manfredch/lyapunov/cad-import

CAD Drawing-to-Scene Workflow

Input: drawings (DXF/DWG files, drawing images/PDFs/screenshots, or a combination), intended use, and site information. Completion: every component has traceable dimension/topology evidence (annotation, contour, or derivation), and render comparison passes. Drawings remain read-only.

CAD input is a special case of scene construction: understand the drawing, reconstruct to scale, then follow scene-construction. Plan first: for a new scene or major reconstruction, write a short plan through environment-planning; this skill turns drawings into dimensions and topology. Do not invent structures or dimensions absent from the drawing.

Cross-check photos separately: drawings provide checkable dimensions, topology, and elevations; photos provide appearance, perspective, material appearance, and calibrated dimension estimates when scale references or camera pose are known (grading and derivation chains are in photo-reconstruction). When elevations/sections/details are missing, supplement with photos and public sources through environment-research; maps/public dimensions can also calibrate estimates. Identify adopted sources and affected components. Resolve conflicts by checking the object, date, and drawing version rather than treating drawings as unconditionally correct. Separate annotations from estimates: label annotated values as annotated, and explain derivations/estimates instead of mixing them with measurements.

1. DXF: Read Actual Coordinates and Units with cad_inspect

When cad_inspect is visible in the current session, use it first. The product Blender plugin registers it with an isolated interpreter explicitly configured by LYAPUNOV_CAD_PYTHON, independent of Blender/Isaac Python.

• Read actual DXF (ASCII and Binary): original/unknown units, layers, block definitions/references (including space and parent block for nested references), actual curve parameters (endpoints/vertices and bulge, circle centres/radii/angles, SPLINE knots/weights), dimensions, and unreadable entities.
• Do not guess units: an undeclared unit means units.known=false and metresPerUnit=null. Specify a unit explicitly for conversion; the result records caller provenance and warns when it conflicts with the file declaration.
• Two consumption surfaces: structured reports for modelling scripts and a default bounded summary. Request full explicitly for raw data; detail counts and per-curve point limits still apply. Only counts are complete.
• Read-only: do not write drawings back, perform Boolean/solid modelling, or invent CAD data from images.

Interpret actual configuration errors: an unconfigured interpreter means this deployment has not supplied the isolated interpreter (see development docs; initial installation does not require an offline cache). Missing ezdxf means dependencies are absent. A missing converter for DWG means conversion to DXF is required, not that the file is bad DXF. Converter configuration uses plugin dwgConverter, LYAPUNOV_DWG_CONVERTER (with _KIND/_ARGS), or GNU LibreDWG dwg2dxf on PATH. ODA File Converter operates on directories; this implementation has not verified that branch and explicitly rejects it with `DWG_CONVERTER_UNSUPPORTED`. Do not present it as available; see docs/DRAWING_INPUT.md. Follow actual errors to alternatives (image reading, research, or requesting a DXF version). Do not install these dependencies into Blender/Isaac Python, and do not infer that the product cannot read DXF merely because one machine lacks configuration. Record actual unresolved capability gaps in delivery notes.

DWG/PDF Use the Same Native Routing: drawing_inspect and cad_convert

When drawing_inspect is visible, route by file header, not extension, without manually guessing the format first.

• DWG: use the configured converter to create a new DXF, preserving the source. Compare source sha256 before/after; sourceUnchanged must be true. Parse with the same ezdxf path. Report actual losses through conversion.losses / lossCategories and delivery notes. An unreadable converted result returns DWG_CONVERTED_UNREADABLE and is not published. Use cad_convert for a separately reusable DXF; it defaults to <workspace>/cad-converted/<name>.dxf and does not overwrite unless overwrite=true.
• Vector PDF: return actual path operators including CTM transforms, text/page coordinates, and page dimensions. Page units are 1/72-inch drawing points, not metres. Metric output requires an explicit scale, such as 1:100; otherwise metresKnown=false and no metric dimensions are returned.
• Scanned/mixed PDF: export embedded images as PNG and deliver them directly as images in context. Pages with /Rotate or CropBox prefer a full-page preview; original image pixels are not equivalent in these cases, and the receipt explains why. This remains image reading, with dimensions graded as below.

2. Build Geometry in Task Scripts

scene_import does not support dxf/dwg; it supports glb/splat/mjcf/urdf/blend/usd. Build geometry in task scripts, using blender_run python_script or connected mcp__blender__*. See architectural-world for modelling/export contracts.

• Metres = drawing-unit value multiplied by `units.metresPerUnit`. With unknown units, retain relative proportions and disclose them, or convert using explicitly supplied units.
• DXF model space lies in XY with thickness along Z, matching Lyapunov/Blender Z-up; no axis swap is needed. Convert coordinates:"ocs" with an extrusion direction other than (0,0,1) to WCS first.
• closed=true LWPOLYLINE/POLYLINE vertex order defines a closed boundary with the last point returning to the first. Interpret wall/plot/other boundaries using layers and annotations; 2D base elevation is in elevation. Do not use `bounds` as a rectangular wall contour: it loses an L-shaped recess.
• Bulge segments are circular arcs with bulge = tan(θ/4). Transform block instances back to world coordinates through the reference chain using row-major right multiplication (v' = v*M); model a repeatedly referenced block once.
• Dimension measurements can be cross-checked in drawing units. Manually overridden annotation text does not change the measured value; explicitly state which source governs.

3. Drawing Images, PDFs, and Screenshots

Read supplied images for room contours/zones, door/window positions and opening directions, wall thickness, key dimensions, fixed furniture, and symbols. Reconstruct in Blender using the dimensions read. Choose the execution surface by task: batch/reproducible work uses blender_run; interactive iteration uses connected mcp__blender__* with execute_blender_code and get_viewport_screenshot. If MCP is unavailable, fall back to blender_run instead of waiting indefinitely. See architectural-world for metric Z-up, component collision, hollow interiors, and export.

Vector PDF can also provide parsed evidence: drawing_inspect returns actual operators and text/coordinates, but no metric values without scale. Scanned/mixed PDF full-page PNGs enter image context; grade dimensions from image reading and explain the pixel-to-metre chain. Bitmap drawings (photos/screenshots) likewise use image reading regardless of extension.

4. Map CAD Symbols to Interior Assets

Determine symbols from legends, layers, and context before finding matching doors/windows/furniture/equipment/vegetation in built-in or public assets. If none exists, create it to confirmed dimensions. Placement/orientation follows the drawing; do not move walls/openings or change annotated dimensions on your own.

Evidence and Capability Boundaries

• Grade dimensions: explicit drawing annotations, then scale-derived values labelled as derived, then customary defaults labelled as modelling assumptions. State the grade of each value. Cross-check CAD against photos, maps, and public dimensions; resolve conflicts by object/date/drawing version, without unconditional source-category priority.
• Report only visible rooms/doors/windows/clearances. List hidden gaps and supplement with photos (photo-reconstruction), focused research (environment-research), or map/public-dimension calibration before asking the user.
• Compare reconstruction renders side by side with the original drawing, checking walls/openings/proportions/furniture layout; list and correct differences.
• Structural calculations, building-code compliance review, and formal construction documents are outside this capability. Deliver a scene model, not a construction document. Requests to turn a drawing into an illustrative rendering use image generation; verify currently configured service availability through asset-generation.

File Boundary (ENV-04)

• Use bash/read/write and other file tools for task-workspace scripts, references, intermediate outputs, and exports.
• Product state must use domain tools (scene_*/asset_*/sim_*/robot_*, etc.). Do not change product state by editing files on disk, and do not modify product source. Capabilities come from skill contracts and actual product-tool results.