{ "@context": { "@language": "en", "@vocab": "https://schema.org/", "citeAs": "cr:citeAs", "column": "cr:column", "conformsTo": "dct:conformsTo", "cr": "http://mlcommons.org/croissant/", "rai": "http://mlcommons.org/croissant/RAI/", "data": { "@id": "cr:data", "@type": "@json" }, "dataType": { "@id": "cr:dataType", "@type": "@vocab" }, "dct": "http://purl.org/dc/terms/", "prov": "http://www.w3.org/ns/prov#", "equivalentProperty": "cr:equivalentProperty", "examples": { "@id": "cr:examples", "@type": "@json" }, "extract": "cr:extract", "field": "cr:field", "isArray": "cr:isArray", "arrayShape": "cr:arrayShape", "fileProperty": "cr:fileProperty", "fileObject": "cr:fileObject", "fileSet": "cr:fileSet", "format": "cr:format", "includes": "cr:includes", "isLiveDataset": "cr:isLiveDataset", "jsonPath": "cr:jsonPath", "key": "cr:key", "md5": "cr:md5", "parentField": "cr:parentField", "path": "cr:path", "recordSet": "cr:recordSet", "references": "cr:references", "regex": "cr:regex", "repeated": "cr:repeated", "replace": "cr:replace", "samplingRate": "cr:samplingRate", "sc": "https://schema.org/", "separator": "cr:separator", "source": "cr:source", "subField": "cr:subField", "transform": "cr:transform" }, "@type": "sc:Dataset", "name": "Yonder", "description": "A 4.65-million-frame drone-perspective dataset spanning 167 indoor 3D environments (all from HSSD, all with semantic annotations), with stereo RGB, depth, IR, LiDAR-360, and semantic segmentation captured at 387,527 navmesh-sampled waypoints (12 yaws each). Released alongside a closed-loop benchmark designed to expose the cross-simulator generalization gap: the failure of perception trained on one simulator to transfer to a different simulator, even when both target the same task.", "conformsTo": "http://mlcommons.org/croissant/1.0", "license": "https://creativecommons.org/licenses/by-nc/4.0/", "url": "https://huggingface.co/datasets/astralhf/yonder", "version": "1.0.0", "datePublished": "2026-05-01", "creator": { "@type": "Organization", "name": "Anonymous (withheld for NeurIPS 2026 double-blind review)" }, "publisher": { "@type": "Organization", "name": "Anonymous (withheld for NeurIPS 2026 double-blind review)" }, "keywords": [ "robotics", "drone navigation", "vision-language navigation", "open-vocabulary detection", "embodied AI", "Habitat-Sim", "Isaac Sim", "cross-simulator transfer", "closed-loop benchmark" ], "isLiveDataset": false, "citeAs": "@inproceedings{anonymous2026yonder, title={Yonder: A 4.65M-Frame Drone Navigation Dataset and the Cross-Simulator Generalization Gap}, author={Anonymous Author(s)}, booktitle={NeurIPS Datasets and Benchmarks Track}, year={2026}}", "rai:dataCollection": "Generated by flying a simulated Holybro x500v2 quadcopter through 3D scenes in Habitat-Sim. Waypoint locations are sampled from the scene navmesh with adaptive densification; 12 yaw orientations are captured per waypoint. Sensor capture is fully deterministic given (scene_id, seed). No human subjects were involved in any phase of data collection. No real-world imagery was collected.", "rai:dataCollectionType": "simulated", "rai:dataCollectionTimeFrame": "2026-01-01/2026-03-31", "rai:dataAnnotationProtocol": "Semantic segmentation annotations are obtained directly from the source 3D scene meshes (HSSD). Each pixel inherits the instance ID of the underlying mesh face. Bounding boxes are derived programmatically: for each unique semantic instance ID with rendered area > 100 pixels, the axis-aligned bounding box of its pixel set is recorded. All 167 HSSD scenes carry semantic instance labels (median 38 unique instances per scene, range 14-80). No human annotators were involved.", "rai:dataAnnotationPlatform": "n/a (programmatic, no human annotators)", "rai:dataAnnotationAnalysis": "Bounding box quality was spot-checked across 50 randomly sampled waypoints from 10 scenes. Common failure modes: (1) over-segmentation of articulated objects (e.g., a chair's legs become separate boxes from its seat where mesh authoring split them); (2) tight bounding boxes around partially-occluded objects can be visually misleading. These are inherent to mesh-authored semantics and are documented as a known limitation.", "rai:dataReleaseMaintenancePlan": "The dataset is released as a static v1.0.0 snapshot. Errata, additional split files, and supplementary metadata may be added under semver patch/minor releases; the rendered NPZ data itself will not be modified post-release. Issues and clarifications are tracked on the HuggingFace Hub repository. After NeurIPS deanonymization, a maintainer contact will be added to the dataset card.", "rai:dataLimitations": "Yonder is a synthetic-only dataset. Performance on Yonder does not transfer to real-world imagery without explicit sim-to-real treatment. The source scenes (HSSD) are biased toward Western residential interiors; transfer to other interior styles, outdoor settings, or real-world drone footage is not validated. Semantic annotations come from HSSD's mesh-authored instance IDs and inherit known limitations (over-segmentation of articulated objects; occasionally misleading boxes around partially-occluded instances).", "rai:hasSyntheticData": true, "prov:wasGeneratedBy": "Rendered in Habitat-Sim 0.3.x using a simulated Holybro x500v2 quadcopter. Waypoints were sampled from each scene's navmesh with adaptive density; 12 yaw orientations were captured per waypoint across stereo RGB, depth, IR, LiDAR-360, and semantic-segmentation sensors. Rendering is fully deterministic given (scene_id, seed). Source 3D scenes are HSSD (167 scenes). No human subjects, no real imagery, no PII.", "rai:dataSocialImpact": "Drone-perspective perception models trained on this dataset could in principle be applied to surveillance, person-identification, or autonomous-targeting systems. Yonder contains no real persons or PII, but the perception capabilities it enables are dual-use. We discourage use of Yonder-trained models for any application that identifies specific real persons or supports lethal-autonomy systems. The CC-BY-NC license also restricts commercial military/surveillance applications by default.", "rai:dataBiases": "Scene-distribution bias: all scenes are HSSD residential interiors (kitchens, bedrooms, living rooms with Western design conventions). Underrepresented: institutional/industrial interiors, non-Western residential styles, outdoor or transitional spaces, low-light or unusual-lighting conditions. Object-distribution bias: object categories are skewed toward those well-represented in HSSD's mesh library (furniture, common household objects); rare or specialized objects are underrepresented. Geographic bias: scene authors and 3D-scan subjects are concentrated in North America and Western Europe.", "rai:dataUseCases": "Recommended: training drone-perspective perception models with closed-loop validation in the deployment simulator; studying cross-simulator generalization; benchmarking visual-language navigation when paired with a closed-loop evaluator. Discouraged: end-to-end navigation policy training (no expert trajectories provided); reporting fine-tuning gains based solely on Yonder's offline evaluation split. Disallowed: surveillance, biometric identification, or any application that identifies specific real persons.", "rai:personalSensitiveInformation": "None. Yonder is rendered from synthetic 3D scenes with no real persons, no faces, no PII, and no biometric data. The simulated drone trajectories contain only synthetic poses and sensor readings.", "distribution": [ { "@type": "cr:FileObject", "@id": "yonder-repo", "name": "yonder-repo", "description": "The Yonder dataset hosted on the HuggingFace Hub.", "contentUrl": "https://huggingface.co/datasets/astralhf/yonder", "encodingFormat": "git+https", "sha256": "main" }, { "@type": "cr:FileSet", "@id": "waypoint-npzs", "name": "waypoint-npzs", "description": "Per-waypoint