MobilityBench: 10,000+ Real-World Route-Planning Queries and Tool Calls

MobilityBench contains between 10,000 and 100,000 anonymized mobility queries sourced from Amap for evaluating route-planning agents. Developed by GD-ML and released in early 2026, the dataset provides a structured task taxonomy and ground truth tool calls for real-world navigation scenarios.

Use Cases

• Evaluating route-planning agents using the structured tool call ground truth
• Training question-answering models on mobility-specific task taxonomies
• Benchmarking agentic workflows for real-world navigation queries

Strengths

• 10,000 to 100,000 record scale
• Sourced from real-world anonymized Amap mobility queries
• Includes structured ground truth for tool calls

Limitations

• Progressive release status limits immediate access to the full corpus
• Potential geographic bias toward regions serviced by the primary data source

Provenance

Source

Amap via GD-ML

Collection Method

Anonymized mobility queries organized with a task taxonomy

Freshness

Last updated March 2026.

Geography

United States