ln2697/lead-123d

---

license: mit pretty_name: LEAD 123D viewer: false tags:

• autonomous-driving
• carla
• py123d
• end-to-end-driving
• multimodal

---

LEAD 123D

A privileged expert drives CARLA Leaderboard 2.0 routes and writes every modality it sees to the standardized Py123D format: cameras, lidar, radar, the full scene state and the maps to place it on. Code, loaders, visualizers and closed-loop benchmarks live in kesai-labs/lead.

Policy learning is one application. The logs carry the whole scene, so perception, world models, video generation and self-play RL read the same files.

Contents

Path	Size	Logs
logs/normal_view	651 GB	8,930
logs/perturbated_view	416 GB	8,930
maps/carla	789 MB	—

logs/<view>/<scenario_type>/<log_name>/ holds one route, across 43 scenario types. perturbated_view is the same routes with the camera rig rotated and translated, which is what lets a policy learn to recover from its own drift. maps/carla holds the 12 CARLA towns, Town01 through Town15.

The training cache the LEAD repository can read is not part of this dataset. It is derived from these logs and policy-specific; scripts/common/build_cache.sh recomputes it.

Modalities

Each log directory holds 31 Arrow files:

• Cameras, six positions (pcam_f0, pcam_b0, pcam_l0, pcam_l1, pcam_r0, pcam_r1) × RGB, depth, semantic and instance segmentation
• Lidar (lidar.lidar_top) and radar (radar.radar_merged)
• Scene state: ego state, 3D box detections, traffic light detections
• Metadata: custom.driving_meta, sync

Point clouds and scene state are written at the native 20 Hz tick rate. Camera streams are written on every fifth tick (4 Hz) to keep storage tractable.

Download

The dataset is roughly a terabyte over half a million files, so fetch the part you need. logs/normal_view and maps alone are enough to train and evaluate.

pip install "huggingface-hub[hf_xet]"
hf download ln2697/lead-123d --repo-type dataset --local-dir lead-123d \
    --include 'logs/normal_view/*' 'maps/*' 'config.yaml'

Downloads resume, so an interrupted one is continued by running it again. Drop --include for everything, perturbated views included.

Usage

The logs are plain Py123D, so the standard API reads them without LEAD:

from py123d.api.scene.arrow.arrow_scene_builder import ArrowSceneBuilder
from py123d.api.scene.scene_filter import SceneFilter
from py123d.common.execution.thread_pool_executor import ThreadPoolExecutor

# Enumeration is lazy; nothing is decoded here.
scenes = ArrowSceneBuilder(
    logs_root="lead-123d/logs",
    maps_root="lead-123d/maps",
).get_scenes(
    SceneFilter(
        split_names=["normal_view"],
        future_num_iterations=40,  # 2 s at the 20 Hz tick rate
        required_scene_modalities=["camera:all@initial"],  # drop partial rigs
    ),
    ThreadPoolExecutor(),
)

# Modalities are read per frame, by iteration index.
ego = scenes[0].get_ego_state_se3_at_iteration(0)  # EgoStateSE3
boxes = scenes[0].get_box_detections_se3_at_iteration(0)  # BoxDetectionsSE3

To train on it, point PY123D_DATA_ROOT at the download and follow the LEAD README.

Citation

@inproceedings{Nguyen2026CVPR,
  author    = {Long Nguyen and Micha Fauth and Bernhard Jaeger and Daniel Dauner and Maximilian Igl and Andreas Geiger and Kashyap Chitta},
  title     = {LEAD: Minimizing Learner-Expert Asymmetry in End-to-End Driving},
  booktitle = {Conference on Computer Vision and Pattern Recognition (CVPR)},
  year      = {2026},
}

@article{Dauner2026ARXIV,
  author  = {Dauner, Daniel and Charraut, Valentin and Berle, Bastian and Li, Tianyu and Nguyen, Long and Wang, Jiabao and Jing, Changhui and Igl, Maximilian and Caesar, Holger and Ivanovic, Boris and Geiger, Andreas and Chitta, Kashyap},
  title   = {123D: Unifying Multi-Modal Autonomous Driving Data at Scale},
  journal = {arXiv preprint arXiv:2605.08084},
  year    = {2026},
}