docs/data_dictionary.md

Data Dictionary

EgoSpatial-CL is distributed as tar shards. Each shard contains one or more complete episode directories. Each episode is a synchronized sequence of front RGB images, semantic-segmentation images, LiDAR point clouds, ego-vehicle state/control rows, and episode metadata.

Dataset-Level Structure

shards/<Town>/<split>/<Town>_<split>_XXXXXX.tar

After extraction, each tar shard reconstructs episode directories using this layout:

<Town>/<split>/episode_XXXX/
  images/
  labels/
  lidar/
  meta.json
  state.csv
  index.csv

The source dataset before packaging uses the same episode contract under each town directory:

<Town>/episode_XXXX/
  images/
  labels/
  lidar/
  meta.json
  state.csv
  index.csv

Release-level episode identifiers:

• episode_id is a town-local integer identifier inherited from the original episode directory naming scheme.
• episode is the town-local directory name, for example episode_0000.
• episode_uid in benchmark_manifest.json is the unique release-level identifier formatted as <Town>/<split>/<episode>.
• global_episode_index in benchmark_manifest.json is a unique integer index assigned in manifest order.

Core Counts

7 towns
7 compact domains
10 spatial slots per town/domain pair
490 episodes
1,000 synchronized saved samples per episode
490,000 synchronized samples
392 train episodes
49 validation episodes
49 test episodes
401 tar shards

Towns

Town	Episodes	Notes
Town01_Opt	70	CARLA optimized town map.
Town02_Opt	70	CARLA optimized town map.
Town04_Opt	70	CARLA optimized town map.
Town05_Opt	70	CARLA optimized town map.
Town06	70	CARLA town map.
Town07	70	CARLA town map.
Town10HD_Opt	70	CARLA optimized high-definition town map.

Compact Domains

domain_id	domain_name	Weather condition	NPC variant	Requested vehicles	Requested walkers
0	clear__NPC00_none	clear	no NPC traffic	0	0
1	night__NPC00_none	night	no NPC traffic	0	0
2	fog__NPC00_none	fog	no NPC traffic	0	0
3	hardrain__NPC00_none	hard rain	no NPC traffic	0	0
4	clear__NPC11_veh_high	clear	vehicle traffic	50	0
5	hardrain__NPC11_veh_high	hard rain	vehicle traffic	50	0
6	clear__NPC31_both_high	clear	vehicle and pedestrian traffic	50	200

vehicles_spawned and walkers_spawned in meta.json may be lower than requested values when valid spawn points or navigation positions are unavailable in a town/episode.

Modalities

Path	Type	Cardinality per episode	Description
images/	PNG sequence	1,000 files	Front RGB camera frames. File names are zero-padded simulator frame IDs, for example 000804.png.
labels/	PNG sequence	1,000 files	Semantic-segmentation camera output aligned with RGB frames and saved after CARLA CityScapesPalette conversion.
lidar/	NumPy arrays	1,000 files	LiDAR point clouds stored as .npy arrays. File names match the synchronized simulator frame ID.
state.csv	CSV	1,000 data rows plus header	Ego-vehicle pose, kinematics, control state, IMU, and GNSS values.
index.csv	CSV	1,000 data rows plus header	Per-frame mapping between synchronized RGB, semantic-segmentation, and LiDAR files.
meta.json	JSON	1 file	Episode-level simulator, weather, ego vehicle, sensor, seed, and NPC metadata.

Synchronization

Each saved sample corresponds to one simulator frame selected after advancing capture_stride synchronous ticks. For release v1.0.0:

fixed_delta_seconds = 0.05
capture_stride_frames = 10
nominal saved-sample interval = 0.5 simulation seconds
samples_per_episode = 1,000

The frame value is the CARLA simulator frame ID. It is used consistently in state.csv, index.csv, image file names, semantic-label file names, and LiDAR file names.

index.csv

index.csv links every saved frame to the modality files inside the same episode directory.

