cs.ROSep 28, 2026
SaveMarsLab: A Martian Rover Simulator for Planetary Rover Autonomous Navigation
Organizations: Department of Robotics and Mechatronics Engineering, DGIST, Daegu, Republic of Korea
Abstract
Future Mars missions will require rover autonomy that can operate across unstructured terrain, changing illumination, atmospheric dust, and limited communication. Simulation is a practical way to study these conditions before deployment, but existing Mars-relevant resources differ in scope, including mission-oriented simulators, fixed analog datasets, task-specific environments, and open robotics interfaces. In this context, we present MarsLab, an open-source, ROS2-native Mars rover simulator for autonomy and navigation algorithm development. MarsLab combines HiRISE-derived and procedural terrain with customizable rock, crater, solar-illumination, and atmospheric-dust settings, and runs a Perseverance-class rover model in NVIDIA Isaac Sim. The runtime publishes RGB, depth, RGB-D point clouds, LiDAR, IMU, wheel odometry, and Ground Truth (GT) pose data through standard ROS2 topics. We demonstrate MarsLab with Simultaneous Localization and Mapping (SLAM) benchmarks across sensing modalities, dust levels, scene geometry, and route length, and with Visual Place Recognition (VPR) benchmarks over repeated Mars Base traversals under illumination and dust changes. The results illustrate how controlled scene variation and shared GT trajectories can be used to compare trajectory-level estimation and image-level place recognition within the same simulator. Our Project Page: https://kimhoyun-robotair.github.io/MarsLab/.
Figures & tables
| Simulator | Year | Scene gen. | Engine | Goal | Scenes/Seqs. | Rock | Dust | Light | Sens. | ROS2 | Open |
| ROAMS [ 9 ] | 2004 | Mars-like terrain & mission simulation | Custom / JPL DARTS-DSHELL | Rover mobility, dynamics, onboard SW V&V | Mission/testbed scenarios; Monte Carlo possible | ✓ | ✗ | Stereo cameras; IMU; encoders; sun sensors | ✗ | ||
| ENav [ 10 ] | 2020 | Mars 2020 ENav test terrains | ROS-based + HDSim/RSVP | Autonomous nav V&V | Mars 2020 test scenarios / Monte Carlo runs | ✓ | ✗ | Stereo/NavCam point cloud; camera rendering | ✗ | ✗ | |
| Giubilato et al. [ 11 ] | 2020 | Simulated Martian / planetary-like environment | ROS/Gazebo | Visual & LiDAR SLAM evaluation | Long/Short simulated rover sequences / datasets | ✓ | ✗ | Cam / 3D L; stereo + LiDAR seqs. | ✗ | ✓ | |
| MarsSim [ 12 ] | 2023 | Multiscale Mars terrain simulation | ROS/Gazebo | High-fidelity physical & visual rover simulation | Pahrump Hills-like and generated scenes; released seqs. NR | ✓ | ✓ | RGB/visual; locomotion data; D/L NR | ✗ | ✗ | |
| ISMRS [ 13 ] | 2024 | Digital-twin and asset-based scenes | Isaac Sim | Mars rover simulation + ML data synthesis | Real-world recordings + synthetic data | ✓ | ✗ | ✓ | RGB/stereo; 2D LiDAR; IMU; wheel odom; seg. | ✓ | ✗ |
| RLRoverLab [ 14 ] | 2024 | Synthetic rover training scenes | Isaac Lab | RL navigation / manipulation / control | RL task scenes | ✗ | Task-dependent; RGB cam, height scan, Isaac obs. | ✗ | ✓ |
TABLE I : Concise survey of Mars rover simulation platforms and benchmark support. Column abbreviations and symbols: gen. = generation; seqs. = sequences; sens. = sensor modalities; ✓ = reported/core capability; = partial, indirect, or not primary; ✗ = not reported; D = depth; L = 3D LiDAR; seg. = semantic or instance labels; NR = not reported.
| Jezero plain | Main Crater | Mars Base | Grand Canyon |
| Jezero rim | Crater interior | Base assets | Canyon wall |
| (a) | (b) | (c) | (d) |
Fig. 3 : Reference MarsLab scenes. MarsLab can generate additional terrain instances, and these four reference scene pairs are provided as scenes for evaluation. The green rectangle in each upper image indicates the approximate area represented by the corresponding lower scene: (a) a plain scene from Jezero crater, (b) a main-crater scene from Arcadia Planitia, (c) a Mars Base scene that augments the Jezero terrain with a habitat asset, and (d) a Grand Canyon scene from Elysium Planitia. All scenes are built on HiRISE-derived terrain, providing diverse geometry, texture, and structure for SLAM and place recognition stress testing.
| Fig. | Reference scene | Scene size | Rock | Dust | Illumination | Target Use |
| (a) | Jezero plain/rim | km | ✓ | ✓ | ✓ | Navigation, terrain following, and low-texture localization baselines. |
| (b) | Main Crater | km | ✓ | ✓ | ✓ | Repetitive-terrain failure analysis and rocky slope traversal. |
| (c) | Mars Base | km | ✓ | ✓ | ✓ | Landmark- and feature-rich localization and VPR benchmarking. |
| (d) | Grand Canyon | km | ✓ | ✓ | ✓ | Long-range drift evaluation over extended canyon terrain with strong elevation changes and complex geometry. |
TABLE II : MarsLab reference scenes. Compact summary of the four evaluation scenes in Fig. 3 . Rock distribution, illumination, and atmospheric dust denote user-customizable scene parameters, including rock size–frequency/spatial-density settings, solar azimuth/elevation, and dust optical depth.
