Datasets:
Sensor And Simulation Configuration
This document records the sensor, simulator, traffic, weather, and synchronization settings used for the EgoSpatial-CL v1.0.0 dataset release. Values are derived from the current generation scripts, release manifests, and episode-level meta.json files.
Release Summary
Dataset: EgoSpatial-CL
Version: v1.0.0
Simulator: CARLA 0.9.15
Execution mode: Docker, headless/offscreen rendering
Synchronous mode: enabled
Fixed delta seconds: 0.05
Capture stride: 10 simulator ticks
Saved samples per episode: 1,000
Nominal saved-sample interval: 0.5 simulation seconds
Episodes: 490
Synchronized samples: 490,000
The dataset records rendered camera data; therefore CARLA no_rendering_mode_capture is false for all release episodes.
Generation Runtime
The v1.0.0 generation audit records the following container image digest:
sha256:1b5417ab03a91d2bad581d41fabeb91a8f3836130d7a112bcc8b4d2dac7dcfcd
Episode metadata in the frozen dataset records:
host = 127.0.0.1
port = 2000
synchronous_mode = true
fixed_delta_seconds = 0.05
no_rendering_mode_capture = false
These host/port values describe the local collection setup, not a requirement for using the released dataset.
Temporal Synchronization
All sensors were configured with sensor_tick = 0.05, matching the fixed simulator step. A saved sample is produced after advancing the simulation by the capture stride:
fixed_delta_seconds = 0.05 seconds
sensor_tick = 0.05 seconds
capture_stride = 10 ticks
saved-sample interval = 0.05 x 10 = 0.5 simulation seconds
For each saved sample, state.csv, index.csv, RGB, semantic segmentation, and LiDAR share the same CARLA frame identifier.
Sensor Suite
| Sensor / signal | CARLA actor or source | Stored output | Release status |
|---|---|---|---|
| Front RGB camera | sensor.camera.rgb |
images/<frame>.png |
enabled |
| Semantic-segmentation camera | sensor.camera.semantic_segmentation |
labels/<frame>.png |
enabled |
| LiDAR | sensor.lidar.ray_cast |
lidar/<frame>.npy |
enabled |
| IMU | sensor.other.imu |
columns in state.csv |
enabled |
| GNSS | sensor.other.gnss |
columns in state.csv |
enabled |
| Ego pose/kinematics/control | ego vehicle actor | columns in state.csv |
enabled |
The release does not include dense depth, optical flow, 2D bounding boxes, 3D bounding boxes, tracking IDs, multi-view cameras, radar, or instance segmentation.
Front RGB Camera
The front RGB camera is attached to the ego vehicle.
CARLA blueprint: sensor.camera.rgb
image width: 800 pixels
image height: 600 pixels
field of view: 90 degrees
sensor_tick: 0.05 seconds
postprocess effects: disabled when supported by the CARLA blueprint
Transform relative to the ego vehicle:
| Component | Value | Unit |
|---|---|---|
location.x |
1.5 | meters |
location.y |
0.0 | meters |
location.z |
1.7 | meters |
rotation.pitch |
-5.0 | degrees |
rotation.yaw |
0.0 | degrees |
rotation.roll |
0.0 | degrees |
Saved format:
directory: images/
filename pattern: %06d.png using the CARLA frame id
color conversion: CARLA raw BGRA -> RGB
file format: PNG
Semantic-Segmentation Camera
The semantic-segmentation camera is aligned with the front RGB camera.
CARLA blueprint: sensor.camera.semantic_segmentation
image width: 800 pixels
image height: 600 pixels
field of view: 90 degrees
sensor_tick: 0.05 seconds
transform: same as front RGB camera
Saved format:
directory: labels/
filename pattern: %06d.png using the CARLA frame id
conversion before saving: carla.ColorConverter.CityScapesPalette
file format: PNG
Important interpretation note: the release stores color-encoded semantic PNGs after CityScapesPalette conversion. They are not single-channel integer class-index rasters. Users requiring class-index masks should map palette colors back to semantic classes using the CARLA semantic-label definitions for CARLA 0.9.15.
