial distance from the camera.

Set depth_output: false to disable synthetic depth and cloud outputs. This differs from depth_sensor.enabled, which enables added depth noise. The supplied Perseverance profile outputs depth with that noise disabled. Curiosity, Opportunity, and Spirit disable both synthetic depth output and added depth noise.

For a research experiment with seeded synthetic depth, merge this into the scenario. On Curiosity, Opportunity, or Spirit, this explicitly adds a simulated depth modality that their supplied profiles do not provide:

rover:
  sensors:
    seed: 42
    camera:
      depth_output: true
      depth_sensor:
        enabled: true
        noise_mean: 0.0
        noise_sigma: 0.005

The depth image and noisy organized XYZ cloud share the same sampled noise and acquisition timestamp. Additional cameras publish RGB and CameraInfo under /rover/cameras/<name>/; that does not imply additional flight depth instruments.

Extra camera workload

Curiosity adds 16 named camera views to the primary camera; Opportunity and Spirit add eight each. To isolate the primary camera while debugging, use:

rover:
  sensors:
    cameras: []

This replaces the entire additional-camera list for the run while retaining the primary camera. Remove the override to restore the profile’s full camera set. Camera count and resolution affect rendering and publication workload; check actual rates on your hardware.

Mounts, LiDAR, and reproducibility

Sensor poses use metres and local_orientation_rpy_deg in degrees, unlike the rover spawn orientation in radians. When a camera uses an existing USD camera prim, update its authored pose and matching YAML together; startup rejects a mismatch.

An integer sensor seed derives separate noise streams for wheel odometry, noisy IMU, and optional depth noise. null selects nondeterministic noise. A fixed seed does not guarantee deterministic GPU physics or rendering, and depth noise also depends on acquisition timestamps.

Curiosity, Opportunity, and Spirit already set lidar_2d: null and lidar_3d: null; no LiDAR publisher is expected for these supplied profiles. For 3-D LiDAR, use an installed profile_name and supported variant; keep legacy profile_json_path and usd_profile null. Keep runtime.enable_motion_bvh enabled for moving RTX geometry; this does not deskew the resulting scan.

ROS 2 integration

Topics

All four supplied profiles use the rover namespace and the same primary RGB, command, IMU, odometry, ground-truth, and joint-state names. Available image and range topics depend on the selected profile.

Interface Topic
Velocity command, geometry_msgs/Twist /rover/cmd_vel
Primary RGB / calibration /rover/rgb/image_raw · /rover/rgb/camera_info
Depth / XYZ cloud — Perseverance by default /rover/depth/image_raw · /rover/depth/points
Raw / noisy IMU /rover/imu · /rover/imu_noisy
2-D / 3-D LiDAR — Perseverance by default /rover/scan · /rover/lidar/points
Wheel odometry /rover/odom
Ground truth /rover/GT_Trajectory
Joint states / robot description /rover/joint_states · /rover/robot_description
Simulation time /clock

For Curiosity, the primary navcam_left_a is exposed through /rover/rgb/image_raw, and an additional Mastcam image is /rover/cameras/mastcam_left/image_raw. For Opportunity and Spirit, the primary is navcam_left, while the right view is /rover/cameras/navcam_right/image_raw. Each image has a matching camera_info topic; the primary is not duplicated under the named-camera path.

Set rover.ros2.namespace without surrounding slashes. The entries in rover.ros2.topics are relative topic names, also without leading or trailing slashes.

TF ownership and wheel odometry

Publisher Responsibility
ROS companion base_link to the selected URDF root, articulation TF, and robot description
MarsLab sensor bridge Fixed sensor transforms relative to their configured mounting links; normally Body_Chassis, with camera-specific parents where configured
MarsLab wheel odometry Wheel estimate and optional dynamic odom to base_link TF
Ground-truth publisher Evaluation pose in map / base_link_gt; no TF publication

Curiosity cameras attached to Body_MastTilt or Body_ArmTurret follow the corresponding articulated mounting link. Their fixed mounting transform must not be reparented to Body_Chassis. Opportunity and Spirit use fixed chassis-relative camera poses.

Wheel odometry uses six wheel encoders and four measured steering angles, with geometry shared by the controller. It applies configured slip and angular-rate noise and integrates planar motion; it does not read ground truth.

If an external localization system owns dynamic odom to base_link TF, set:

wheel_odom:
  publish_tf: false

This disables MarsLab’s dynamic odometry TF, not the odometry topic itself. rover.wheel_odometry.enabled controls the estimate. If you disable the estimate, also disable its TF publication.

QoS and timing

The supplied camera publishers use best-effort delivery. A subscriber requesting reliable delivery will not match them. Use compatible reliability and durability settings for each consumer; keep latched/static data transient-local where appropriate.

