"ALAR: A Multimodal Underwater Dataset for Harsh-Domain Perception"

"The ALAR dataset is a synthetic multimodal underwater perception dataset designed to support research on imaging sonar, underwater vision, multimodal learning, and sim-to-real transfer in harsh sensing conditions. The dataset was generated with ALAR, a runtime-configurable extension of HoloOcean that enables reproducible scene population with custom underwater assets and explicit sonar refresh after scene updates, ensuring that runtime-inserted geometry remains acoustically observable. Data were acquired with a simulated BlueROV2 platform equipped with a front RGB camera, a bottom RGB camera, an imaging sonar, and auxiliary telemetry.The current release contains 27,022 synchronized samples collected over 292 runs and organized into three acquisition settings: World1, a manually arranged and object-centric environment for structured capture; World2, a runtime-generated environment for reproducible coverage-oriented evaluation; and Only Objects, an object-only setting without seabed context for controlled target-signature analysis. The dataset includes six semantic categories overall: seafloor, tube, submarine, mine, anchor, and torpedo, with sensor-specific annotations where needed to account for the different sensing footprints of the bottom camera and the imaging sonar.Each sample includes front and bottom RGB images in PNG format, imaging sonar data as 256 x 256 float32 matrices in NPZ format, frame-level metadata in CSV format, run-level metadata in YAML format, and reproducible scene configuration files in JSON format. The dataset is intended for benchmarking multimodal underwater perception, domain shift analysis across acquisition regimes, and synthetic-to-real experimentation."