Example:

frame,rgb_path,semseg_path,lidar_path
804,images/000804.png,labels/000804.png,lidar/000804.npy

Field	Type	Unit / format	Description
frame	integer	CARLA frame ID	Simulator frame associated with the synchronized sample.
rgb_path	string	relative path	Path from the episode directory to the front RGB PNG frame.
semseg_path	string	relative path	Path from the episode directory to the semantic-segmentation PNG frame. Empty only if semantic segmentation was disabled; in release v1.0.0 it is populated.
lidar_path	string	relative path	Path from the episode directory to the LiDAR .npy frame. Empty only if LiDAR was disabled; in release v1.0.0 it is populated.

state.csv

state.csv stores one synchronized ego-state row per saved sample. Values are recorded after the simulator tick that produced the synchronized sensor measurements.

Coordinate and rotation fields follow CARLA/Unreal conventions unless stated otherwise. World positions are in meters. Vehicle rotation angles are in degrees. Ego velocity is in meters per second. Ego acceleration and IMU acceleration are in meters per second squared. Actor angular velocity from get_angular_velocity() is in degrees per second; IMU gyroscope values are in radians per second.

Field	Type	Unit / range	Description
frame	integer	CARLA frame ID	Simulator frame associated with this saved sample.
sim_time	float	seconds	Elapsed simulation time from the CARLA world snapshot timestamp.
ts_platform	float	Unix timestamp, seconds	Host/platform wall-clock time when the row was written. Use sim_time for simulation-time analyses.
ego_x	float	meters	Ego vehicle world-frame x position.
ego_y	float	meters	Ego vehicle world-frame y position.
ego_z	float	meters	Ego vehicle world-frame z position.
ego_roll	float	degrees	Ego vehicle roll angle from the CARLA transform.
ego_pitch	float	degrees	Ego vehicle pitch angle from the CARLA transform.
ego_yaw	float	degrees	Ego vehicle yaw/heading angle from the CARLA transform.
vel_x	float	m/s	Ego vehicle world-frame velocity x component.
vel_y	float	m/s	Ego vehicle world-frame velocity y component.
vel_z	float	m/s	Ego vehicle world-frame velocity z component.
acc_x	float	m/s^2	Ego vehicle world-frame acceleration x component returned by the actor state.
acc_y	float	m/s^2	Ego vehicle world-frame acceleration y component returned by the actor state.
acc_z	float	m/s^2	Ego vehicle world-frame acceleration z component returned by the actor state.
angvel_x	float	deg/s	Ego vehicle angular-velocity x component returned by the actor state.
angvel_y	float	deg/s	Ego vehicle angular-velocity y component returned by the actor state.
angvel_z	float	deg/s	Ego vehicle angular-velocity z component returned by the actor state.
throttle	float	[0, 1]	Applied vehicle throttle command.
steer	float	[-1, 1]	Applied steering command. Negative and positive values correspond to opposite steering directions in CARLA's vehicle-control convention.
brake	float	[0, 1]	Applied brake command.
hand_brake	integer	0 or 1	Whether the hand brake is active.
reverse	integer	0 or 1	Whether reverse gear/direction is active.
gear	integer	CARLA gear value	Current vehicle gear reported by the control object.
imu_accel_x	float	m/s^2	IMU-frame linear acceleration x component from sensor.other.imu.
imu_accel_y	float	m/s^2	IMU-frame linear acceleration y component from sensor.other.imu.
imu_accel_z	float	m/s^2	IMU-frame linear acceleration z component from sensor.other.imu; includes gravity according to CARLA IMU behavior.
imu_gyro_x	float	rad/s	IMU-frame angular velocity x component from sensor.other.imu.
imu_gyro_y	float	rad/s	IMU-frame angular velocity y component from sensor.other.imu.
imu_gyro_z	float	rad/s	IMU-frame angular velocity z component from sensor.other.imu.
imu_compass	float	radians	Compass orientation with respect to CARLA north.
gnss_lat	float	degrees	GNSS latitude from sensor.other.gnss.
gnss_lon	float	degrees	GNSS longitude from sensor.other.gnss.
gnss_alt	float	meters	GNSS altitude from sensor.other.gnss.

