Appendix I: Dataset Conformance Testing (Informative)

This appendix documents systematic conformance testing performed against five public reference datasets. The results validated the completeness and expressiveness of the SpatialDDS 1.5 sensing, mapping, coordination, and spatial events profiles and directly informed several normative additions to this specification.

Motivation

Sensor-data specifications risk becoming disconnected from real-world workloads if they are designed in isolation. To guard against this, the SpatialDDS 1.5 profiles were validated against five complementary datasets that together exercise the full signal-to-semantics pipeline and multi-agent coordination:

Dataset Focus Modalities Stressed
nuScenes (Motional / nuTonomy) Perception → semantics Camera (6×), lidar, radar detections (5×), 3D annotations, coordinate conventions
DeepSense 6G (ASU Wireless Intelligence Lab) Signal → perception Raw radar I/Q tensors, 360° cameras, lidar, IMU, GPS-RTK, mmWave beam vectors
S3E (Sun Yat-sen University / HKUST) Multi-agent coordination 3 UGVs × (lidar, stereo, IMU), UWB inter-robot ranging, RTK-GNSS, collaborative SLAM
ScanNet (TU Munich / Princeton) Indoor scene understanding RGB-D depth frames, 3D surface mesh, instance segmentation (NYU40), room-level zones, 20 scene types
LaMAR (CVG ETH Zurich / Microsoft) AR localization with radio assistance WiFi scans (wifi.txt), Bluetooth scans (bt.txt), RGB images, AR trajectories, visual localization metadata

nuScenes was chosen because it stresses sensor diversity, per-detection radar fields rarely found in other corpora (compensated velocity, dynamic property, RCS), and rich annotation metadata (visibility, attributes, evidence counts). DeepSense 6G was chosen because it stresses signal-level data (raw FMCW radar cubes, phased-array beam power vectors) and ISAC modalities absent from traditional perception datasets. S3E was chosen because it is the first collaborative SLAM dataset with UWB inter-robot ranging and exercises the multi-agent capabilities — map lifecycle, inter-map alignment, range-only constraints, and fleet discovery — that differentiate SpatialDDS from single-vehicle frameworks such as ROS 2. ScanNet was chosen because it is the definitive indoor RGB-D scene understanding benchmark, uniquely exercises depth sensing (DEPTH16) and the Spatial Events extension (room zones, object-in-room events, per-class occupancy counts), and validates the semantics profile's instance segmentation types against a rich 40-class indoor vocabulary. LaMAR was chosen because it provides paired visual and radio observations (WiFi/Bluetooth) for AR localization and directly tests whether radio fingerprints can be represented as first-class typed streams rather than ad hoc JSON metadata.

The goal was not to certify particular datasets but to answer two concrete questions: Can every field, enum, and convention in each dataset's schema be losslessly mapped to SpatialDDS 1.5 IDL without workarounds or out-of-band agreements? And for multi-agent scenarios: Can the full coordination lifecycle — from independent mapping through inter-map alignment — be expressed using the standard types?

Methodology

For each dataset, a conformance harness was constructed as a self-contained Python 3 script that:

  1. Mirrors the SpatialDDS 1.5 IDL as Python data structures (enum values, struct field lists, normative prose flags).
  2. Mirrors the dataset schema as synthetic data (sensor names, field lists, data shapes).
  3. Runs targeted checks, each producing a verdict:
Verdict Meaning
PASS Dataset field maps losslessly to an existing SpatialDDS type or enum value.
GAP A mapping exists conceptually but the required SpatialDDS type or field does not yet exist.
MISSING No SpatialDDS construct exists for the dataset field; a new profile is needed.
  1. Reports a per-modality scorecard.

Neither nuScenes nor DeepSense 6G harness requires network access, a DDS runtime, or a dataset download. Both operate as static schema-vs-schema dry runs, reproducible in any CI environment. The S3E (§I.3) and ScanNet (§I.4) conformance sections were performed as manual schema analyses following the same check structure; scripted harnesses are planned for a future revision.

I.1 nuScenes Conformance

Reference Dataset

nuScenes (Motional / nuTonomy) is a multimodal autonomous driving dataset containing:

Dimension Value
Scenes 1,000 (20 s each)
Cameras 6 surround-view (FRONT, FRONT_LEFT, FRONT_RIGHT, BACK, BACK_LEFT, BACK_RIGHT)
Lidar 1 x 32-beam spinning (Velodyne HDL-32E), ~34 k points/scan
Radar 5 x Continental ARS 408 (FRONT, FRONT_LEFT, FRONT_RIGHT, BACK_LEFT, BACK_RIGHT)
3D annotations 1.4 M oriented bounding boxes, 23 object classes
Annotation metadata visibility tokens, attribute tokens, per-box lidar/radar point counts
Coordinate convention Right-handed; quaternions in (w, x, y, z) order

Checks Performed (27)