| Stage | User-provided input | Customizable variation |
| Terrain base | HiRISE DEM/orthomosaic | Crop region, terrain scale, and generation seed. |
| Surface clutter | Rock assets and rock size–frequency parameters | Rock density, diameter range, spatial sampling, and placement seed. |
| Crater field | Mars crater production-function parameters | Crater density, diameter range, morphology, and resurfacing masks. |
| Illumination | Solar-position and irradiance configuration | Sun azimuth, elevation, shadow direction, and scene shading. |
| Atmosphere | Mars dust optical-depth setting | Dust optical depth , sky brightness, and image contrast. |
TABLE III : Customizable MarsLab scene generation. Beyond the reference scenes, MarsLab lets users build HiRISE-based scenes by controlling terrain extent, surface clutter, crater fields, illumination, and atmospheric dust. The table summarizes the scene-generation factors exposed as user-controlled variables.
Fig. 4 : Procedurally generated rock and crater scenes. Rendered MarsLab terrain showing (a) the rock size–frequency distribution ( Eq. 1 ) and (b) crater density and morphometry ( Eq. 2 ).
Fig. 5 : Controllable solar illumination. The sun azimuth and elevation are explicit experimental variables; changing the solar position redistributes cast shadows and shading across the same terrain, directly affecting feature visibility for vision-based navigation algorithms.
Fig. 6 : Atmospheric-dust sweep. MarsLab uses the optical-depth parameter in Eq. 3 as a controlled evaluation condition for vision-based algorithms: increasing brightens the sky and ground, lowers image contrast, and reduces visibility while leaving the terrain geometry unchanged.
| ORB-SLAM (RGB) | RTAB-Map (RGB-D+Odom) | MOLA (LiDAR) | |||
| nominal | |||||
| Mars Base | |||||
| Main Crater Interior | ✗ tracking lost | ||||
| Grand Canyon | ✗ tracking lost | ✗ tracking lost | |||
Fig. 8 : Multi-scene SLAM benchmark: estimated vs. ground-truth trajectories. Top-down view of every run—the estimated trajectory (red) versus ground truth (black) after Umeyama alignment, over the reconstructed point-cloud map (coloured by height); rows are scenes, columns are sensing modality and dust optical depth. Grey tiles ( ✗ ) mark monocular ORB-SLAM runs that lost tracking and produced no usable trajectory. Scene names and GT trajectory lengths are listed in Table V ; GT trajectory geometry is shown in Fig. 10 ; per-run ATE in Table IV ; reconstructed 3D maps in Fig. 9 .
| ORB-SLAM | RTAB-Map | MOLA | |||
| Scene | LiDAR | ||||
| Mars Base | 0.59 | 4.70 | 0.48 | 0.64 | 0.20 |
| Main Crater Interior | 40.28 | ✗ | 8.84 | 10.61 | 0.13 |
| Grand Canyon | ✗ | ✗ | 12.28 | 19.08 | 0.38 |
TABLE IV : ATE RMSE (m) across MarsLab scenes and SLAM stacks. Camera methods are run at low ( ) and high ( ) dust; LiDAR is run once per scene. Per-row best in bold ; ✗ marks a tracking failure.
Fig. 10 : GT trajectories. Compact overview of the three indexed ground-truth trajectories used in the SLAM benchmark.
| Scene | Length | GT Trajectory Role |
| Mars Base | m | Feature- and landmark-rich trajectory for testing localization in structured habitat scenes. |
| Main Crater Interior | m | Rock-rich trajectory for testing localization under repetitive terrain and sparse visual features. |
| Grand Canyon | m | Long-distance trajectory for testing drift across extended canyon terrain with large elevation changes. |
TABLE V : SLAM benchmark GT trajectory metadata. Each GT trajectory is authored in a MarsLab scene, replayed by the rover runtime, and logged from Isaac Sim pose streams in TUM format for ATE/RPE evaluation.
| Query | Retrieved (R@1) | Query | Retrieved (R@1) |
| (i) Day/Dark, | (ii) Dust, to | ||
Fig. 11 : Visual place recognition: query vs. retrieved. Qualitative examples for the two Table VI protocols: (i) day-lit queries matched to a dark database at and (ii) low-dust daytime queries at matched to high-dust daytime database images at . Each pair shows the query frame and the top-1 retrieved database frame (R@1).
| Method | Day/Dark, | Dust, | ||||
| R@1 | R@5 | R@10 | R@1 | R@5 | R@10 | |
| NetVLAD [ 33 ] | 77.90 | 95.35 | 98.36 | 85.02 | 98.18 | 99.33 |
| AnyLoc [ 34 ] | 64.04 | 92.21 | 97.17 | 74.89 | 95.71 | 99.06 |
| BoQ [ 35 ] | 91.44 | 99.19 | 99.80 | 94.65 | 99.37 | 99.78 |
TABLE VI : Visual place recognition recall on the MarsLab Mars-base benchmark using the strict m positive radius. Results are Recall@ (%). Best values in each protocol/metric are shown in bold.