metadata
license: cc-by-sa-4.0
Dataset Card for Adv-nuSc
Dataset Description
Overview
The Adv-nuSc dataset is a collection of adversarial driving scenarios generated by the Challenger framework, designed to evaluate the robustness of autonomous driving (AD) systems. It builds upon the nuScenes validation set, introducing intentionally challenging interactions that stress-test AD models with aggressive maneuvers like cut-ins, sudden lane changes, tailgating, and blind spot intrusions.
- Curated by: Zhiyuan Xu, Bohan Li, Huan-ang Gao, Mingju Gao, Yong Chen, Ming Liu, Chenxu Yan, Hang Zhao, Shuo Feng, Hao Zhao
- Affiliation: Tsinghua University; Geely Auto
- License: CC-BY-SA-4.0
- Project Page: Challenger
Dataset Structure
The dataset consists of:
- 156 scenes (6,115 samples), each 20 seconds long
- Multiview video data from 6 camera perspectives
- 3D bounding box annotations for all objects
Key statistics:
- 12,858 instances
- 254,436 ego poses
- 225,085 total annotations
Usage
The Adv-nuSc dataset is in nuScenes format. However, a few minor modifications are needed to evaluate common end-to-end autonomous driving models on it. Please follow instructions in Eval E2E.
Creation Process
Source Data
- Built upon the nuScenes validation set (150 scenes)
- Uses nuScenes' original sensor data and annotations as foundation
Adversarial Generation
- Trajectory Generation: Uses diffusion models to create diverse adversarial maneuvers
- Physics Simulation: Simulates trajectories with an LQR controller and kinematic bicycle model
- Multi-round Refinement: Iteratively improves trajectories using:
- Drivable area compliance checks
- Collision avoidance
- Adversarial challenge scoring
- Neural Rendering: Produces photorealistic multiview videos using MagicDriveDiT
Filtering
Scenarios are filtered to ensure:
- No collisions between adversarial and other vehicles
- Adversarial vehicle remains within 100m × 100m area around ego
- Meaningful interaction with ego vehicle occurs
Intended Use
- Robustness evaluation of autonomous driving systems
- Stress-testing end-to-end AD models (e.g., UniAD, VAD, SparseDrive, DiffusionDrive)
- Identifying failure modes in perception, prediction, and planning modules
Limitations
- Currently focuses on single adversarial vehicles (though extendable to multiple)
- Open-loop evaluation (no reactive ego agent)
- Minor rendering artifacts compared to real sensor data
Ethical Considerations
Safety
- Intended for research use in controlled environments only
- Should not be used to train real-world systems without additional safety validation
Privacy
- Based on nuScenes data which has already undergone anonymization
- No additional privacy concerns introduced by generation process