Radar — Detection Path (6 checks)
ID Check Description
R-01 Detection-centric profile RadDetection struct exists with per-detection xyz, velocity, RCS, dyn_prop.
R-02 Per-detection velocity Cartesian velocity_xyz (preferred) + scalar v_r_mps (fallback), both with has_* guards.
R-03 Ego-compensated velocity velocity_comp_xyz field for ego-motion-compensated velocity.
R-04 Dynamic property enum RadDynProp covers all 7 nuScenes values (UNKNOWN through STOPPED).
R-05 Per-detection RCS rcs_dbm2 field in dBm² with has_rcs_dbm2 guard.
R-06 Sensor type enum RadSensorType differentiates SHORT_RANGE, LONG_RANGE, IMAGING_4D, etc.
Vision (5 checks)
ID Check Description
V-01 RigRole coverage RigRole enum includes FRONT, FRONT_LEFT, FRONT_RIGHT, BACK, BACK_LEFT, BACK_RIGHT.
V-02 Pre-rectified images Normative prose documents dist = NONE with model = PINHOLE semantics.
V-03 Image dimensions CamIntrinsics.width / height are REQUIRED; zero values are malformed.
V-04 Keyframe flag VisionFrame.is_key_frame boolean.
V-05 Quaternion reorder §2 table maps nuScenes (w,x,y,z) to SpatialDDS (x,y,z,w).
Lidar (6 checks)
ID Check Description
L-01 BIN_INTERLEAVED encoding CloudEncoding value for raw interleaved binary with normative record layout table.
L-02 Per-point timestamps PointLayout.XYZ_I_R_T and XYZ_I_R_T_N with normative prose for the t field.
L-03 Metadata guards LidarMeta uses has_range_limits, has_horiz_fov, has_vert_fov guards.
L-04 Timestamp presence flag LidarFrame.has_per_point_timestamps signals per-point timing in the blob.
L-05 t_end computation Normative guidance for computing t_end from t_start + 1/rate_hz or max(point.t).
L-06 Ring field PointLayout.XYZ_I_R carries ring as uint16.
Semantics (5 checks)
ID Check Description
S-01 Size convention Normative: size[0] = width (X), size[1] = height (Z), size[2] = depth (Y). nuScenes (w,l,h) -> (w,h,l) mapping documented.
S-02 Attributes Detection3D.attributes as sequence<MetaKV, 8> with has_attributes guard.
S-03 Visibility Detection3D.visibility float [0..1] with has_visibility guard.
S-04 Evidence counts num_lidar_pts + num_radar_pts with has_num_pts guard.
S-05 Quaternion reorder §2 table covers annotation quaternion conversion.
Common / Core (5 checks)
ID Check Description
C-01 Quaternion table §2 convention table covering GeoPose, ROS 2, nuScenes, Eigen, Unity, Unreal, OpenXR, glTF.
C-02 FQN guidance FrameRef { uuid, fqn } semantics documented; UUID is authoritative.
C-03 Local-frame coverage §3.3.4 covers local-only deployments.
C-04 has_* pattern consistency All new optional fields use the has_* guard pattern uniformly.
C-05 Sequence bounds Standard bounds table: SZ_MEDIUM (2048), SZ_SMALL (256), SZ_XL (32768), SZ_LARGE (8192).

Results

All 27 nuScenes checks pass.

Modality Checks Pass Remaining Gaps
Radar (detections) 6 6 0
Vision 5 5 0
Lidar 6 6 0
Semantics 5 5 0
Common / Core 5 5 0
Total 27 27 0

I.2 DeepSense 6G Conformance

Reference Dataset

DeepSense 6G (Arizona State University, Wireless Intelligence Lab) is a large-scale multi-modal sensing and communication dataset containing:

Dimension Value
Scenarios 40+ across 12+ locations
Snapshots 1.08 M+ synchronized samples
FMCW Radar 76–81 GHz, 3 Tx × 4 Rx, complex I/Q tensor [4×256×128], 10 Hz
3D Lidar Ouster OS1-32, 32×1024, 120 m range, 865 nm, 10–20 Hz
Camera ZED2 stereo (960×540) + Insta360 ONE X2 360° (5.7K)
GPS-RTK 10 Hz, ≤1 cm accuracy (RTK fix), DOP + satellite metadata
IMU 6-axis, 100 Hz
mmWave Comm 60 GHz phased array, 64-beam codebook, 90° FoV, 10 Hz
Deployment types V2I, V2V (4× arrays/vehicle), ISAC indoor, drone

The dataset was chosen because it stresses signal-level data (raw FMCW radar cubes consumed directly by ML pipelines), 360° camera rigs, and ISAC modalities (beam power vectors, blockage state) absent from perception-focused datasets.

Checks Performed (41)

