Dynamic Indoor 3D Point Cloud and Mesh Dataset with Semantic and Instance Labels for Wireless Propagation and Ray Tracing

Overview This dataset provides synchronized 3D scene representations and wireless propagation data for a dynamic indoor environment. It is intended to support research at the intersection of three-dimensional scene understanding, radio propagation modeling, ray tracing, sensing-assisted communications, and machine-learning-based wireless prediction. Scene and radio configuration The simulated environment represents an indoor industrial room with approximate dimensions of 9.5 m x 9.5 m x 3.0 m. The scene contains walls, floor and ceiling structures, furniture, industrial equipment, robotic platforms, two articulated robotic manipulators, a humanoid robot, a quadruped robot, an unmanned aerial vehicle, and a moving human actor. The scene geometry is separated into static, articulated dynamic, and actor-related components. Static geometry represents the fixed environment, while articulated robot meshes and actor geometry are updated for each frame according to the recorded poses and motion states. A single transmitter, identified as tx_ap, is positioned at [0.0, 0.0, 2.4] m and operates with a transmit power of 30 dBm. Six receivers are associated with different robotic and human entities. The receiver placement captures wireless links with different heights, surrounding geometries, mobility patterns, and blockage conditions. Wireless propagation is simulated at a carrier frequency of 28 GHz. The ray-tracing configuration includes line-of-sight propagation and specular reflections with a maximum propagation depth of two interactions. Diffuse reflection and refraction are disabled. The solver uses up to 20,000 samples per source and supports up to 10,000 paths per source. Temporal sequence The release contains 2,446 temporally ordered scene states sampled at 10 Hz, corresponding to approximately 244.6 seconds of synchronized dynamic-scene evolution. The first recorded timestamp is 0.1 s, the final timestamp is 244.6 s, and the temporal interval between consecutive frames is 0.1 s. For every sampled state, the static environment is combined with the current poses and geometries of the articulated robots and the moving human actor. The complete wireless dataset contains 14,676 frame-RX records, corresponding to 2,446 frames evaluated for six receivers. For predictive experiments using a one-second horizon, the sequence provides 2,436 valid source-target frame pairs and 14,616 aligned frame-RX samples. Three-dimensional data Each scene state is represented by a ground-truth point cloud containing exactly 100,000 points. The complete release therefore contains 244,600,000 labeled 3D points. Each point contains the following properties: x, y, and z spatial coordinates; semantic class label; instance identifier; material identifier; object-type identifier; source-type identifier. These annotations enable semantic, instance-aware, material-aware, and object-aware analysis of the dynamic scene. The dataset additionally provides polygon meshes used to construct the ray-tracing scenes. These include merged static meshes grouped by radio-material class, reusable model meshes, articulated robot-link meshes, and frame-specific actor geometry. The static environment is represented by 11 merged material meshes: cardboard; ceiling board; chipboard; glass; human skin; metal; panel wall; plastic; textile; vinyl tile; wood. Each Sionna RT scene contains 11 static shapes, 21 articulated dynamic shapes, and one actor shape. The package includes 89 PLY geometry files, together with supporting DAE, OBJ, STL, and GLB assets. Static, dynamic, and actor components remain distinguishable throughout the dataset structure. Voxelized scene representations The dataset contains a complete voxelized representation for all 2,446 scene frames. The voxel size is 0.04 m, and the voxel grid dimensions are 250 x 250 x 88 cells. Each voxel representation includes: voxel coordinates; occupancy values; semantic information; material information; instance information; instance-boundary information. The voxel data contain 21 fine semantic classes and 10 material channels. Ray-tracing and wireless data Each frame includes a portable Sionna RT XML scene description referencing the corresponding package-local geometry. The dataset contains 2,446 XML scene files. All geometry references use relative paths, and all required mesh files are included in the archive. The associated wireless products include: frame and receiver identifiers; timestamps; transmitter and receiver positions; carrier frequency; transmit power; number of propagation paths; minimum and maximum propagation delay; delay spread; aggregate path gain; aggregate path gain in decibels; simulated received power in dBm; solver validity information; diagnostic and error fields; binary wireless-event labels; frame and RX identifiers for cross-modal alignment. The provided wireless data represent compact link-level propagation summaries. The archive does not contain complete path-level delay and power lists, path angles, detailed interaction sequences, complete channel coefficients, or full channel impulse response tensors. The received-power values represent absolute simulated received power in dBm. Temporal wireless labels The dataset contains labels defined using a prediction horizon of 10 frames, corresponding to one second. The provided labels describe: changes in the number of propagation paths; disappearance of propagation paths; received-power reductions of 0.5 dB, 1 dB, and 2 dB; increase in delay spread; wireless adaptation triggers based on 1 dB and 2 dB thresholds. The supplied data split contains 10,170 training samples, 4,326 test samples, and 120 temporally excluded samples separating the training and test periods. Compact object-aware representations The release includes complete compact descriptors aligned with the 14,616 predictive frame-RX samples. Four representation variants are provided: G, containing geometry-based information; GS, containing geometry and semantic information; GI, containing geometry and instance information; GSI, containing geometry, semantic, and instance information. The feature dimensions are: G: 65 features; GI: 715 features; GS: 1,430 features; GSI: 2,080 features. Additional arrays contain 16-dimensional link-context information and a 16-dimensional current-beam one-hot representation. Dataset structure The archive contains the following main directories: sionna_xml, containing the per-frame Sionna RT scenes; rt_results, containing wireless propagation summaries and temporal labels; inputs, containing frozen Panda and UR5 pose logs and their manifest; gt_scene_pointclouds, containing the labeled ground-truth point clouds; frames, containing per-frame meshes, transforms, and scene manifests; features, containing compact descriptors and voxel representations; config, containing experiment and taxonomy configuration; geometry, containing static, dynamic, and supporting geometry assets. All included data products are aligned using frame identifiers, receiver identifiers, timestamps, and package-relative file references. Generator and source code The dataset was generated using a Gazebo-based dynamic-scene workflow connected to Sionna RT for wireless ray tracing. The generator, processing scripts, validation tools, and workflow documentation are available in the following public repository: https://github.com/Teleinfrastructure-Research-Lab/gazebo-sionna-pipeline The repository provides the processing workflow used to construct synchronized geometry, point-cloud data, ray-tracing scenes, wireless propagation summaries, voxel representations, and compact object-aware features. The dataset archive can be inspected independently because the required runtime geometry, configuration files, pose inputs, scene descriptions, and experiment metadata are included using portable package-relative paths. Potential applications The dataset can directly support research on: semantic, instance-aware, and material-aware 3D scene analysis; temporal propagation-path change prediction; received-power-drop prediction; delay-spread-change prediction; wireless adaptation-trigger prediction; semantic and instance-aware wireless learning; comparison of geometry-only and object-aware scene representations; multimodal fusion of point-cloud, voxel, mesh, and wireless data. With additional processing, the dataset may support: 3D-aware radio propagation prediction; dynamic blockage analysis; radio environment maps; sensing-assisted and environment-aware communications; wireless digital twins; environment-aware wireless control; benchmarking machine-learning models for dynamic wireless environments.