Warehouse LiDAR Point Cloud Dataset

Autonomous mobile robots (AMRs) are transforming warehouse logistics, but navigation in dense warehouse environments requires precise 3D perception. Unlike outdoor driving where roads provide clear structure, warehouses present narrow aisles, dynamic obstacles (forklifts, workers, carts), varying shelf configurations, and the repetitive visual patterns that confuse camera-based systems. LiDAR provides the geometric precision these robots need.

Warehouse environments have unique characteristics that outdoor LiDAR datasets cannot capture: highly structured but variable shelving geometry, pallets at multiple heights, narrow passages between rack rows, loading dock transitions between indoor and outdoor conditions, and the reflective surfaces of shrink-wrapped pallets and metal shelving. Training data must come from real warehouse environments to capture these domain-specific challenges.

The warehouse automation market is projected to exceed $30 billion by 2027, driven by labor shortages and e-commerce growth. Companies building AMRs, autonomous forklifts, and inventory drones need training data at scale that covers the diversity of real warehouse configurations — not just the single demonstration facility most companies use for development.

Claru deploys mobile LiDAR platforms (robot-mounted or hand-carried Ouster sensors) through operating warehouse facilities during normal business hours. Collection covers distribution centers, fulfillment centers, cold storage, retail backrooms, and manufacturing warehouses to ensure representation of diverse facility types and operational conditions.

Each collection session captures point clouds at 10-20Hz as the platform traverses warehouse aisles, cross-aisles, staging areas, loading docks, and transition zones. Sessions include normal operations with active forklifts, workers, and carts providing realistic dynamic obstacle data. Multiple passes at different times of day capture varying occupancy levels.

Facility diversity spans small retail backrooms to million-square-foot distribution centers across different racking systems (selective, drive-in, push-back, carton flow), floor types (concrete, epoxy, mezzanine), and climate conditions (ambient, refrigerated, frozen). Each facility contributes unique layout characteristics that broaden the training distribution.

AMR navigation systems train on the annotated warehouse LiDAR to build 3D occupancy maps, identify navigable paths, detect dynamic obstacles, and plan collision-free trajectories through dense warehouse aisles. The diversity of facility layouts ensures navigation policies generalize across different racking configurations and aisle widths.

Inventory monitoring systems use the shelf geometry and pallet detection annotations to train models that automatically assess inventory levels from mobile robot LiDAR scans. These systems need to detect pallets at multiple rack heights, estimate fill levels, and identify empty locations — capabilities that require training data from real warehouse environments with authentic pallet arrangements.

Safety systems for autonomous forklifts train on the dynamic object tracking data to detect and predict the movements of workers, manual forklifts, and carts sharing warehouse aisles. The real-world dynamic obstacle data captures the unpredictable movement patterns of human workers that simulation cannot faithfully reproduce.