Robust 2D Traversability Mapping for Construction AMRs via Failure-Mode-Aware Fusion of LiDAR Geometry and Monocular Semantics
Autonomous Mobile Robots (AMRs) on active construction sites face severe navigational challenges: geometry-based traversability mapping (e.g., LiDAR) misses visually hazardous but geometrically flat surfaces like wet mud and ponding concrete, while abrupt geometry on drivable speed-breakers and inclines produces phantom obstacles. We propose a real-time, failure-mode-aware multimodal traversability pipeline on an NVIDIA Jetson AGX Orin, where LiDAR is the primary geometric safety estimate and monocular semantics act as a selective, class- and confidence-gated corrective signal. The representation retains distinct traversable classes, namely flat road, terrain, and rocky terrain, while flagging construction hazards. We also release a multimodal construction-site dataset from a custom AMR: four closed-loop ROS 2 sequences from two active sites (RGB, depth, LiDAR, IMU, GPS-RTK, odometry) plus 506 annotated frames across 28 semantic classes. By projecting LiDAR onto dense semantic masks, resolving sparsity via morphological in-painting, and applying failure-mode-aware fusion with Patchwork++, the system corrects complementary geometric failure modes for a local AMR costmap.