drivingbench/driving-dataset-1

DrivingBench Driving Dataset 1

A person drives a Tesla Model 3 through cone courses in parking lots, at walking-to-jogging pace (mostly 0.5–2 m/s). A comma device records both road cameras and the car's state. Each episode is one pass of one course: it starts stopped, follows the corridor of cones through its turns, and stops in the finish area.

62 passes, 108 minutes: four lots, six course layouts, each driven in both directions; 31 passes in daylight, 18 at dusk and 13 at night.

Format

LeRobot v3.0 at 10 Hz. Each episode is stored once, as recorded.

Feature	Shape	Meaning
observation.images.front	380×672×3	Narrow road camera (about 61°×37°), half resolution
observation.images.wide	380×672×3	Wide fisheye road camera, uncorrected
observation.state	2	Measured path curvature (1/m, left positive) and speed (m/s)
action	2	Curvature and speed at this step
observation.yaw_rate_rad_s	1	Yaw rate (rad/s, left positive)
action.acceleration_mps2	1	Longitudinal acceleration (m/s²)
ego.pose	3	x (m, forward), y (m, left), yaw (rad), relative to the episode's first frame

• The action is what the car actually did at that step, limited to what it can execute: curvature within the steering lock (about ±0.19/m) and speed up to 3.5 m/s. Curvature describes the path, not a wheel angle, so it carries over to other vehicles. Build future-action chunks with delta_timestamps.
• Episode boundaries. Each pass starts 1 s before the car first moves and ends 3 s after its final stop.
• Instruction. Every frame's task is the same: "Follow the cone course, stay between the cones, and stop in the finish area."

Episodes

Episodes are in recording order, except that the last 8 are held out (meta/info.json splits: train 0–53, test 54–61): passes of courses whose other passes are in train. meta/passes.json labels each episode with its lot (A–D), layout, direction (clockwise or counterclockwise, the sense of its net turn), pass number within that course, lighting, start time and length.

from lerobot.datasets.lerobot_dataset import LeRobotDataset

train = LeRobotDataset(
    "drivingbench/driving-dataset-1",
    episodes=list(range(54)),
    delta_timestamps={"action": [k / 10 for k in range(10)]},  # the next second of actions
)

To augment by mirroring, flip both images, then negate curvature, yaw rate, and the pose's y and yaw. Speed is unchanged.

License

CC BY 4.0. License plates on vehicles, and people close enough to recognize, are blurred. If you use the dataset, please link drivingbench.com.