Off-road autonomous navigation demands reliable 3D perception for robust obstacle detection in challenging unstructured terrain. While LiDAR is accurate, it is costly and power-intensive. Monocular depth estimation using foundation models offers a lightweight alternative, but its integration into outdoor navigation stacks remains underexplored. We present an open-source off-road navigation stack supporting both LiDAR and monocular 3D perception without task-specific training. For the monocular setup, we combine zero-shot depth prediction (Depth Anything V2) with metric depth rescaling using sparse SLAM measurements (VINS-Mono). Two key enhancements improve robustness: edge-masking to reduce obstacle hallucination and temporal smoothing to mitigate the impact of SLAM instability. The resulting point cloud is used to generate a robot-centric 2.5D elevation map for costmap-based planning. Evaluated in photorealistic simulations (Isaac Sim) and real-world unstructured environments, the monocular configuration matches high-resolution LiDAR performance in most scenarios, demonstrating that foundation-model-based monocular depth estimation is a viable LiDAR alternative for robust off-road navigation. By open-sourcing the navigation stack and the simulation environment, we provide a complete pipeline for off-road navigation as well as a reproducible benchmark. Code available at https://github.com/LARIAD/Offroad-Nav.
Figures & tables
Fig. 1 : Autonomous navigation with the wheeled ground robot in the real and simulation off-road environment.
Fig. 2 : Diagram of our navigation pipeline. It takes as input either a LiDAR point could or monocular camera images with a SLAM sparse point cloud as well as the robot localization. It returns the velocity commands run by the robot.
Fig. 3 : Estimated rescaled depth maps and the extracted 3 D point clouds from the top RGB image. The left side shows results without edge filtering, while the right side shows results with edge filtering applied.
Fig. 4 : Top views of the simulation experiments. From left to right, the easy, medium and hard simulated environments, respectively. In each environment, the goal points of the 10 , 20 and 30 meters scenarios is indicated with an orange, blue and green star, respectively. The reference remotely operated trajectories are given in dashed lines and the ones performed by the robot are shown in solid lines. From top to bottom: the robot trajectories are obtained using the LiDAR (sim-tuned), the LiDAR (real-params) and Mono-VINS (STCD only), respectively.
Fig. 5 : Top view for the real-world experimental area where obstacles are placed on an off-road terrain. The first row shows the LiDAR experiments and the second one, the runs with the monocular camera. From left to right: the easy, medium and hard scenario. Obstacles are represented in white circles and rectangles. The dashed lines correspond to the remotely operated reference trajectories.
Easy
Medium
Hard
Average
SR ↑
SPL ↑
DR ↑
SR ↑
SPL ↑
DR ↑
SR ↑
SPL ↑
DR ↑
SR ↑
SPL ↑
DR ↑
10m goal
LiDAR (sim-tuned)
1.00
0.70
1.00
1.00
0.73
1.00
0.40
0.31
0.76
0.80
0.58
0.92
LiDAR (real-params)
0.90
0.66
0.95
0.80
0.47
0.93
0.40
0.29
0.72
0.70
0.47
0.87
Mono-VINS (STCD only)
1.00
0.73
1.00
0.80
0.59
0.89
0.50
0.39
0.78
0.77
0.57
0.89
20m goal
TABLE I : Navigation performance across simulated scenarios and goal distances.
Easy
Medium
Hard
Average
SR ↑
SPL ↑
DR ↑
SR ↑
SPL ↑
DR ↑
SR ↑
SPL ↑
DR ↑
SR ↑
SPL ↑
DR ↑
LiDAR (real-params)
1.00
0.73
1.00
1.00
0.68
1.00
1.00
0.67
1.00
1.00
0.69
1.00
Mono-VINS
1.00
0.51
1.00
1.00
0.63
1.00
1.00
0.63
1.00
1.00
0.59
1.00
TABLE II : Navigation performance across real scenarios with the Barakuda robot.
Module
LiDAR (Hz)
Mono (Hz)
3D perception (CPU/GPU)
20
10
Elevation mapping (GPU)
16
6
Path planning (CPU)
12
12
TABLE III : Frequency of each module with the Jetson ORIN (MAXN profile) for the LiDAR-based and Monocular-based configurations
Edge masking
Smoothing
Easy
Medium
Hard
Avg.
✗
✗
0.50
0.48
0.19
0.39
✓
0.54
0.61
0.30
0.48
✓
✗
0.63
0.74
0.18
0.52
✓
0.54
0.76
0.45
0.58
TABLE IV : Average SPL performances with different configurations for the depth rescaling
Interdisciplinary Program in Artificial Intelligence, Seoul National University, Seoul, 08826, Korea. · Department of Electrical and Computer Engineering, Seoul National University, Seoul, 08826, Korea.