Radar — Tensor Path (8 checks)
ID Check Description
DT-01 Tensor meta struct RadTensorMeta exists with axes, voxel_type, layout, physical_meaning.
DT-02 Complex sample type SampleType.CF32 covers complex I/Q data.
DT-03 Channel axis RadTensorLayout.CH_FAST_SLOW maps raw FMCW [Rx, samples, chirps].
DT-04 MIMO antenna config num_tx, num_rx, num_virtual_channels with has_antenna_config guard.
DT-05 Waveform params bandwidth_hz, center_freq_hz, samples_per_chirp, chirps_per_frame with guard.
DT-06 Frame blob transport RadTensorFrame.hdr.blobs[] carries the raw cube; size computable from axes × sample size.
DT-07 Sensor type RadSensorType covers FMCW radar as MEDIUM_RANGE or IMAGING_4D.
DT-08 StreamMeta extrinsics T_bus_sensor (PoseSE3) + nominal_rate_hz for hand-eye calibration and 10 Hz cadence.
Vision (7 checks)
ID Check Description
DV-01 Standard camera PixFormat.RGB8 + CamIntrinsics.width/height cover ZED2 at 960×540.
DV-02 Camera extrinsics VisionMeta.base → StreamMeta.T_bus_sensor for hand-eye calibration.
DV-03 Camera model CamModel.PINHOLE for ZED2 pre-rectified output.
DV-04 Frame rate StreamMeta.nominal_rate_hz = 10 (downsampled from 30 Hz).
DV-05 360° rig roles RigRole.PANORAMIC and EQUIRECTANGULAR for Insta360 ONE X2 in V2V scenarios.
DV-06 Keyframe flag VisionFrame.is_key_frame boolean.
DV-07 Compression codec Codec enum covers JPEG/H264/H265/AV1.
Lidar (7 checks)
ID Check Description
DL-01 Lidar type LidarType.MULTI_BEAM_3D for Ouster OS1-32 (spinning, 32 rings).
DL-02 Ring count + FOV LidarMeta.n_rings, has_horiz_fov, has_vert_fov with guards.
DL-03 Range limits has_range_limits + max_range_m = 120 m.
DL-04 Point layout PointLayout.XYZ_I_R for x, y, z, intensity, ring.
DL-05 Cloud encoding CloudEncoding.BIN_INTERLEAVED for raw binary transport.
DL-06 Sensor wavelength LidarMeta.wavelength_nm with has_wavelength guard (865 nm).
DL-07 Frame rate StreamMeta.nominal_rate_hz covers 10–20 Hz.
IMU (4 checks)
ID Check Description
DI-01 6-axis sample ImuSample with accel (Vec3, m/s²) + gyro (Vec3, rad/s).
DI-02 Noise densities ImuInfo.accel_noise_density + gyro_noise_density + random walk params.
DI-03 Frame reference ImuInfo.frame_ref for sensor-to-bus mounting.
DI-04 Timestamp + sequence ImuSample.stamp + .seq for 100 Hz temporal ordering.
GPS (6 checks)
ID Check Description
DG-01 Position GeoPose.lat_deg/lon_deg/alt_m for GPS-RTK coordinates.
DG-02 Orientation GeoPose.q (QuaternionXYZW) for heading-derived orientation.
DG-03 Timestamp GeoPose.stamp for 10 Hz GPS samples.
DG-04 Covariance GeoPose.cov for positional uncertainty (RTK ≤1 cm).
DG-05 GNSS quality NavSatStatus provides DOP, fix type, and satellite count with has_dop guard.
DG-06 Speed over ground NavSatStatus.speed_mps + course_deg with has_velocity guard.
mmWave Beam (8 checks)
ID Check Description
DB-01 Beam power vector RfBeamFrame.power (sequence) carries per-beam received power. 64 entries for DeepSense exhaustive sweep. Provisional rf_beam profile (K-B1).
DB-02 Codebook metadata RfBeamMeta.n_beams (64), n_elements (16), center_freq_ghz (60.0), fov_az_deg (90), codebook_type.
DB-03 Optimal beam index RfBeamFrame.best_beam_idx (uint16) with has_best_beam guard. Ground-truth label: beam maximizing SNR.
DB-04 Blockage status RfBeamFrame.is_blocked (boolean) + blockage_confidence (float 0..1) with has_blockage_state guard.
DB-05 Multi-array set RfBeamArraySet.arrays (sequence) batches per-array frames. overall_best_array_idx + overall_best_beam_idx for cross-array best beam. Covers V2V 4-array rig.
DB-06 Sparse sweep indices RfBeamFrame.beam_indices maps power[i] to codebook position for PARTIAL/TRACKING sweeps. BeamSweepType enum: EXHAUSTIVE, HIERARCHICAL, TRACKING, PARTIAL.
DB-07 Power unit convention RfBeamMeta.power_unit (PowerUnit enum: DBM, LINEAR_MW, RSRP) declares units for RfBeamFrame.power.
DB-08 Stream linkage RfBeamFrame.stream_id matches RfBeamMeta.stream_id for meta/frame correlation.

Note: All mmWave Beam checks validated against the provisional sensing.rf_beam profile (Appendix E). Types are subject to breaking changes.

Semantics (4 checks)
ID Check Description
DS-01 2D bounding boxes Detection2D.bbox + class_id covers 8 DeepSense object classes.
DS-02 Sequence index FrameHeader.frame_seq for sample ordering.
DS-03 Class ID Detection2D.class_id (string) maps all DeepSense class labels.
DS-04 Beam/blockage labels RfBeamFrame.best_beam_idx and .is_blocked/.blockage_confidence carry ISAC-specific ground-truth labels. Covered by provisional rf_beam profile.

Results

All 44 DeepSense 6G checks pass. GNSS diagnostics are covered by NavSatStatus, and mmWave Beam checks pass against the provisional rf_beam profile (Appendix E).