Image Files

RGB frames are saved as PNG files after converting CARLA camera raw data from BGRA to RGB. The release uses:

image width = 800 pixels
image height = 600 pixels
field of view = 90 degrees
sensor tick = 0.05 seconds
camera transform relative to ego = x 1.5 m, y 0.0 m, z 1.7 m, pitch -5 deg, yaw 0 deg, roll 0 deg

Semantic-Segmentation Files

Semantic-segmentation frames are saved as PNG files aligned with the RGB camera. During collection, the CARLA semantic camera output is converted with carla.ColorConverter.CityScapesPalette before saving. Therefore, release labels are color-encoded semantic images rather than single-channel integer class-index masks.

Users needing class-index tensors should convert the palette colors to class IDs consistently with CARLA's semantic-label definitions for the simulator version used in the release.

LiDAR Files

Each LiDAR file is a NumPy .npy array written with allow_pickle=False.

shape = (N, 4)
dtype = float32
columns = x, y, z, intensity

Column	Type	Unit / range	Description
x	float32	meters	Point x coordinate in the LiDAR sensor frame.
y	float32	meters	Point y coordinate in the LiDAR sensor frame.
z	float32	meters	Point z coordinate in the LiDAR sensor frame.
intensity	float32	unitless	LiDAR return intensity reported by CARLA.

N can vary across frames because it depends on ray returns, scene geometry, and simulator conditions.

Release v1.0.0 LiDAR generation parameters:

sensor type = sensor.lidar.ray_cast
transform relative to ego = x 0.0 m, y 0.0 m, z 2.0 m
range = 50.0 m
points_per_second = 100000
channels = 32
upper_fov = 10.0 deg
lower_fov = -30.0 deg
rotation_frequency = 20 Hz
sensor_tick = 0.05 seconds

meta.json

meta.json stores episode-level metadata. The table below describes the principal fields observed in release v1.0.0.

Field	Type	Description
created_utc	string	UTC timestamp when the episode metadata file was written.
carla_version	string	Simulator version used for generation.
host	string	CARLA server host used during collection.
port	integer	CARLA RPC port used during collection.
map	string	CARLA map loaded for the episode.
seed	integer	Episode seed. Release rule: seed = 12345 + 17 * episode_id.
fixed_delta_seconds	float	Fixed synchronous simulation step in seconds.
synchronous_mode	boolean	Whether synchronous mode was enabled.
no_rendering_mode_capture	boolean	Whether CARLA no-rendering mode was used during capture. It is false for release v1.0.0 because camera data are recorded.
weather	string	High-level weather/domain label used by the collector.
weather_params	object	CARLA weather parameters actually recorded from the world.
weather_overrides	object or null	Explicit weather overrides applied beyond the named preset, if any.
ego	object	Ego-vehicle blueprint, spawn transform, autopilot flag, and traffic-manager settings.
npcs	object	Requested/spawned background vehicles and walkers plus NPC traffic-manager parameters.
sensors	object	Sensor availability, ticks, camera geometry, and camera transform.
shutdown	object	Collector shutdown/parking settings used after episode generation.

meta.json.weather_params

Field	Type	Unit / range	Description
cloudiness	float	percentage-like CARLA value	Cloud coverage parameter.
precipitation	float	percentage-like CARLA value	Rain intensity parameter.
precipitation_deposits	float	percentage-like CARLA value	Wet-road/rain-deposit parameter.
wetness	float	percentage-like CARLA value	Surface wetness parameter.
wind_intensity	float	percentage-like CARLA value	Wind intensity parameter.
fog_density	float	percentage-like CARLA value	Fog density parameter.
fog_distance	float	meters / CARLA world parameter	Fog distance parameter.
fog_falloff	float	CARLA world parameter	Fog falloff parameter.
sun_altitude_angle	float	degrees	Sun altitude angle. Negative values correspond to night-like lighting.
sun_azimuth_angle	float	degrees	Sun azimuth angle.
scattering_intensity	float	CARLA world parameter	Atmospheric scattering intensity.
mie_scattering_scale	float	CARLA world parameter	Mie scattering scale.
rayleigh_scattering_scale	float	CARLA world parameter	Rayleigh scattering scale.
dust_storm	float	percentage-like CARLA value	Dust storm parameter.