LiDAR
LiDAR point clouds are attached to the ego vehicle and saved once per synchronized sample.
CARLA blueprint: sensor.lidar.ray_cast
sensor_tick: 0.05 seconds
rotation_frequency: 20 Hz
points_per_second: 100000
channels: 32
range: 50.0 meters
upper_fov: 10.0 degrees
lower_fov: -30.0 degrees
Transform relative to the ego vehicle:
| Component | Value | Unit |
|---|---|---|
location.x |
0.0 | meters |
location.y |
0.0 | meters |
location.z |
2.0 | meters |
rotation.pitch |
0.0 | degrees |
rotation.yaw |
0.0 | degrees |
rotation.roll |
0.0 | degrees |
Saved format:
directory: lidar/
filename pattern: %06d.npy using the CARLA frame id
NumPy dtype: float32
NumPy shape: (N, 4)
columns: x, y, z, intensity
allow_pickle: false
N can vary across frames because it depends on the simulated scene geometry and valid LiDAR returns. Point coordinates are stored in the LiDAR sensor frame.
IMU
The IMU is attached to the ego vehicle at the actor origin:
CARLA blueprint: sensor.other.imu
sensor_tick: 0.05 seconds
relative transform: location (0, 0, 0), rotation (0, 0, 0)
stored in: state.csv
Stored IMU fields:
| Field group | Columns | Unit |
|---|---|---|
| Linear acceleration | imu_accel_x, imu_accel_y, imu_accel_z |
m/s^2 |
| Angular velocity | imu_gyro_x, imu_gyro_y, imu_gyro_z |
rad/s |
| Compass | imu_compass |
radians |
The IMU acceleration follows CARLA IMU behavior and includes gravity in the sensor-frame measurement.
GNSS
The GNSS sensor is attached to the ego vehicle at the actor origin:
CARLA blueprint: sensor.other.gnss
sensor_tick: 0.05 seconds
relative transform: location (0, 0, 0), rotation (0, 0, 0)
stored in: state.csv
Stored GNSS fields:
| Field | Unit | Description |
|---|---|---|
gnss_lat |
degrees | Latitude reported by CARLA GNSS. |
gnss_lon |
degrees | Longitude reported by CARLA GNSS. |
gnss_alt |
meters | Altitude reported by CARLA GNSS. |
Ego Vehicle
All release episodes use the same ego vehicle blueprint:
ego blueprint: vehicle.tesla.model3
autopilot: true
Traffic Manager port: 8000
Ego Traffic Manager configuration recorded across all release episodes:
| Setting | Value | Unit / type |
|---|---|---|
distance_to_leading |
3.5 | meters |
speed_diff_pct |
60.0 | percent |
ignore_lights_pct |
0.0 | percent |
ignore_signs_pct |
0.0 | percent |
ignore_walkers_pct |
0.0 | percent |
auto_lane_change |
false | boolean |
hybrid_physics |
false | boolean |
hybrid_radius |
70.0 | meters |
The ego vehicle state and control values are recorded in state.csv.
NPC Traffic And Walkers
NPC settings depend on the compact domain. Requested NPC counts are:
| Domain | Requested vehicles | Requested walkers |
|---|---|---|
clear__NPC00_none |
0 | 0 |
night__NPC00_none |
0 | 0 |
fog__NPC00_none |
0 | 0 |
hardrain__NPC00_none |
0 | 0 |
clear__NPC11_veh_high |
50 | 0 |
hardrain__NPC11_veh_high |
50 | 0 |
clear__NPC31_both_high |
50 | 200 |
Common NPC configuration:
npc_vehicle_filter = vehicle.*
npc_min_dist_to_ego = 15.0 meters
tm_npc_speed_diff_pct = 20.0
tm_npc_auto_lane_change = false
Walker settings for pedestrian domains:
pedestrians_cross_factor = 0.6
walker_speed_min = 1.2 m/s
walker_speed_max = 2.0 m/s
For no-pedestrian domains, pedestrians_cross_factor is recorded as the collector default 0.1, and walker speeds are recorded as defaults even though no walkers are requested.
vehicles_spawned and walkers_spawned can be lower than requested counts because spawning depends on available map spawn points, minimum distance constraints, and valid navigation locations.
