shdragron/pdm-lite-train-R1_Normal

PDM-Lite R1_Normal

Expert driving in CARLA 0.9.15 on the full route set of carla_garage's PDM-Lite data (CARLA Leaderboard 2.0 scenarios, the scenario set Bench2Drive also uses), driven by the rule-based PDM-Lite expert and recorded with two camera rigs at once: BEVAD's six nuScenes-style cameras and the eight-camera PCCR R1 rig, plus a 32-channel LiDAR and the simulator's labels at 5 Hz.

Collection in progress (2026-10-30): 4467 of 5545 routes attempted, 4100 successful routes uploaded so far; new routes are added as they finish. The numbers below are those of 2026-10-30.

Routes (successful episodes)	4,100
Frames (5 Hz)	613,571 (34.1 h of driving)
Images	8,589,994 (14 per frame: 6 BEVAD + 8 R1), JPEG 800 x 448
Size	383 GB in 4,100 tar files
Scenario types	38 scenario types + routes without a scenario (noScenarios)
Towns	Town01, Town02, Town03, Town04, Town05, Town10HD, Town12, Town13
Weather	22 CARLA presets; 48% of the routes in night mode (vehicle lights on, carla_garage's rule: low sun, heavy cloud or fog)
Ego vehicle	vehicle.lincoln.mkz_2020

The Dataset Viewer above shows the preview table: one key frame of up to four routes per scenario, with the front camera, both surround views and the expert's control. index.csv lists every uploaded route (the viewer reads one format per repository, so it is not a viewer table).

Samples

All six cameras of BEVAD (rgb) and all eight of R1 (rgb_r1) at the same instant; in the middle of the R1 view the LiDAR (grey) and the world-state boxes in a bird's-eye view, forward up (white: ego, blue: vehicles, orange: pedestrians, aqua: bicycles and motorcycles, grey outline: static objects; rings every 10 m).

DynamicObjectCrossing · Town05 · WetCloudyNoon (data/lb1_split/DynamicObjectCrossing/Town05_Scenario3_44.tar, frame 203)

More scenes (BEVAD cameras and the bird's-eye view, ±40 m):

BEVAD cameras	LiDAR + boxes
SignalizedJunctionLeftTurn · Town12 · HardRainNight · frame 73	data/50x38_Town12/SignalizedJunctionLeftTurn/1002_0.tar
ConstructionObstacle · Town12 · MidRainyNoon · frame 58	data/50x38_Town12/ConstructionObstacle/2266_0.tar
HighwayCutIn · Town13 · ClearSunset · frame 40	data/50x36_Town13/HighwayCutIn/1283_16.tar

The samples/ folder has the same views for all six scenes.

Sensors

Positions in CARLA's vehicle frame of the actor (x forward, y right, z up; metres, degrees). Every camera is rendered at the listed resolution and stored at 800 x 448, JPEG quality 75.

camera	folder	x, y, z	yaw	pitch	FoV	rendered
CAM_FRONT	rgb	0.60, 0.00, 1.60	0	5	90	1600 x 900
CAM_FRONT_LEFT	rgb	0.27, -0.55, 1.60	-55	0	90	1600 x 900
CAM_FRONT_RIGHT	rgb	0.27, 0.55, 1.60	55	0	90	1600 x 900
CAM_BACK	rgb	-2.00, 0.00, 1.60	180	-5	90	1600 x 900
CAM_BACK_LEFT	rgb	-0.32, -0.55, 1.60	-110	-5	90	1600 x 900
CAM_BACK_RIGHT	rgb	-0.32, 0.55, 1.60	110	-5	90	1600 x 900
CAM_BUMPER_FRONT	rgb_r1	2.36, 0.00, 0.68	0	0	65	1280 x 720
CAM_FENDER_FRONT_RIGHT	rgb_r1	1.09, 0.99, 0.72	41	0	65	1280 x 720
CAM_SIDE_RIGHT	rgb_r1	-0.22, 0.74, 1.40	90	0	65	1280 x 720
CAM_FENDER_BACK_RIGHT	rgb_r1	-1.57, 0.99, 0.84	135	0	65	1280 x 720
CAM_BUMPER_BACK	rgb_r1	-2.44, 0.00, 0.64	180	0	65	1280 x 720
CAM_FENDER_BACK_LEFT	rgb_r1	-1.63, -0.99, 0.86	-135	0	65	1280 x 720
CAM_SIDE_LEFT	rgb_r1	-0.20, -0.78, 1.40	-90	0	65	1280 x 720
CAM_FENDER_FRONT_LEFT	rgb_r1	1.23, -0.95, 0.78	-43	0	65	1280 x 720

LiDAR LIDAR_TOP (point_cloud/, LAZ): 32 channels, range 100 m, vertical FoV -30° to +10°, at (-0.46, 0.0, 1.86). IMU and GNSS are part of the episode metadata. The full rig with every attribute is meta/rig_r1_normal.json.

