TruckDrive

Warning

Experimental Dataset Support

TruckDrive support is currently experimental and may still change. If you run into issues, please open a bug report on GitHub Issues.

TruckDrive is a long-range autonomous highway driving dataset designed for heavy-truck safety, perception, prediction, and planning research. It targets high-speed highway operation, where reliable scene understanding hundreds of meters ahead is required for anticipatory planning and safe braking.

The py123d integration supports multi-camera and multi-lidar scene conversion, per-frame ego states, 3D bounding boxes (where available), and per-log lane-map objects.

For extensive details about the dataset contents, sensor setup, and companion tools, refer to the official TruckDrive repository: torc-ai/TruckDrive.

Overview

Download

Hugging Face (gated)

Code

py123d.parser.truckdrive (parser/downloader integration in this repository)

License

Please refer to the dataset’s official license terms.

Available splits

truckdrive_train, truckdrive_val, truckdrive_test

Available Modalities

Name

Available

Description

Ego Vehicle

✓

Vehicle pose and inferred dynamics from trajectory and calibration data. See EgoStateSE3.

Map

✓

Per-log map objects parsed from lane-line and lane-segment annotations (including lane topology). See Lane and RoadLine.

Bounding Boxes

✓

3D box detections with TruckDrive label mapping for train/val scenes. See BoxDetectionsSE3.

Traffic Lights

X

Not currently exposed as a dedicated detection modality.

Cameras

✓

Multi-camera rig (Leopard cameras) with calibrated pinhole intrinsics/extrinsics. See Camera.

Lidars

✓

One merged AEVA stream plus three Ouster lidars. See Lidar.

Download

TruckDrive is distributed as a gated Hugging Face dataset. After requesting access, export a token and download selected scenes with:

pip install py123d[hf]
export HF_TOKEN=hf_...

py123d-download dataset=truckdrive \
    'dataset.downloader.scenes=[scene_28_1]'

By default, the downloader fetches camera, lidar, poses, calibrations, and annotation archives and extracts them into the expected on-disk layout.

Installation

The parser itself is included in py123d. Install the hf extra if you want to use the built-in Hugging Face downloader:

pip install py123d[hf]

Conversion

Local mode (already downloaded scenes):

export TRUCKDRIVE_DATA_ROOT=/path/to/TruckDrive
py123d-conversion dataset=truckdrive

# Convert a custom scene list:
py123d-conversion dataset=truckdrive \
    'dataset.parser.scene_names=[scene_28_1,scene_35_1]'

Streaming mode (download + convert in one run):

export HF_TOKEN=hf_...
py123d-conversion dataset=truckdrive-stream \
    'dataset.parser.scene_names=[scene_28_1]'

In streaming mode, scenes are downloaded to a managed temporary directory, converted, then cleaned up.

Dataset Issues

  • truckdrive_test scenes currently lack the ground-truth trajectory/annotations needed for full log conversion, so the current parser skips test logs in get_log_parsers().

  • Sensor frequencies and annotation completeness can vary by scene; verify assumptions for downstream training/evaluation.

Citation

If you use TruckDrive, please cite the original dataset publication and follow the official citation instructions from the dataset maintainers.