PDM-Lite-DVS

PDM-Lite-DVS is an independently collected synthetic CARLA 0.9.15 event-camera dataset generated with the rule-based PDM-Lite expert on route configurations published by carla_garage. The release lineage is 5,545 public route XMLs → 5,503 recordings in the frozen local pool → 257 rejected recordings → 5,246 published recordings. “Public route release” refers only to the route configurations: the sensor measurements are an independent collection. This is not the official carla_garage sensor dataset and is not affiliated with or endorsed by its authors.

This release contains the complete portable, adapted DVS representation and derived BEV occupancy targets available for a frozen pool of 5,246 accepted recordings. It does not contain RGB, LiDAR, CARLA binaries, Bench2Drive assets, or the original per-frame pickle files.

Code release

Dataset construction, preprocessing, training, and evaluation code will be published in SamanthaZhang-stu/ReflexWorldModel. This GitHub repository is the designated code-release location for the project.

Corpus and quality filtering

The frozen collection pool contained 5,503 recorded routes. The same FullPosttrain quality policy used by the local PDM-Lite training pipeline excluded 257 recordings, leaving 5,246:

Outcome	Routes
Accepted	5,246
Rejected Completed routes with non-exempt collision/infraction outcomes	235
Rejected expert timeouts	17
Rejected blocked-expert routes	5

Accepted recordings have a perfect score/status, or complete with only the pipeline's exempt minimum-speed and minor outside-route-lane infractions. The 5,503-recording frozen pool is distinct from both the 5,545 upstream route configuration files and later recollections on the source machine.

The accepted recordings cover 39 scenario types and these towns:

Town	Recordings
Town01	126
Town02	90
Town03	307
Town04	524
Town05	548
Town10HD	157
Town12	1,798
Town13	1,696

Splits

Recordings are assigned deterministically at canonical physical-route level, stratified by scenario type. Recollections of the same physical route remain in one split.

Split	Recordings	Canonical physical routes	WM-eligible frames
train	4,132	2,899	225,218
validation	538	368	28,636
test	576	368	29,355
total	5,246	3,635	283,209

The split lists and audited route-level metadata are under metadata/. The Dataset Viewer reads the lightweight route-level indexes under viewer/; the large event and label arrays remain external payload files referenced by each row rather than being embedded in the table. The Viewer split named validation maps to the on-disk payload directories events/val/ and labels/val/.

DVS collection and representation

The simulated front DVS camera used a resolution of 768×768, 60° field of view, translation [0.35, 0, 2.25] metres, rotation [0, 0, 0] degrees, positive_threshold=0.3, negative_threshold=0.3, and a raw save interval of 250 ms. When a raw window exceeded 40,000 events it was uniformly randomly subsampled without replacement. The adapter then stable-sorted events by absolute timestamp and re-binned them into the 50 ms training representation published here.

Each route is stored under events/{split}/{route_id}/:

File	Contents
win_x.npy	Event x coordinate, uint16, range 0–767
win_y.npy	Event y coordinate, uint16, range 0–767
win_t.npy	Fractional event time within its 50 ms bin, float32, [0,1)
win_p.npy	Event polarity, uint8, values 0 or 1
win_small.npz	ptr, pose, vel, n_win, W, and H
pe1_align.npz	T_abs_ns, m_causal, valid, t0, and n_win

ptr has length n_win + 1 and indexes the four flat event arrays. pose has shape [n_win, 3] and stores world x, world y, and yaw in degrees; vel has shape [n_win, 1]. The source-machine raw_dir field was deliberately removed from pe1_align.npz to make this publication portable and to avoid leaking local paths; all numerical fields are preserved.

Derived occupancy labels

Each route under labels/{split}/{route_id}/ contains:

File	Contents
fi.npy	Frame indices, integer array
riskgt.npy	Current binary BEV obstacle occupancy, float16, [N,64,64]
riskgt_fut1.npy	Binary occupancy at the simulator frame nearest +500 ms
riskgt_fut2.npy	Binary occupancy at the simulator frame nearest +1,000 ms
cur.npy	Zero-byte legacy cache sentinel; not a NumPy array

Here fi.npy is int32, N = len(fi), and occupancy values are 0 or 1. The grid covers [-32,32) m in both axes at 1 m/cell; rows follow ego-forward x and columns follow ego-right y. The shared labels/front_mask.npy marks the 60° forward field used by downstream risk models. These are obstacle-occupancy targets rasterized from synthetic CARLA bounding-box metadata, not calibrated collision probabilities.

Example loading code:

from pathlib import Path
import numpy as np

route = next((Path("events/train")).iterdir())
x = np.load(route / "win_x.npy", mmap_mode="r")
y = np.load(route / "win_y.npy", mmap_mode="r")
t = np.load(route / "win_t.npy", mmap_mode="r")
p = np.load(route / "win_p.npy", mmap_mode="r")
small = np.load(route / "win_small.npz")

a, b = small["ptr"][0], small["ptr"][1]
first_50ms_window = (x[a:b], y[a:b], t[a:b], p[a:b])

Provenance and modifications

• PDM-Lite expert data were independently generated in CARLA using route configurations from carla_garage.
• The event-camera path follows the LEAD-style single-camera DVS format.
• Raw 250 ms events were converted to flat 50 ms-window arrays.
• The frozen 5,503-route recorded pool was quality-filtered to 5,246 routes.
• Local absolute paths were removed from alignment metadata for publication.
• No Bench2Drive dataset assets are included.

License and attribution

Original payload that the publisher is entitled to license is released under the Creative Commons Attribution 4.0 International license. This does not relicense PDM-Lite/DriveLM, carla_garage, CARLA, CARLA Leaderboard, ScenarioRunner, or other third-party materials. Upstream carla_garage route/code materials are MIT licensed. See DATA_LICENSE.md and THIRD_PARTY_NOTICES.md.

Limitations and responsible use

This is synthetic CARLA data, not real-world sensor data. Filtering for successful expert behavior introduces survivorship bias. Traffic-manager seed 0 was reused for each route, reducing stochastic diversity relative to an independently seeded collection. The PDM-Lite policy, scenario distribution, repeated physical routes, simulator rendering, synthetic DVS noise model, thresholds, and event cap introduce further biases; capping also censors event density, and the sampling seed and pre-cap counts cannot be recovered from this adapted release. It should not be treated as an unbiased estimate of real driving risk and must not be used by itself to control a real vehicle or other safety-critical system. No real people or real-world personally identifying information are represented.

Upstream references

@inproceedings{Sima2024ECCV,
  title={DriveLM: Driving with Graph Visual Question Answering},
  author={Chonghao Sima and Katrin Renz and Kashyap Chitta and Li Chen and Hanxue Zhang and Chengen Xie and Jens Bei{\ss}wenger and Ping Luo and Andreas Geiger and Hongyang Li},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024}
}

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