What a frame contains

Each route is one episode in bevad-sim's EpisodeIo format: ep_data.zstd (zstandard-compressed pickle of a CoreContainer) next to the sensor folders. Arrays have a leading batch axis of 1, then time (5 Hz), then the slot.

field	content
world_state	every tracked object incl. the ego: 4x4 pose (world), size, category (car, truck, bus, ambulance, pedestrian, bicycle, motorcycle, construction, roadway item), attribute (parking, driving, emergency, open door, ...), track id, visibility, validity
tce	traffic lights and signs: pose, size, control box, state (red / yellow / green / off, sign type)
odometry	GNSS, compass, angular velocity, acceleration
action.simulator_control	the expert's throttle, steer, brake per frame
routing_information	target route and route commands, navigation goal per frame
rgb, rgb_r1, point_cloud	per camera / LiDAR: file per frame (named by simulator frame id), timestamps, intrinsics, extrinsics (ego -> sensor)
episode_meta	scenario type, town, route file, weather parameters, ego size

Labels and point clouds are in the ego frame of bevad-sim: x forward, y left, z up, origin at the centre of the ego's bounding box (0.75 m above the ground for this car). Static map data is not stored; the towns are CARLA's.

Layout

data/<set>/<scenario>/<route>.tar      one route; unpacks to <set>/<scenario>/<route>/
    summary.json                         seed, frames, route score, sizes
    collect.log                          the collection log of that route
    <route>-<id>/ep_data.zstd            labels (see above)
    <route>-<id>/rgb/CAM_*/<frame>.jpg   BEVAD cameras
    <route>-<id>/rgb_r1/CAM_*/<frame>.jpg  R1 cameras
    <route>-<id>/point_cloud/LIDAR_TOP/<frame>.laz
index.csv                              uploaded routes: order, set, scenario, town, weather, night, seed, frames, score, size, tar
meta/rig_r1_normal.json                sensor rig
meta/routes_index.csv                  all 5545 route files with seed and weather
preview/, samples/                     the viewer table and the figures of this card

Loading

from huggingface_hub import hf_hub_download
import tarfile

p = hf_hub_download("shdragron/pdm-lite-train-R1_Normal", "data/lb1_split/DynamicObjectCrossing/Town05_Scenario3_44.tar", repo_type="dataset")
tarfile.open(p).extractall("pdm_lite")

# pip install -e <github.com/dmholtz/bevad>/bevad-sim    (bevad-sim reads the episode format)
import glob
from bevad_sim.data_interface.episode_io import EpisodeIo

ep_dir = glob.glob("pdm_lite/lb1_split/DynamicObjectCrossing/Town05_Scenario3_44/*/")[0]
ep = EpisodeIo.read_episode(ep_dir, load_payload=False)       # labels only; images and LiDAR stay on disk
ws = ep.world_state
print(ws.transform.shape)                                     # (1, frames, objects, 4, 4)
cam = ep.rgb.sensor_names.index("CAM_FRONT")
frame_id = int(ep.rgb.frame_ids[0, 0, cam])                   # -> {ep_dir}rgb/CAM_FRONT/<frame_id>.jpg

ep_data.zstd is a pickle: open it only from this repository (or other sources you trust).

Collection

• Expert: PDM-Lite (carla_garage, team_code/autopilot.py) through BEVAD's bevad-sim port, with carla_garage's red-light stop distance (6 m) and vehicle lights in night mode restored, so it drives as carla_garage's original.
• Routes: every route file of carla_garage's PDM-Lite data set (50x38_Town12, 50x36_Town13, lb1_split; 5545 files, one route each), in carla_garage's shuffled order (seed 42); the traffic-manager seed counts up by one per route.
• Weather: one of the first 22 CARLA presets per route, as carla_garage's shuffle_weather, fixed by the route seed.
• Episodes: simulator at 20 Hz, recorded every 4th tick (5 Hz). A route that fails (scenario could not be spawned, simulator crash, 90 min time-out) gets one more try on a fresh server; only routes whose episode succeeded are here.

