pamasan/duckad-data

DuckAD Driving Dataset

Expert driving demonstrations for DuckAD, an end-to-end vision-based driving model, collected in CARLA on custom Duckietown-style maps. A rule-aware expert driver was rolled out under six traffic/obstacle scenarios; every frame pairs a front camera image with the expert's future trajectory, a high-level navigation command, and ground-truth bird's-eye-view (BEV) semantics.

• 214,200 training frames from two maps (duckietown_04, duckietown_05), six scenarios × 35,700 frames each.
• 10,200 test frames from an unseen map (duckietown_06) rendered under unseen HDRI lighting, for cross-map generalization evaluation.
• The simulator runs synchronously at 20 Hz and each timestep is captured from three camera rigs (center, left, right — side rigs give recovery-style viewpoint diversity with correspondingly transformed trajectory labels), so the training set corresponds to roughly one hour of expert driving.

Note that this dataset needs to be scaled 1/20 for the real Duckietown!

Scenarios

scenario	Traffic	Duckies
baseline	–	–
traffic_only	✓	–
ducks_roadside	–	roadside
ducks_roadside_traffic	✓	roadside
ducks_obstacle	–	on-road obstacles
ducks_obstacle_traffic	✓	on-road obstacles

Fields

Field	Type	Description
image	224×224 RGB image	front (fisheye) camera frame
seg	224×224 grayscale image	binary foreground mask: 255 = Duckietown foreground (road surface, lane markings, signs, bots, duckies), 0 = replaceable background. Used for background-swap augmentation.
bev	64×64 grayscale image	ground-truth BEV semantics as raw CARLA semantic tag ids (see below)
command	class label	navigation command: DEFAULT (lane follow), LEFT, STRAIGHT, RIGHT (junction maneuvers)
speed	float32	ego speed in m/s
trajectory	10×2 float32	future spatial waypoints at 1 m arc-length spacing, meters in the ego frame (x forward, y left)
temporal_trajectory	10×2 float32	future ego positions sampled every 0.3 s (3.33 Hz), same ego frame — encodes the speed profile
scenario / map / episode / frame	strings / int	provenance metadata

BEV tag ids: 29 center_lane, 30 side_lane, 31 asphalt, 32 stop_lane, 33 sign, 34 bot, 35 duck; any other id (e.g. 11 = terrain) is background. For training we remap these to 8 contiguous classes (background=0 + the 7 above).

Usage

from datasets import load_dataset

ds = load_dataset("pamasan/duckad-data", split="train")
sample = ds[0]
sample["image"]                # PIL.Image, 224x224 RGB front camera
sample["seg"]                  # PIL.Image, 224x224 foreground mask (255 = Duckietown foreground)
sample["bev"]                  # PIL.Image, 64x64 BEV semantics (CARLA tag ids, see table above)
sample["command"]              # int class label: 0 DEFAULT, 1 LEFT, 2 STRAIGHT, 3 RIGHT
sample["speed"]                # ego speed, m/s
sample["trajectory"]           # 10 spatial waypoints [x, y], 1 m spacing, meters in ego frame
sample["temporal_trajectory"]  # 10 future ego positions [x, y] sampled every 0.3 s -> speed profile

import numpy as np
bev_ids = np.array(sample["bev"])                     # 64x64 tag ids
fg_mask = np.array(sample["seg"]) > 0                 # boolean foreground mask

Collection

Data was collected with a rule-aware expert (lane following, junction turns, stopping for duckies and traffic) driving in synchronous CARLA at 20 Hz (fixed_delta_seconds=0.05). Each recording segment respawns the ego at a new random location; traffic vehicles are driven by the CARLA Traffic Manager. Junction approaches on the training maps are labeled with the expert's chosen LEFT/STRAIGHT/RIGHT command; everywhere else the command is DEFAULT.

Notes

• Frames are stored as captured — no augmentation is baked in.
• The test split is for cross-map/lighting generalization; it uses a map and HDRI environment that never appear in train.

License

Released under CC-BY 4.0. All visible content (maps, duckiebot vehicles, props) is custom-built for this dataset.