DriveBench-Embeddings: 298,326 Driving Scene Vectors

Author: Nikhil Upadhyay | MSc Business Analytics | Dublin Business School Model: Trazemag/DriveBench

Overview

Pre-computed 256-dimensional DriveBench embeddings for all 298,326 clips from the NVIDIA PhysicalAI-AV dataset across 25 countries.

Use these as features for any downstream driving task without running the model.

File

File	Size	Shape
drivebench_embeddings.npz	282 MB	(298326, 256)

Usage

import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "Trazemag/DriveBench-Embeddings",
    "drivebench_embeddings.npz",
    repo_type="dataset")

data = np.load(path)
embeddings = data["embeddings"]  # (298326, 256) float32

# Match to clip IDs using PRECOG-Labels dataset
# huggingface.co/datasets/Trazemag/PRECOG-Labels

What each embedding captures

Each 256-dim vector encodes:

• Danger context (AUC 0.84 on held-out countries)
• Geographic driving patterns (6 regions, 25 countries)
• Time-of-day risk (peak danger 13:00-15:00)
• Radar sensor health (AUC 1.00)
• Traffic density and scene complexity

Citation

@misc{upadhyay2026drivebench,
  title  = {DriveBench: General-Purpose Driving Scene Encoder
            via Multi-Task Safety-Focused Pre-training across 25 Countries},
  author = {Upadhyay, Nikhil},
  year   = {2026},
  url    = {https://github.com/TrazeMaG/PRECOG-AV}
}