Distribution

Sets and towns

set	town	routes	frames	hours
50x36_Town13	Town13	1376	221,095	12.3
50x38_Town12	Town12	1428	218,363	12.1
lb1_split	Town01	90	10,333	0.6
lb1_split	Town02	58	6,606	0.4
lb1_split	Town03	213	30,771	1.7
lb1_split	Town04	399	45,131	2.5
lb1_split	Town05	426	63,029	3.5
lb1_split	Town10HD	110	18,243	1.0

Scenarios (39)

scenario	routes	frames	hours
VehicleTurningRoute	439	81,148	4.5
noScenarios	290	28,416	1.6
SignalizedJunctionLeftTurn	269	21,195	1.2
ControlLoss	209	39,031	2.2
DynamicObjectCrossing	191	38,722	2.2
SignalizedJunctionRightTurn	184	15,288	0.8
OppositeVehicleRunningRedLight	182	14,680	0.8
ParkedObstacle	85	18,436	1.0
AccidentTwoWays	83	17,047	0.9
NonSignalizedJunctionRightTurn	83	8,695	0.5
ParkingExit	82	11,463	0.6
HazardAtSideLaneTwoWays	82	18,565	1.0
NonSignalizedJunctionLeftTurn	81	8,909	0.5
ParkedObstacleTwoWays	81	12,773	0.7
PriorityAtJunction	81	9,756	0.5
PedestrianCrossing	81	15,040	0.8
ParkingCutIn	78	13,786	0.8
OppositeVehicleTakingPriority	78	8,075	0.4
BlockedIntersection	78	11,042	0.6
VehicleTurningRoutePedestrian	78	20,724	1.2
InvadingTurn	77	8,638	0.5
VehicleOpensDoorTwoWays	77	12,451	0.7
StaticCutIn	76	8,175	0.5
InterurbanActorFlow	76	8,240	0.5
HardBreakRoute	75	13,504	0.8
InterurbanAdvancedActorFlow	75	7,923	0.4
HighwayExit	74	11,628	0.6
ConstructionObstacle	73	11,031	0.6
YieldToEmergencyVehicle	71	11,246	0.6
ConstructionObstacleTwoWays	71	13,508	0.8
ParkingCrossingPedestrian	71	12,345	0.7
MergerIntoSlowTraffic	69	11,952	0.7
EnterActorFlow	69	7,966	0.4
Accident	68	9,709	0.5
MergerIntoSlowTrafficV2	67	13,591	0.8
HazardAtSideLane	65	17,021	0.9
HighwayCutIn	61	12,439	0.7
CrossingBicycleFlow	44	5,524	0.3
EnterActorFlowV2	26	3,889	0.2

Weather (22 presets)

weather	routes	frames
ClearNoon	212	30,494
SoftRainSunset	211	31,641
Default	208	31,665
WetCloudyNoon	206	31,978
SoftRainNoon	203	32,406
WetNoon	199	30,164
CloudyNight	198	28,364
SoftRainNight	191	31,424
MidRainSunset	191	29,658
WetCloudySunset	186	25,679
MidRainyNight	185	29,094
WetNight	182	26,775
HardRainNight	180	27,414
ClearSunset	180	23,933
CloudySunset	178	29,295
ClearNight	178	25,938
CloudyNoon	178	27,035
MidRainyNoon	176	27,639
HardRainNoon	173	23,633
WetCloudyNight	167	24,616
DustStorm	163	22,453
HardRainSunset	155	22,273

Limitations

• Only successful expert episodes: there are no failures, near misses outside the scenarios, or recovery from off-route states.
• One ego vehicle (lincoln mkz 2020) and one rig placement; the R1 rig is mounted on this car, not on PCCR's original vehicle (see meta/rig_r1_normal.json for the exact positions).
• Labels come from the simulator, not from annotators: the boxes are CARLA's actor bounding boxes.

Acknowledgements

CARLA, Bench2Drive, carla_garage / PDM-Lite, BEVAD / bevad-sim, PCCR camera rigs.