NARRATE: New Multimodal Driving Dataset for Explainable AI
The newly launched dataset, NARRATE, aims to enhance explainable AI in the realm of automated driving. It features 2,050 annotated events collected from 35 skilled drivers and instructors operating on public roads in Australia. This dataset includes synchronized streams of visual data, localization, motion, and LiDAR, along with free-text explanations provided either during the drive or afterward. NARRATE offers action labels and scenario-context labels across six high-level and 32 detailed categories, as well as span-level Situational Awareness (SA) annotations for Perception, Comprehension, and Projection. This initiative seeks to fill the gap left by existing language-annotated driving datasets, which often rely on observer-written, post-hoc, simulation-based, or sensor-generated explanations. The relevant paper can be found on arXiv under the identifier 2608.14767.
Key facts
- NARRATE is a multimodal real-world Australian driving dataset.
- It includes 2,050 annotated events from 35 experienced drivers and driving instructors.
- Data was collected on public roads.
- Each event is grounded in synchronised visual, localisation, motion, and LiDAR streams.
- Explanations were elicited from drivers in-vehicle and/or post-drive.
- The dataset includes action labels and scenario-context labels (six high-level and 32 fine-grained categories).
- Span-level Situational Awareness annotations cover Perception, Comprehension, and Projection.
- The paper is available on arXiv (2608.14767).
Entities
Institutions
- arXiv
Locations
- Australia