Lidar2OSM: Distributed Urban Mapping with Bandwidth-Efficient Geospatial Descriptors 📡🗺️🤖

This repository supports a project investigating the feasibility of converting raw lidar scans into OpenStreetMap (OSM)-style representations for distributed multi-robot operations. The aim is to empower autonomous robot teams to efficiently generate and accurately share low-dimensional maps of their environments, particularly in bandwidth-limited urban areas with constrained GPS signals.

The sheer complexity of LiDAR data, with its high-dimensional semantic richness, could introduce noise or irrelevant details that don't actually help with multi-robot map matching but instead make the process more difficult. This difficulty becomes exacerbated when considering transient objects and varying observational perspectives. Counterintuitively, utilizing OSM-style data, though simpler, might highlight the key features that are most useful for peer-to-peer map associations without overwhelming an underlying algorithm. The maximum clique technique is particularly effective at exploiting these core relationships in the filtered data.

Included Projects

This repository includes the semantic_clipper project, which is based on work originally developed by the MIT Aerospace Controls Lab. The code is used under the MIT License, and the original license is included in the semantic_clipper/ directory.

CU-MULTI Dataset

There are a limited number of datasets available that use a mobile ground-based robot in urban scenarios with both accurate GPS data, lidar, and IMU. It is for this reason we demonstrate our findings on two major on-road datasets, KITTI-360 and NuScenes. However, we fill this gap with the CU-MULTI Dataset, a multi-robot dataset collected in an off-road urban environment consisting of two large environments on the University of Colorado Boulder's Main Campus.

Running Lidar2OSM (Python)

This repo uses a standard src/ Python package layout. Install the package in editable mode once, then you can run scripts/CLI without setting PYTHONPATH.

Install (recommended)

conda env create -f environment.yaml
conda activate lidar2osm_env
python -m pip install -e .

One-config workflow

Edit the single repo-root config:

• config.yaml

Run the pipelines

# Data relabeling pipeline (runs the existing scripts under src/core/data/)
lidar2osm --data_pipeline --config config.yaml

# Training
lidar2osm --training_pipeline --config config.yaml

# Inference
lidar2osm --inference_pipeline --config config.yaml

You can also use module invocation:

python -m lidar2osm --data_pipeline --config config.yaml

Run individual scripts (still supported)

After pip install -e ., you can run the underlying scripts directly:

# Data pipeline scripts
python src/core/data/create_global_sem_map_octree.py --dataset_path /mnt/semkitti/cu-multi-data/ --num_scans 10 --global_voxel 0.1
python src/core/data/extend_building_offsets.py --dataset_path /mnt/semkitti/cu-multi-data/
python src/core/data/relabel_maps.py --dataset_path /mnt/semkitti/cu-multi-data/
python src/core/data/relabel_dense_semantics.py --dataset_path /mnt/semkitti/cu-multi-data/
python src/core/data/relabel_scans.py --dataset_path "/mnt/semkitti/cu-multi-data/"

# Model Testing
python src/core/models/train.py --config config.yaml
python src/core/models/infer.py --config config.yaml