meta.json.ego

Field	Type	Description
blueprint	string	Ego vehicle actor blueprint used in the simulator.
transform_spawn.location.{x,y,z}	float	Ego spawn location in world coordinates, meters.
transform_spawn.rotation.{pitch,roll,yaw}	float	Ego spawn rotation in degrees.
autopilot	boolean	Whether CARLA Traffic Manager autopilot controlled the ego vehicle.
traffic_manager_port	integer or null	Traffic Manager port used by the ego autopilot.
traffic_manager.distance_to_leading	float	Target distance to leading vehicle, meters.
traffic_manager.ignore_lights_pct	float	Percentage probability for ignoring traffic lights.
traffic_manager.ignore_signs_pct	float	Percentage probability for ignoring traffic signs.
traffic_manager.ignore_walkers_pct	float	Percentage probability for ignoring walkers.
traffic_manager.speed_diff_pct	float	Traffic Manager speed difference percentage relative to speed limits.
traffic_manager.auto_lane_change	boolean	Whether automatic lane changing was enabled for the ego vehicle.
traffic_manager.hybrid_physics	boolean	Whether Traffic Manager hybrid physics was enabled.
traffic_manager.hybrid_radius	float	Hybrid physics radius, meters.

meta.json.npcs

Field	Type	Description
npc_seed	integer	Deterministic seed used for NPC vehicle and walker spawning.
vehicles_requested	integer	Number of NPC vehicles requested by the domain configuration.
vehicles_spawned	integer	Number of NPC vehicles successfully spawned.
walkers_requested	integer	Number of pedestrians requested by the domain configuration.
walkers_spawned	integer	Number of pedestrians successfully spawned.
pedestrians_cross_factor	float	CARLA pedestrian crossing factor used for the episode.
npc_vehicle_filter	string	Blueprint filter used for NPC vehicle selection.
npc_min_dist_to_ego	float	Minimum spawn distance from ego vehicle, meters.
tm_npc_speed_diff_pct	float	Traffic Manager speed difference percentage for NPC vehicles.
tm_npc_auto_lane_change	boolean	Whether automatic lane changing was enabled for NPC vehicles.
walker_speed_min	float	Minimum walker speed, m/s.
walker_speed_max	float	Maximum walker speed, m/s.

meta.json.sensors

Field	Type	Description
rgb.w	integer	RGB image width in pixels.
rgb.h	integer	RGB image height in pixels.
rgb.fov	float	RGB camera field of view in degrees.
rgb.tick	float	RGB sensor tick in seconds.
rgb.transform.location.{x,y,z}	float	RGB camera position relative to the ego vehicle, meters.
rgb.transform.rotation.{pitch,roll,yaw}	float	RGB camera rotation relative to the ego vehicle, degrees.
semseg	boolean	Whether semantic-segmentation images were recorded.
lidar	boolean	Whether LiDAR point clouds were recorded.
imu.tick	float	IMU sensor tick in seconds.
gnss.tick	float	GNSS sensor tick in seconds.

Release Manifest Files

dataset_info.json

Dataset-level summary used for release metadata.

Field	Type	Description
schema_version	string	Version of the release metadata schema.
name	string	Machine-readable dataset release name.
pretty_name	string	Human-readable short dataset name.
title	string	Full citable dataset title.
version	string	Dataset release version.
license	string	Hugging Face license identifier for the dataset.
repository	string	Public Hugging Face dataset URL.
code_repository	string	Public code repository URL.
doi_status	string	DOI state; remains pending until final release freeze.
release_status	string	Release preparation state. release_candidate_v1_0_0 indicates final local review before DOI and public freeze.
format	string	Distribution format, tar_shards for v1.0.0.
episode_order	string	Episode ordering convention used by the packed release.
num_towns	integer	Number of towns.
num_domains_per_town	integer	Number of compact domains per town.
num_base_tasks_expected	integer	Number of town/domain combinations.
num_episodes	integer	Number of complete episodes.
expected_samples_per_episode	integer	Expected saved synchronized samples per episode.
expected_samples_total	integer	Expected total synchronized samples.
num_shards	integer	Number of tar shards.
total_shard_bytes	integer	Sum of tar-shard byte sizes.
splits	object	Train/validation/test episode counts using technical labels train, val, and test.
split_counts_by_town	object	Split counts for each town.
split_aliases	object	Human-readable aliases for technical split labels, for example val = validation.
episode_identifier_scope	object	Scope and uniqueness rules for episode identifiers.
modalities	array	Modalities included in the release.
simulator	string	Simulator version.
provenance	object	Generation provenance summary.
notes	array	Release notes.

benchmark_manifest.json

Episode-level and shard-level manifest. Despite the field name benchmark_name, the v1.0.0 public artifact is described primarily as a dataset with benchmark-ready splits.

