Autoware in Construction: Gap Analysis and LiDAR Perception Toward Off-Road Autonomous Driving

Technology Gaps in Autonomous Driving Between Public Roads and Construction Sites

We identified key differences between public-road and off-road construction environments in each module of the Autoware stack, as summarized in Table I . Challenges in construction environments are interdependent across these modules. Therefore, achieving construction autonomy requires system-wide integration rather than the isolated replacement of individual algorithms.

Lessons Learned from Field Development of LiDAR-based Perception

Among the technology gaps in Table I , this section presents our LiDAR-based perception pipeline: LiDAR sensor $\rightarrow$ sensor driver $\rightarrow$ preprocessing filters $\rightarrow$ multi-LiDAR concatenation $\rightarrow$ postprocessing filters $\rightarrow$ ground segmentation $\rightarrow$ dust filtering $\rightarrow$ object detection, developed by extending the existing Autoware pipeline.

Roadmap toward End-to-End autonomous driving in construction

Current learning-based perception modules in Autoware have been trained primarily on public-road datasets [ 14 ] . For example, a standard LiDAR-based CenterPoint model detects five object classes: cars, trucks, buses, bicycles, and pedestrians.

Construction sites, however, contain a substantially different range of objects, including heavy construction machinery, service vehicles, workers wearing high-visibility clothing, traffic cones, rocks, and berms.

Sources and Demonstrations

Figures are reproduced from Otsuki et al. See the full paper for experimental details and the project page, when the paper links one, for demonstrations and videos.