Modality Checks Pass Gap Missing Notes
Radar (tensor) 8 8 0 0 —
Vision 7 7 0 0 Includes 360° rig roles
Lidar 7 7 0 0 Includes sensor wavelength
IMU 4 4 0 0 —
GPS 6 6 0 0 NavSatStatus covers GNSS diagnostics
mmWave Beam 8 8 0 0 Provisional rf_beam profile (K-B1)
Semantics 4 4 0 0 Beam labels via rf_beam
Total 44 44 0 0 100% coverage

Deferred Items

DeepSense 6G conformance has no remaining schema gaps. Future ISAC extensions (e.g., CSI/CIR profiles) remain under discussion; see Appendix K for the maturity promotion criteria.

I.3 S3E Conformance (Multi-Robot Collaborative SLAM)

Reference Dataset

S3E (Sun Yat-sen University / HKUST) is a multi-robot multimodal dataset for collaborative SLAM containing:

Dimension Value
Robots 3 UGVs (Alpha, Blob, Carol) operating simultaneously
LiDAR 1 × 16-beam 3D scanner (Velodyne VLP-16) per robot, 10 Hz
Stereo cameras 2 × high-resolution color cameras per robot
IMU 9-axis, 100–200 Hz per robot
UWB Inter-robot Ultra-Wideband ranging (pairwise distances at ~10 Hz)
GNSS Dual-antenna RTK receiver per robot (ground truth)
Environments 13 outdoor + 5 indoor sequences
Trajectory paradigms 4 collaborative patterns (concentric circles, intersecting circles, intersection curve, rays)
Format ROS 2 bag files; ground truth as TUM-format pose files

The dataset was chosen because it is the first C-SLAM dataset to include UWB inter-robot ranging, exercises multi-agent map building with inter-robot loop closures, and represents a scenario class (heterogeneous multi-robot coordination) where SpatialDDS's Mapping extension, Discovery profile, and multi-source pose graph types provide capabilities absent from ROS 2's nav_msgs and sensor_msgs.

Checks Performed (38)

Per-Robot Sensing — LiDAR (5 checks)
ID Check Description
SL-01 LiDAR meta LidarMeta with sensor_type, rate_hz, point_layout covers Velodyne VLP-16.
SL-02 Point layout PointLayout.XYZ_I_R_T carries x, y, z, intensity, ring, time — matches Velodyne binary format.
SL-03 Per-robot topic isolation Topic template spatialdds/<scene>/lidar/<sensor_id>/frame/v1 with per-robot sensor_id (e.g., alpha/vlp16).
SL-04 CloudEncoding BIN_INTERLEAVED covers raw binary point cloud blobs.
SL-05 RigRole RigRole.TOP covers single roof-mounted LiDAR.
Per-Robot Sensing — Vision (4 checks)
ID Check Description
SV-01 Stereo pair Two VisionFrame streams per robot with RigRole.LEFT / RigRole.RIGHT.
SV-02 Camera intrinsics CameraMeta with fx, fy, cx, cy, dist_model, dist_coeffs covers calibrated stereo cameras.
SV-03 Per-robot namespacing Topic spatialdds/<scene>/vision/<sensor_id>/frame/v1 isolates per-robot camera streams.
SV-04 Timestamp sync VisionFrame.stamp synchronized to common timebase via hardware PPS trigger.
Per-Robot Sensing — IMU (3 checks)
ID Check Description
SI-01 9-axis sample ImuSample with accel (Vec3, m/s²) + gyro (Vec3, rad/s) covers 6-axis; MagSample covers magnetometer.
SI-02 High-rate ordering ImuSample.seq monotonic counter handles 100–200 Hz temporal ordering.
SI-03 Extrinsic calibration Sensor-to-body transform publishable as FrameTransform (LiDAR-IMU, camera-IMU extrinsics).
Per-Robot Sensing — GNSS/RTK (3 checks)
ID Check Description
SG-01 RTK fix type GnssFixType.RTK_FIXED covers dual-antenna RTK ground truth receiver.
SG-02 GeoPose output GeoPose with lat_deg, lon_deg, alt_m, quaternion covers RTK-derived global pose.
SG-03 NavSatStatus NavSatStatus with fix_type, num_satellites, hdop, vdop covers receiver diagnostics.
Inter-Robot Ranging — UWB (4 checks)
ID Check Description
SU-01 Range edge type mapping::EdgeType.RANGE explicitly models UWB range-only constraint (scalar distance, no orientation).
SU-02 Range fields mapping::Edge.range_m + range_std_m carry measured distance and uncertainty.
SU-03 Cross-map provenance has_from_map_id / has_to_map_id populated on RANGE edges because UWB connects nodes in different robots' maps.
SU-04 Range-assisted alignment AlignmentMethod.RANGE_COARSE covers initial inter-map alignment derived solely from UWB distances.
Core Pose Graph (5 checks)
ID Check Description
SC-01 Per-robot nodes core::Node with map_id per robot (e.g., alpha-map, blob-map, carol-map), @key node_id unique per keyframe.
SC-02 Odometry edges core::Edge with type = ODOM connects sequential keyframes within each robot's map.
SC-03 Intra-robot loop closures core::Edge with type = LOOP for within-map loop closures (e.g., concentric circle paradigm).
SC-04 Versioning Node.seq monotonic per source; Node.graph_epoch increments after global re-optimization.
SC-05 Multi-source coexistence Three simultaneous source_id values on core::Node and core::Edge topics — one per robot.
Mapping Extension — Multi-Agent (8 checks)
ID Check Description
SM-01 Map lifecycle MapMeta per robot with state progressing: BUILDING → OPTIMIZING → STABLE.
SM-02 Map kind MapMeta.kind = POSE_GRAPH for each robot's SLAM output.
SM-03 Inter-robot loop closures mapping::Edge with type = INTER_MAP and has_from_map_id / has_to_map_id populated.
SM-04 MapAlignment MapAlignment with T_from_to expressing the inter-map transform after cross-robot alignment.
SM-05 Alignment revision MapAlignment.revision increments as more inter-robot edges accumulate and the alignment refines.
SM-06 Evidence trail MapAlignment.evidence_edge_ids[] references the specific cross-map edges supporting the alignment.
SM-07 MapEvent notifications MapEvent with MAP_ALIGNED event when two robots' maps are first linked.
SM-08 Concurrent map builds Three MapMeta samples simultaneously active (keyed by map_id), demonstrating multi-map lifecycle.
Discovery & Coordination (3 checks)
ID Check Description
SD-01 Service announcement Each robot publishes Announce with ServiceKind.SLAM and sensor capabilities in topics[].
SD-02 Spatial coverage Announce.coverage (Aabb3 or geo-bounds) advertises each robot's operational area.
SD-03 Multi-frame NodeGeo After inter-map alignment, NodeGeo.poses[] carries a node's pose in multiple robots' map frames simultaneously (FramedPose array).
Cross-cutting (3 checks)
ID Check Description
SX-01 Quaternion convention §2 table covers ROS 2 (x,y,z,w) to SpatialDDS (x,y,z,w) identity mapping for S3E's ROS 2 bag source.
SX-02 Coordinate frame convention Right-handed; S3E uses right-hand rule per documentation.
SX-03 Time synchronization Hardware PPS-synchronized timestamps map directly to Time { sec, nanosec }.