The configured IMU frequency and LiDAR rotation rates are not the same as measured end-to-end ROS delivery rates. Inspect live topics and timestamps. The companion’s publish_frequency affects robot-state publication, not Isaac physics or sensor sampling.

Evaluation

The paper evaluates SLAM across sensing modalities, dust conditions, scene geometry, and route lengths. It also evaluates visual place recognition on repeated Mars Base traversals under changes in illumination and dust.

Daytime query image of a Mars Base structure
Daytime query image.
Retrieved image of the same Mars Base location under dark conditions
Retrieved location under dark conditions.

For a repeatable comparison, record the scene, rover profile and overrides, sun angles, dust level, renderer settings, sensor seed, and software versions. Connect the algorithm through ROS, record its estimated trajectory alongside ground truth, and use a consistent alignment and evaluation protocol. A dust sweep isolates visual appearance changes in the current model; it does not test dust-induced LiDAR degradation.

The runtime supplies sensor data and reference motion. The paper’s trajectory followers and full benchmarking pipeline are not bundled with it. See the paper for experimental protocols and quantitative results.

Troubleshooting

Symptom What to check
Launcher cannot find Isaac Python Check ISAAC_SIM_PATH; the default installation directory is $HOME/isaacsim.
Startup reports a missing scene or rover asset Check the resolved path and supporting files. Relative paths belong to the declaring YAML.
Configuration fails before the viewer opens Check spelling, types, units, list completeness, and matching rover names. Unsupported keys are rejected.
Companion cannot import configuration dependencies Check that the Python used by ros2 launch has PyYAML 6+ and Pydantic 2+ installed.
No depth or LiDAR topics after switching rover Curiosity, Opportunity, and Spirit disable these outputs by default. Check the selected sensor configuration rather than assuming a bridge failure.
Image topic exists but RViz is blank Set Image reliability to Best Effort and check ROS_DOMAIN_ID and DDS discovery.
Rover stops shortly after a command Publish Twist continuously; the default timeout is 0.5 simulated seconds.
Rover spawns incorrectly after changing scenes Recheck spawn mode, xy, surface support, and positive z_offset.
TF conflicts or an inconsistent robot model Match scenario and rover in both terminals; ensure one owner of dynamic odometry TF.
Flicker at viewport edges Check antialiasing_op: 4 selects DLAA; assess GPU cost with your scene.
Consumers fail after Stop / Play Handle the new episode and simulation timestamps moving backwards.

YAML changes require a restart. Configuration checks do not start the simulator; final contact behavior, rendering, sensor outputs, ROS compatibility, and resets must be checked in the running simulator.

About the captures on this page

The rover loops, four terrain views, dust comparison, driving clip, illumination sweep, and sensor visualizations were acquired from a local MarsLab run on 19 November 2026, using Isaac Sim 5.1 and an RTX 5090. Rover orbit videos move an observation camera while physics remains active. The driving clip uses MarsLab’s controller without animating the rover pose.

The observation camera renders at 1280 × 720 with a 28 mm lens and fixed exposure (ISO 100, f/5, 1/500 s). Unless varied explicitly, τ is 0.25, sun azimuth is 135°, and elevation is 35°. The Canyon and Mars Base landscape stills use a higher overview camera and 55° solar elevation. Additional profile cameras and ROS publication were disabled for these captures; the native primary camera remains active. These presentation settings do not replace the supplied rover defaults.

The depth image is colorized simulator output, and the point-cloud image is a plot of acquired RTX LiDAR points. Each rover profile retains its individual orbit capture. The day/night evaluation pair is retained from the project’s existing research materials.

Capture record · Media and asset credits

Roadmap

The current runtime supports four rover profiles, 2-D and 3-D LiDAR, configurable sensing, a ROS 2 bridge, and independent ground-truth and wheel-odometry streams.

Planned work includes photorealistic scene generation and customizable rock distribution, aerial robots such as Ingenuity, simultaneous multi-robot simulation, richer sensor bias/drift models, improved sky and light rendering, more detailed wheel–soil interaction, integrated evaluation workflows, and containerized deployment. Scene-generation v2.0 is under development and is not part of the current distribution.

This guide was prepared from the MarsLab README, Configuration and Tuning Guide, project documentation, rover technical reports, and current configuration profiles and schemas. For source files and updates, see the MarsLab repository. The repository’s README.md and CONFIGURATION_GUIDE.md are the detailed setup and parameter references.

Citation

@inproceedings{kim2026marslab,
  title = {MarsLab: A Martian Rover Simulator for
           Planetary Rover Autonomous Navigation},
  author = {Kim, Hoyun and Kim, Beomsu and Kim, Giseop},
  booktitle = {International Conference on Space Robotics (iSpaRo)},
  year = {2026},
  eprint = {2609.34702},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2609.34702}
}