Drive-P2D: A Progressive Perception-to-Decision Benchmark for Autonomous Driving VLMs

Drive-P2D is a benchmark containing 6,650 questions designed to evaluate Vision-Language Models across three progressive levels: object-level perception, scene understanding, and driving decision-making in autonomous driving. It was created by ColamentosZJU and is hosted on Hugging Face. The dataset page was last updated on June 4, 2026.

Use Cases

• Benchmarking VLM object detection capabilities based on the Object-level perception questions.
• Evaluating scene understanding and reasoning in driving contexts based on the Scene-level questions.
• Testing driving decision-making logic and safety based on the decision-making level questions.
• Comparing model performance across a progressive hierarchy of autonomous driving tasks.

Strengths

• Contains 6,650 questions, providing a substantial test set.
• Structures evaluation across three progressive levels (Object, Scene, Decision), which suggests a methodical design.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

ColamentosZJU via Hugging Face.

Freshness

Last updated 2026-06-04 17:04:17; freshness should be verified.