Researchers present Handoff-H1, a vision-agent system designed to automate material quantity takeoffs from architectural blueprints by combining computer-vision models with tool-using agents and a hierarchical project foundation.

  • The system integrates purpose-built CV models for primitive extraction, agents with image operations, and a knowledge-base-grounded project structure.
  • Evaluation on the Construction Blueprint Takeoff Benchmark uses 10 residential blueprint sets with 2,009 verified line items.
  • Handoff-H1 reaches a composite score of 81.6%, outperforming frontier agents by roughly 20 points and independent estimators who scored 77.6%.
  • The evaluation harness is public via the open harbor framework, with blueprint sets available for research use.

The system demonstrates that orchestrating vision models and agents can achieve near-human quantity precision and superior coverage compared to professional estimators.