Trailblazer: Learning Off-Road Costmaps for Long-Range Planning

Autonomous navigation in off-road environments is a challenging task, especially for Unmanned Ground Vehicles (UGVs) involved in search and rescue, exploration, and surveillance. Trailblazer is a novel framework that automates the generation of costmaps from multi-modal sensor data, such as satellite imagery, LiDAR, DEMs and OpenStreetMaps, to enable efficient and scalable long-range path planning without manual tuning. By leveraging imitation learning and a differentiable A* planner, Trailblazer adapts to diverse terrains and dynamic environments, offering a robust solution for global planning tasks.

---

Fig. 1: Inputs to Trailblazer from the Texas A&M RELLIS test site. Satellite imagery and LiDAR data are used to generate semantic, height, slope, and intensity maps.

Fig. 2: Trailblazer framework architecture. The encoder-decoder generates costmaps, while the Neural A planner computes paths from the costmaps.*

Fig. 3: Comparison of costmaps generated using LiDAR (4 points/m²) and DEM (1/3 arc-second resolution) data.

For more details, check out the Paper

Getting Started

Before You Begin

• We are providing ROS 1 package. The repository is tested with ROS Noetic.
• All maps and waypoints should be stored in the assets folder.
• Launch files for the RELLIS test site are included.
• The input is preprocessed and saved. The package currently runs the Trailblazer algorithm and generates waypoints in UTM format.
• A requirements.txt file is provided for environment setup.

Manual Installation

Some dependencies need to be installed manually to ensure the proper functioning of the Trailblazer framework. These include:

1. mmcv
2. mmseg

Installation Instructions:

• mmcv: Follow the mmcv GitHub page for detailed installation steps.
• mmseg: Refer to the mmsegmentation documentation for installation instructions.

---

Asset Download

The assets required for RELLIS test site are too large to be uploaded to GitHub. You can download them from the following link:
Download Assets

After Downloading:

• Place the .tif file in the assets/images/rellis folder.
• Place the best.pth file in the scripts/weights/segformer folder.

You can find LiDAR data for your area of interest from: USGS LiDAR Explorer, UK DEFRA Survey data.

Satellite imagery can be downloaded from ARCGIS Imagery servers through QGIS.

OSM data can be downloaded from OpenStreetMaps

---

Running the Package

To run the Trailblazer package, use the following command:

roslaunch camel rellis.launch

Please run roscore and source devel/setup.bash before running the launch file.

Launch file parameters

The launch file parameters are as follows:

1. waypoints_file: Path to the waypoints file. "None" to generate random waypoints.
2. site_name: Name of the test site; ex: cavasos, irwin, camp roberts.
3. cmap_resolution: Costmap resolution in meters.'
4. epsg_code: epsg_code used by the pointcloud. If using PDAl library, use "None".
5. UTM_zone: UTM zone of the region.
6. gps_origin: GPS coordinate in UTM format at origin.
7. image_input: Path to the overhead image.
8. pointcloud_input: Path to the overhead pointcloud/DEM.
9. osm_file: Path to OSM file input.

One can change the cmap_resolution parameter to generate the costmap at different resolutions.

Scripts

The scripts folder contains the following scripts:

1. camel_pathV2.py: The main script that generates the waypoints.
2. cmap_gen.py: Generates the costmap from the input.
3. dijkstra.py: Implements the Dijkstra algorithm.
4. input_gen.py: Generates the input from the overhead image and pointcloud.
5. semantic_seg.py: Implements the semantic segmentation.
6. utils/laz_reader.py: Reads the pointcloud data.
7. utils/dem_reader.py: Reads Digital Elevation Maps and converts them to pointcloud.

Future Updates

The current version supports the RELLIS test site. Additional examples will be added soon.

Future updates will include support for more test sites and enhanced functionality.

Citation

@misc{viswanath2025trailblazerlearningoffroadcostmaps,
      title={Trailblazer: Learning offroad costmaps for long range planning}, 
      author={Kasi Viswanath and Felix Sanchez and Timothy Overbye and Jason M. Gregory and Srikanth Saripalli},
      year={2025},
      eprint={2505.09739},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2505.09739}, 
}