Results

All 38 S3E checks pass.

Modality Checks Pass Remaining Gaps
LiDAR 5 5 0
Vision 4 4 0
IMU 3 3 0
GNSS/RTK 3 3 0
UWB (inter-robot range) 4 4 0
Core Pose Graph 5 5 0
Mapping (multi-agent) 8 8 0
Discovery & Coordination 3 3 0
Cross-cutting 3 3 0
Total 38 38 0

S3E Scenario Narrative (Informative)

The S3E "teaching building" outdoor sequence illustrates the full multi-agent lifecycle:

  1. Bootstrap. Three robots (Alpha, Blob, Carol) power on and each publishes an Announce with ServiceKind.SLAM, their sensor capabilities, and an initial coverage bounding box. Each begins publishing core::Node and core::Edge (ODOM) on the pose graph topics with distinct source_id and map_id values.

  2. Independent mapping. Each robot runs visual-inertial-lidar SLAM independently. MapMeta per robot shows state = BUILDING. Keyframes stream as core::Node; odometry constraints as core::Edge (ODOM); intra-robot loop closures as core::Edge (LOOP). ImuSample, VisionFrame, and LidarFrame are published on per-robot sensor topics.

  3. UWB ranging begins. As robots come within UWB range (~50 m), pairwise distance measurements are published as mapping::Edge with type = RANGE, range_m carrying the measured distance, has_from_map_id / has_to_map_id identifying which robots' maps the linked nodes belong to.

  4. Inter-robot loop closure. When Alpha and Blob's LiDAR scans overlap, a cross-robot loop closure is detected. This is published as mapping::Edge with type = INTER_MAP, match_score carrying the ICP fitness, and from_map_id = "alpha-map", to_map_id = "blob-map".

  5. Map alignment. A MapAlignment is published linking Alpha's and Blob's maps, with method = LIDAR_ICP (or MULTI_METHOD if UWB ranges were fused), T_from_to carrying the inter-map transform, and evidence_edge_ids[] referencing the supporting cross-map edges. MapEvent with MAP_ALIGNED notifies all subscribers.

  6. Multi-frame localization. Once the alignment exists, a geo-referencing service can publish NodeGeo with poses[] containing FramedPoses in both Alpha's and Blob's map frames simultaneously. Consumers (e.g., a planning service) can pick the frame they need.

  7. Graph optimization. After sufficient inter-robot constraints accumulate, a global optimizer runs. All robots' MapMeta.state transitions to OPTIMIZING, then STABLE. graph_epoch increments on all nodes and edges. MapAlignment.revision increments. Consumers watching graph_epoch know to re-fetch the entire graph.

This end-to-end scenario is precisely what ROS 2's nav_msgs and sensor_msgs cannot express: there is no ROS 2 standard for map lifecycle, inter-map alignment, range-only constraints, or multi-agent discovery with spatial coverage.