NPZ files. Each file holds one drone pose with 12 yaw orientations across all sensor modalities.", "containedIn": { "@id": "yonder-repo" }, "encodingFormat": "application/x-npz", "includes": "indoor/drone-data/augmented/*/wp*.npz" }, { "@type": "cr:FileSet", "@id": "scene-manifests", "name": "scene-manifests", "description": "One manifest per scene, recording source dataset, source_id, scene MD5, license note, and waypoint sampling parameters.", "containedIn": { "@id": "yonder-repo" }, "encodingFormat": "application/json", "includes": "indoor/drone-data/augmented/*/manifest.json" } ], "recordSet": [ { "@type": "cr:RecordSet", "@id": "waypoints", "name": "waypoints", "description": "One record per waypoint NPZ, capturing all 12 yaw orientations of all sensor modalities at a single drone pose.", "field": [ { "@type": "cr:Field", "@id": "waypoints/scene_id", "name": "scene_id", "description": "Scene identifier (e.g., hssd-102343992). Maps to the source 3D scene.", "dataType": "sc:Text", "source": { "fileSet": { "@id": "waypoint-npzs" }, "extract": { "fileProperty": "fullpath" }, "regex": "indoor/drone-data/augmented/([^/]+)/wp[0-9]+\\.npz" } }, { "@type": "cr:Field", "@id": "waypoints/waypoint_id", "name": "waypoint_id", "description": "Zero-padded sequential waypoint index within the scene.", "dataType": "sc:Integer", "source": { "fileSet": { "@id": "waypoint-npzs" }, "extract": { "fileProperty": "fullpath" }, "regex": "indoor/drone-data/augmented/[^/]+/wp([0-9]+)\\.npz" } }, { "@type": "cr:Field", "@id": "waypoints/left_rgb", "name": "left_rgb", "description": "Stereo left RGB camera, 12 yaw orientations. Shape (12, 480, 640, 3) uint8. NPZ keys: left_rgb_yaw00 ... left_rgb_yaw11.", "dataType": "sc:ImageObject", "repeated": true }, { "@type": "cr:Field", "@id": "waypoints/right_rgb", "name": "right_rgb", "description": "Stereo right RGB camera, 12 yaw orientations. Shape (12, 480, 640, 3) uint8.", "dataType": "sc:ImageObject", "repeated": true }, { "@type": "cr:Field", "@id": "waypoints/forward_depth", "name": "forward_depth", "description": "Forward-facing depth, meters. Shape (12, 480, 640) float16.", "dataType": "sc:ImageObject", "repeated": true }, { "@type": "cr:Field", "@id": "waypoints/landing_cam", "name": "landing_cam", "description": "Downward-facing landing camera RGB. Shape (12, 480, 640, 3) uint8.", "dataType": "sc:ImageObject", "repeated": true }, { "@type": "cr:Field", "@id": "waypoints/up_ir", "name": "up_ir", "description": "Upward-facing IR camera. Shape (12, 480, 640) uint8.", "dataType": "sc:ImageObject", "repeated": true }, { "@type": "cr:Field", "@id": "waypoints/down_ir", "name": "down_ir", "description": "Downward-facing IR camera. Shape (12, 480, 640) uint8.", "dataType": "sc:ImageObject", "repeated": true }, { "@type": "cr:Field", "@id": "waypoints/lidar360", "name": "lidar360", "description": "360-degree LiDAR point cloud. Shape (1024, 16) float32 meters.", "dataType": "sc:Float", "isArray": true, "arrayShape": "1024,16" }, { "@type": "cr:Field", "@id": "waypoints/position", "name": "position", "description": "Drone position in Habitat-Sim world frame. Shape (3,) float32, meters.", "dataType": "sc:Float", "isArray": true, "arrayShape": "3" }, { "@type": "cr:Field", "@id": "waypoints/orientation", "name": "orientation", "description": "Drone orientation as quaternion. Shape (12, 4) float32.", "dataType": "sc:Float", "isArray": true, "arrayShape": "12,4" }, { "@type": "cr:Field", "@id": "waypoints/semantic_seg", "name": "semantic_seg", "description": "Per-pixel category IDs keyed to the per-scene COCO categories array in annotations/{scene}/annotations.json. Shape (12, 480, 640) uint16, one channel per yaw. Available for all 167 HSSD scenes; stored as separate {scene}_wp{NNNN}_semantics.npz files under semantics/{scene}/ in the repo.", "dataType": "sc:Integer", "isArray": true, "arrayShape": "12,480,640" } ] } ] }