Field	Type	Description
benchmark_name	string	Machine-readable release/protocol identifier kept for compatibility.
dataset_release_name	string	Machine-readable dataset release identifier.
dataset_name	string	Human-readable dataset name.
title	string	Full citable title.
version	string	Dataset release version.
license	string	Dataset license identifier.
repository	string	Public Hugging Face dataset URL.
code_repository	string	Public code repository URL.
doi_status	string	DOI state.
release_status	string	Release preparation state.
compact_domains	array	Domain definitions and requested NPC counts.
episode_order	string	Episode ordering convention.
episodes	array	One record per episode with town, split, domain, spatial slot, spawn index, and weather summary.
shards	array	One record per tar shard with path, size, checksum, split, town, and included episodes.
towns	array	Towns included in the release.
warnings	array	Packaging/manifest warnings; empty for the validated v1.0.0 metadata state.
provenance	object	Generation provenance summary.

benchmark_manifest.json.episodes[]

Each object in episodes describes one complete episode. Release-level uniqueness is provided by episode_uid and global_episode_index.

Field	Type	Description
global_episode_index	integer	Unique release-level integer index assigned in manifest order, from 0 to 489.
episode_uid	string	Unique release-level episode identifier formatted as <Town>/<split>/<episode>.
episode	string	Town-local episode directory name, for example episode_0000.
episode_id	integer	Town-local episode index inherited from the generated episode naming scheme, from 0 to 69 within each town.
town	string	CARLA town for the episode.
split	string	Technical split label: train, val, or test.
domain_id	integer	Compact domain identifier from 0 to 6.
domain_name	string	Compact domain name combining condition and NPC variant.
condition	string	Public condition label such as clear, night, fog, or hardrain.
npc_variant	string	Compact NPC traffic variant.
spatial_slot	integer	Slot index within the 10 selected spawn slots for the town/domain pair.
spawn_index	integer	CARLA spawn-point index used for the ego vehicle.
meta_summary	object	Selected episode metadata copied from meta.json, including map, seed, weather label, and key weather parameters.

benchmark_manifest.json.shards[]

Each object in shards describes one packed tar shard.

Field	Type	Description
path	string	Relative path to the tar shard inside the dataset repository.
sha256	string	SHA256 checksum of the tar shard payload.
bytes	integer	Tar shard size in bytes.
town	string	Town represented by the shard.
split	string	Technical split label represented by the shard: train, val, or test.
num_episodes	integer	Number of complete episode directories stored in the shard.
episodes	array	Episode directory names included in the shard.

shards_manifest.csv

Shard-level tabular manifest.

Field	Type	Description
path	string	Relative path to a tar shard in the HF dataset repository.
sha256	string	SHA256 checksum of the tar shard.
bytes	integer	Tar shard size in bytes.
town	string	Town represented by the shard.
split	string	Split represented by the shard: train, val, or test.
num_episodes	integer	Number of complete episodes stored in the shard.
episodes	string	Pipe-separated list of episode directory names included in the shard.

checksums.sha256

Standard SHA256 checksum file for tar shards.

<sha256>  <relative-shard-path>

spatial_splits.json

Frozen spatial split by town and split. Values are CARLA spawn indices assigned to train, validation, or test.

Missing Values And Caveats

• Release v1.0.0 expects all synchronized modality paths in index.csv to be populated.
• ts_platform is wall-clock time and should not be used as a simulation-time axis.
• Semantic segmentation labels are palette-color PNGs, not class-index rasters.
• This release does not include dense depth, optical flow, 2D/3D bounding boxes, tracking IDs, or multi-view camera annotations.
• The dataset contains simulated data. Simulator-specific coordinate conventions, rendering behavior, traffic-manager behavior, and semantic labels should be reported when comparing against real-world datasets.