I.4 ScanNet Conformance (Indoor Scene Understanding)

Reference Dataset

ScanNet (TU Munich / Princeton) is an RGB-D video dataset of indoor scenes containing:

Dimension Value
Scenes 1,513 (707 unique spaces, multiple rescans)
RGB-D sensor Structure.io depth + iPad color camera
Depth format 16-bit unsigned integer, millimeters, 640×480 @ 30 Hz
Color format JPEG-compressed RGB, 1296×968 @ 30 Hz
Camera poses Per-frame 4×4 camera-to-world extrinsics via BundleFusion
IMU Embedded IMU data in .sens stream
Surface reconstruction Dense triangle mesh (PLY) via BundleFusion
Semantic annotations Instance-level labels (NYU40 label set, 40 classes)
Instance annotations Per-vertex segment IDs + aggregated object instances
Scene types 20 categories (bathroom, bedroom, kitchen, living room, office, etc.)
Axis alignment Per-scene 4×4 gravity-alignment matrix
Coordinate convention Right-handed; +Z up in aligned frame

ScanNet was chosen because it is the definitive indoor RGB-D scene understanding benchmark, exercises depth sensing absent from all three prior conformance datasets, and provides room-level semantic structure that naturally maps to the Spatial Events extension — the only SpatialDDS extension not yet tested by conformance.

Checks Performed (35)

RGB-D Sensing — Color (4 checks)
ID Check Description
NC-01 Color meta VisionMeta with pix = RGB8, codec = JPEG, CamIntrinsics (fx, fy, cx, cy at 1296×968).
NC-02 Color frame VisionFrame per RGB image with frame_seq, hdr.stamp, blob reference to JPEG payload.
NC-03 Per-scene stream isolation Topic spatialdds/<scene_id>/vision/<stream_id>/frame/v1 with unique stream_id per scan.
NC-04 Rig linkage VisionMeta.rig_id shared between color and depth streams for spatial association.
RGB-D Sensing — Depth (5 checks)
ID Check Description
ND-01 Depth meta VisionMeta with pix = DEPTH16, codec = NONE (raw 16-bit), CamIntrinsics for depth camera.
ND-02 Depth pixel format PixFormat.DEPTH16 explicitly identifies 16-bit millimeter depth. Requires SN-1.
ND-03 Depth frame VisionFrame per depth image with frame_seq matching co-located color frame.
ND-04 Invalid depth convention Zero-valued pixels denote no measurement, consistent with DEPTH16 normative note.
ND-05 Depth unit Default millimeter unit; no depth_unit attribute required for ScanNet's Structure.io sensor.
IMU (2 checks)
ID Check Description
NI-01 IMU sample ImuSample with accel (Vec3, m/s²) + gyro (Vec3, rad/s) covers 6-axis IMU embedded in .sens stream.
NI-02 Temporal ordering ImuSample.seq provides monotonic ordering within the scan.
Camera Pose & Frames (4 checks)
ID Check Description
NP-01 Per-frame pose Camera-to-world 4×4 matrix maps to FrameHeader.sensor_pose (PoseSE3: translation + quaternion).
NP-02 Axis-alignment transform Per-scene gravity-alignment matrix published as FrameTransform from sensor frame to aligned frame.
NP-03 Frame hierarchy Aligned frame FQN follows §2.2 pattern: <scene_id>/aligned.
NP-04 Quaternion convention ScanNet uses 4×4 rotation matrices; decomposition to (x,y,z,w) quaternion per §2 convention table.
Mesh Reconstruction (4 checks)
ID Check Description
NM-01 Map kind MapMeta with kind = MESH for BundleFusion surface reconstruction.
NM-02 Map lifecycle MapMeta.state = STABLE for completed reconstructions (offline dataset; no BUILDING phase observed).
NM-03 Mesh payload BlobRef referencing PLY mesh file. SpatialDDS carries mesh references, not inline mesh data.
NM-04 Vertex count metadata MapMeta.attributes carries vertex/face count as MetaKV for consumers to assess mesh complexity.
3D Instance Segmentation — Semantics (6 checks)
ID Check Description
NS-01 3D detection Detection3D per annotated object instance, with class_id from NYU40 label set (e.g., "chair", "table", "door").
NS-02 Instance ID Detection3D.det_id unique per object instance within a scene (maps from ScanNet's objectId).
NS-03 Oriented bounding box Detection3D.center + size + q cover ScanNet's axis-aligned bounding boxes (identity quaternion in aligned frame).
NS-04 Track ID Detection3D.track_id groups the same physical object across multiple rescans of the same space.
NS-05 Visibility Detection3D.visibility (0–1) maps from ScanNet annotation coverage ratio.
NS-06 Class vocabulary class_id as free-form string covers all 40 NYU40 categories without a closed enum — consistent with SpatialDDS's ontology-agnostic design.
Spatial Events — Indoor Zones (6 checks)
ID Check Description
NZ-01 Room as zone SpatialZone per ScanNet scene, with zone_id = scene ID, name = human-readable scene name.
NZ-02 Zone kind ZoneKind.MONITORING for general-purpose room observation (no access restriction implied).
NZ-03 Zone bounds SpatialZone.bounds (Aabb3) enclosing the room extent, derived from mesh bounding box in aligned frame.
NZ-04 Scene type as attribute ScanNet sceneType (bathroom, bedroom, kitchen, etc.) carried as MetaKV in SpatialZone.attributes with namespace = "scene_type", json = {"type": "kitchen"}.
NZ-05 Class filter SpatialZone.class_filter populated with object classes of interest (e.g., ["person", "chair", "table"]) for selective event triggering.
NZ-06 Zone frame SpatialZone.frame_ref references the gravity-aligned frame established by the axis-alignment transform (NP-02).
Spatial Events — Object Events (4 checks)
ID Check Description
NE-01 Zone entry SpatialEvent with event_type = ZONE_ENTRY when a Detection3D instance is first observed within a SpatialZone's bounds.
NE-02 Trigger linkage SpatialEvent.trigger_det_id references the triggering Detection3D.det_id; trigger_class_id carries the NYU40 label.
NE-03 Zone state ZoneState with zone_occupancy count reflecting the number of annotated object instances within the room.
NE-04 Class counts ZoneState.class_counts (sequence of MetaKV) carries per-class occupancy (e.g., {"count": 4} for class "chair").