Weather And Domain Parameters
The public compact domain names use release-level labels (clear, night, fog, hardrain, and NPC variants). The raw meta.json.weather field records the collector preset name. Some release domains, especially night and fog, are defined by weather-parameter overrides in addition to the preset.
Representative recorded weather parameters by compact domain:
| Domain | meta.json.weather |
cloudiness |
precipitation |
precipitation_deposits |
wind_intensity |
fog_density |
fog_distance |
fog_falloff |
sun_altitude_angle |
|---|---|---|---|---|---|---|---|---|---|
clear__NPC00_none |
clear |
5.0 | 0.0 | 0.0 | 10.0 | 2.0 | 0.75 | 0.1 | 45.0 |
night__NPC00_none |
clear |
25.0 | 0.0 | 0.0 | 10.0 | 2.0 | 0.75 | 0.1 | -10.0 |
fog__NPC00_none |
cloudy |
60.0 | 0.0 | 0.0 | 10.0 | 50.0 | 30.0 | 1.0 | 25.0 |
hardrain__NPC00_none |
hardrain |
100.0 | 100.0 | 90.0 | 100.0 | 7.0 | 0.75 | 0.1 | 45.0 |
clear__NPC11_veh_high |
clear |
5.0 | 0.0 | 0.0 | 10.0 | 2.0 | 0.75 | 0.1 | 45.0 |
hardrain__NPC11_veh_high |
hardrain |
100.0 | 100.0 | 90.0 | 100.0 | 7.0 | 0.75 | 0.1 | 45.0 |
clear__NPC31_both_high |
clear |
5.0 | 0.0 | 0.0 | 10.0 | 2.0 | 0.75 | 0.1 | 45.0 |
Additional recorded weather fields include:
wetness
sun_azimuth_angle
scattering_intensity
mie_scattering_scale
rayleigh_scattering_scale
dust_storm
These fields are documented in docs/data_dictionary.md.
Towns And Spatial Slots
The release includes seven towns:
Town01_Opt
Town02_Opt
Town04_Opt
Town05_Opt
Town06
Town07
Town10HD_Opt
Each town uses 10 frozen spawn slots reused across all compact domains:
8 train slots
1 validation slot
1 test slot
The exact spawn indices are stored in spatial_splits.json and episode-level records in benchmark_manifest.json.
Excluded from release v1.0.0:
Town03
Town03_Opt
Town12
Town12 was excluded because the local CARLA 0.9.15 headless Docker setup was unstable when sensors were attached to the ego vehicle. The other excluded towns are outside the frozen v1.0.0 release scope.
Collection Stability Parameters
The final collection workflow used conservative synchronization and shutdown settings:
warmup_ticks = 20
post_load_sleep = 2.0 seconds
map_load_warmup_ticks = 40
sensor_timeout = 20.0 seconds
rpc_timeout = 30.0 seconds
shutdown_sleep = 4.0 seconds
park_server = true
park_fixed_dt = 0.05 seconds
These settings are provenance for dataset generation. Users loading the released data do not need to reproduce the live simulator runtime unless regenerating the dataset.
Interpretation Notes
state.csvactor pose, velocity, acceleration, angular velocity, control, IMU, and GNSS values are synchronized to saved sample frames.- RGB and semantic-segmentation images are both 800 x 600 PNG files aligned by simulator frame.
- Semantic-segmentation PNGs are palette-color images, not integer class-index masks.
- LiDAR files contain variable-length point clouds with shape
(N, 4)and dtypefloat32. - The dataset is simulated; CARLA-specific coordinate, rendering, semantic-label, vehicle-control, and Traffic Manager behavior should be acknowledged in downstream studies.