Proprietary applied R&D · Active development
AutonomousPlan2CAD
A research system for turning architectural images and PDFs into structured, verifiable and editable CAD—with human review for genuine exceptions.
Executive summary
AutonomousPlan2CAD investigates architectural drawing reconstruction as a structured reasoning problem, not a simple raster-to-vector conversion. It combines multimodal/CV observations, a Semantic Building Model, deterministic geometry, dimensional constraints, automated QA and an editable DXF compiler.
- Sector: residential construction, architecture, engineering and CAD workflows
- Classification: proprietary BGPD Labs applied R&D under active development
Why the problem is difficult
A plan contains walls, openings, rooms, labels, fixtures, dimensions, scale, relationships, topology and implicit geometry. A visually plausible drawing can still contain semantic or dimensional errors. The system must understand what each element means, how it relates to others and when the evidence is insufficient.
- Raster-to-vector conversion solves only a portion of the task
- Critical dimensional conflicts must block release
- Human review remains necessary for genuine ambiguity
Simplified architecture
The pipeline separates probabilistic perception from deterministic engineering and preserves evidence and verification state.
- Source image or PDF
- Ingestion and normalization
- Multimodal/CV observations
- Semantic Building Model
- Deterministic geometry and topology
- Dimensional reconciliation
- Automated QA
- Exception review
- CAD compiler
- Editable DXF
What has been demonstrated
An experimental pipeline execution processed the following scale and diversity of geometry. These figures are not accuracy benchmarks.
- Walls
- 297
- Doors
- 40
- Windows
- 30
- Detected rooms
- 23
- Semantic rooms
- 14
- CAD block references
- 22
Entities processed in an experimental execution.
Entities processed in an experimental execution.
Entities processed in an experimental execution.
Detection count, not a correctness rate.
Identified room entities in the execution.
Reusable CAD references generated.
Controls and current status
The system distinguishes observed, inferred, assumed and human-corrected information. It is moving from assisted drafting toward bounded, exception-based autonomy: automate what evidence supports, validate what can be measured and escalate ambiguity. AutonomousPlan2CAD remains active research; it is not presented as a production-ready commercial system or a claim of full autonomy.
- Provenance and verification state
- Wall connectivity, room closure, opening hosting and collision checks
- Blocking release for unresolved critical issues