Results

All 35 ScanNet checks pass.

Modality Checks Pass Remaining Gaps
Color (RGB) 4 4 0
Depth (RGBD) 5 5 0
IMU 2 2 0
Camera Pose & Frames 4 4 0
Mesh Reconstruction 4 4 0
3D Instance Segmentation 6 6 0
Spatial Events — Zones 6 6 0
Spatial Events — Object Events 4 4 0
Total 35 35 0

ScanNet Scenario Narrative (Informative)

The ScanNet "apartment" scan sequence illustrates how SpatialDDS types map to a complete indoor scene understanding pipeline:

  1. Scan ingestion. An operator walks through a kitchen with an iPad running the ScanNet capture app. Color frames are published as VisionFrame (pix=RGB8, codec=JPEG) and depth frames as VisionFrame (pix=DEPTH16, codec=NONE) on paired streams linked by rig_id. ImuSample streams concurrently from the embedded IMU.

  2. Pose estimation. BundleFusion produces per-frame camera poses, published as FrameHeader.sensor_pose on each VisionFrame. The per-scene axis-alignment matrix is published as a FrameTransform from the sensor coordinate system to a gravity-aligned room frame.

  3. Mesh reconstruction. The completed surface mesh is registered as MapMeta with kind = MESH, state = STABLE. The PLY file is referenced via BlobRef. Vertex/face counts are carried in MapMeta.attributes.

  4. Zone definition. The kitchen is defined as a SpatialZone with kind = MONITORING, bounds enclosing the room extent, and attributes carrying scene_type = "kitchen". The frame_ref points to the gravity-aligned frame.

  5. 3D instance detection. Crowdsourced annotations produce Detection3D instances for each labeled object: chairs with class_id = "chair", tables with class_id = "table", a refrigerator with class_id = "refrigerator" — each with an oriented bounding box in the aligned frame.

  6. Spatial events. A zone monitoring service evaluates which Detection3D instances fall within the kitchen SpatialZone's bounds and publishes SpatialEvent (ZONE_ENTRY) for each. ZoneState is published periodically with zone_occupancy = 12 (total instances) and class_counts listing per-class breakdowns.

This pipeline exercises the Spatial Events extension end-to-end — from zone definition through detection to event generation — a capability path untested by nuScenes (no zones), DeepSense 6G (no zones), or S3E (no zones or semantics).

Deferred Items

  • Per-vertex semantic labels. ScanNet provides per-vertex class labels on the reconstructed mesh. SpatialDDS has no per-vertex label type; the labeled mesh PLY is carried as a BlobRef. A future per-vertex or per-point semantic annotation type could make this data first-class.
  • CAD model alignment. ScanNet aligns ShapeNet CAD models to detected objects. The ShapeNet model ID can be carried in Detection3D.attributes as a MetaKV, but there is no first-class CAD reference type.
  • 2D projected labels. ScanNet provides per-frame 2D semantic/instance label images. These can be published as VisionFrame with a label-specific stream_id and pix = RAW16 (16-bit label IDs), but a dedicated label pixel format is not defined.

I.5 LaMAR Conformance (AR Localization with Radio Fingerprints)

Reference Dataset

LaMAR (CVG ETH Zurich / Microsoft) is an AR localization benchmark containing aligned visual and radio observations:

Dimension Value
Captures Indoor/outdoor smartphone + HoloLens trajectories
Visual data RGB image streams for retrieval/localization
Radio data WiFi scans (wifi.txt) and Bluetooth scans (bt.txt)
Targets Visual localization recall under radio-assisted retrieval
Notable finding WiFi/BT-assisted retrieval improves localization recall (+4.6% to +17.5%)

LaMAR was selected to validate radio-assisted AR workflows and close the prior LM-1 gap where radio observations were carried only as ad hoc MetaKV JSON payloads.

