CASCADE: Causal Spatio-Temporal Annotations for 2,000+ Driving Clips

2,000+ 20-second driving clips from the NVIDIA Physical AI Dataset for AV are annotated with causally-linked, temporally-ordered structures. The dataset, created by NVIDIA, explicitly links environments, agents, traffic control, and actions with causal and co-reference relations. This allows complex driving scenarios to be queried as structured graphs.

Use Cases

• Training causal inference models based on annotated causal relations between agents and actions.
• Developing spatio-temporal reasoning systems based on temporally-ordered annotations of driving scenarios.
• Querying complex driving events as structured graphs based on the dataset's explicit graph representation.
• Benchmarking autonomous vehicle perception systems on scenarios with annotated agent interactions and traffic control.

Strengths

• Over 2,000 annotated driving clips, each 20 seconds long.
• Annotations include explicit causal and co-reference relations, enabling structured graph queries.
• Data is derived from the NVIDIA Physical AI Dataset for AV, suggesting a high-fidelity source.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Last updated 2026-07-16 10:58:03; freshness should be verified.

Provenance

Source

NVIDIA Physical AI Dataset for AV

Collection Method

Clips drawn from a source dataset and annotated with causal structures.

Freshness

Last updated 2026-07-16 10:58:03.