Checks Performed (22)

Radio Profile Coverage (12 checks)
ID Check Description
LM-01 Typed per-scan container RadioScan carries one scan event with sensor_id, radio_type, scan_seq, and stamp.
LM-02 Typed per-observation container RadioObservation carries one transmitter measurement (identifier, measurement_kind, value).
LM-03 WiFi identifier format BSSID maps to lowercase colon-separated identifier.
LM-04 BLE identifier format Beacon UUID/MAC maps to canonical identifier.
LM-05 RSSI representation RSSI maps to measurement_kind = RSSI, value in dBm.
LM-06 WiFi frequency/channel frequency_mhz, band, and channel map with has_* guards.
LM-07 BLE major/minor iBeacon major/minor maps with has_major_minor.
LM-08 BLE Tx power Advertised Tx power maps with has_tx_power.
LM-09 Scan duration Variable scan-window duration maps to scan_duration_s.
LM-10 Aggregation window ±window aggregation maps to aggregation_window_s.
LM-11 Sensor metadata RadioSensorMeta captures capability flags and adapter metadata.
LM-12 Schema tag schema_version set to spatial.sensing.radio/1.5.
Discovery and QoS Integration (5 checks)
ID Check Description
LD-01 Registered type Discovery type registry includes radio_scan.
LD-02 QoS profile RADIO_SCAN_RT available for radio scan topics.
LD-03 Topic naming Topic pattern spatialdds/<scene>/radio/<sensor_id>/scan/v1 is valid under §3.3.1.
LD-04 Meta durability RadioSensorMeta uses RELIABLE + TRANSIENT_LOCAL semantics.
LD-05 Optional fields Radio optional values consistently follow has_* guard pattern.
Interop and Privacy (5 checks)
ID Check Description
LP-01 Multi-technology support A device can publish separate WiFi and BLE scan streams with shared timebase.
LP-02 Fingerprint matching readiness Canonical identifier formats support stable join keys across sessions.
LP-03 Pose association Optional sensor_pose + pose_frame_ref supports radio-visual alignment.
LP-04 Privacy guidance Identifier anonymization guidance documented for sensitive deployments.
LP-05 No algorithm coupling Profile transports observations only; no positioning algorithm mandated.

Results

All 22 LaMAR checks pass.

Modality Checks Pass Gap Missing Notes
Radio profile 12 12 0 0 LM-1 closed via RadioScan/RadioSensorMeta
Discovery + QoS 5 5 0 0 radio_scan + RADIO_SCAN_RT integrated
Interop + privacy 5 5 0 0 Identifier and anonymization guidance documented
Total 22 22 0 0 100% coverage

Deferred Items

  • CSI/CIR first-class payloads. CSI_REF currently points to external payloads. A future extension may define typed CSI/CIR transport.
  • Multi-band coexistence metadata. Additional fields for scan policy and dwell-time scheduling may be needed for dense AP environments.

Reproducing the Tests

The nuScenes and DeepSense 6G conformance harnesses are self-contained Python 3 scripts with no external dependencies.

nuScenes harness (scripts/nuscenes_harness_v2.py):

python3 scripts/nuscenes_harness_v2.py

Mirrors the SpatialDDS 1.5 IDL structures as Python dictionaries and checks them against the nuScenes schema. Produces a plain-text report and a JSON results file.

DeepSense 6G harness (scripts/deepsense6g_harness_v3.py):

python3 scripts/deepsense6g_harness_v3.py

Validates 44 checks across 7 modalities (radar tensor, vision, lidar, IMU, GPS, mmWave beam, semantics). The mmWave beam checks validate against the provisional rf_beam profile (Appendix E). Produces a plain-text report and a JSON results file.

S3E conformance: The 38 S3E checks documented in §I.3 were performed as a manual schema-vs-schema analysis. A scripted harness (scripts/s3e_harness_v1.py) following the same pattern as the nuScenes and DeepSense 6G scripts is planned for a future revision.

ScanNet conformance: The 35 ScanNet checks documented in §I.4 were performed as a manual schema-vs-schema analysis. A scripted harness (scripts/scannet_harness_v1.py) is planned for a future revision.

LaMAR conformance: The 22 LaMAR checks documented in §I.5 were performed as a manual schema-vs-schema analysis against the published wifi.txt and bt.txt field layouts and the radio-assisted retrieval workflow described by the benchmark. A scripted harness (scripts/lamar_harness_v1.py) is planned for a future revision.

No harness requires network access, a DDS runtime, or a dataset download. Implementers are encouraged to adapt the harnesses for additional reference datasets (e.g., Waymo Open, KITTI, Argoverse 2, RADIal, SubT-MRS, ScanNet, LaMAR) to validate coverage for sensor configurations or multi-agent scenarios not already covered.

Limitations

This testing validates schema expressiveness -- whether every dataset field has a lossless SpatialDDS mapping. It does not validate:

  • Wire interoperability -- actual DDS serialization/deserialization round-trips.
  • Performance -- throughput, latency, or memory footprint under real sensor loads.
  • Semantic correctness -- whether a particular producer's mapping preserves the intended meaning of each field.
  • Multi-dataset coverage -- datasets with different sensor configurations (e.g., solid-state lidar, event cameras, ultrasonic sensors) or deployment patterns (e.g., multi-floor hierarchical spaces, aerial-ground cooperation, dense pedestrian tracking) may surface additional gaps. S3E covers three-robot outdoor coordination; ScanNet covers single-room indoor scenes. Larger fleet sizes, degraded-communication environments, multi-floor buildings, and heterogeneous robot types (ground + aerial) remain untested.

These areas are appropriate targets for future conformance work.