Deterministic 3D CAD CUA & RLVR for Frontier AI
Frontier models can navigate 2D web forms, but fail on 85.0%–90.0% of industrial 3D mechanical blueprints. Train and evaluate Computer-Use & Reasoning Agents with human SolidWorks / Creo trajectories, ISO-128 2D blueprints, and verify_step.py B-Rep geometric verifiers.
85.0%+
Frontier CAD-CUA Failure Rate
Claude 4.5, GPT-5.2 & Gemini 3 on 20 ISO-128 blueprints
62 Episodes
Live CAD-CUA Dataset (HF)
11,792 lossless frames · 237,579 GUI events · 6 CAD suites
215 CAD Files
Paired B-Rep Deliverables
115 native (.sldprt/.prt/.fcstd) + 98 .STEP + 48 blueprints
<10ms
Deterministic .STEP Verifier
verify_step.py BBox IoU, Volume, Euler-Poincaré & Bore Spectrum
CAD-CUA Deterministic .STEP RLVR Benchmark
Evaluating Claude 4.5, GPT-5.2, and Gemini 3 across 20 ISO-128 4-View Reference Engineering Blueprints (data/benchmark.jsonl). Candidate .STEP B-Rep solids are scored deterministically against human mechanical engineer ground truth via verify_step.py.
Why 2D Web CUA Agents Break Inside 3D Parametric Engineering Environments
Unlike 2D DOM buttons, industrial CAD requires reading Third-Angle orthographic projections, maintaining 6-DoF camera orientation, and constructing watertight CLOSED_SHELL B-Rep topology. Current frontier models swap 3D principal axes, omit internal counterbores, and generate self-intersecting fillets that fail manifold verification 85.0% to 90.0% of the time.
| Rank & Model / Agent | Pass@1 (20 Blueprints) | Failure Rate | Mean Reward (R) | BBox IoU | Topology | Primary Observed Failure Mode |
|---|---|---|---|---|---|---|
1 Human CAD Engineer (Ground Truth)Expert Baseline (SolidWorks / Creo) | 100.0% (20/20) | 0.0% | 1.000 | 1.000 | 1.000 | Watertight ISO-10303-21 B-Rep manifold solid |
2 Claude 4.5 (Computer Use / CAD Code)Frontier CUA + Code | 15.0% (3/20) | 85.0% | 0.463 | 0.543 | 0.415 | Missing internal counterbores, fillet self-intersections |
3 GPT-5.2 (Vision + CAD Code)Frontier Reasoning | 10.0% (2/20) | 90.0% | 0.391 | 0.476 | 0.335 | 3D axis swap, non-manifold shells, bore pitch drift |
4 Gemini 3 (Multimodal Spatial + CAD)Frontier Multimodal | 10.0% (2/20) | 90.0% | 0.371 | 0.449 | 0.324 | Unclosed sketch loops, missing radial hole arrays |
Reproducible with python3 verify_step.py --ground-truth gt.step --candidate pred.step included in the public release.
Inspect Real 2D Blueprints, CUA GUI Trajectories & B-Rep Rewards
Construct a watertight 3D B-Rep solid matching 2D Reference Blueprint CAD-000020 (outer envelope 100.00 × 67.50 × 35.00 mm, 78 B-Rep faces: 14 planar, 38 cylindrical) within ISO 2768-mK tolerances.
python3 verify_step.py \ --ground-truth models/cad-000020.zip \ --candidate models/cad-000020.zip
Where Frontier Models Break Down: ISO-128 Blueprint vs. 3D CAD Ground Truth & B-Rep Diagnostics
Omitted Concentric Counterbores & Extrusion Depth Drift
All three frontier models extrude the outer rectangular block envelope but fail to subtract the concentric inner trunnion counterbores on the Y-Z plane, reducing cylindrical surface count from 42 to <15 and violating ISO 2768-mK bounding tolerances.
- ▸Target Topology: 91 ADVANCED_FACE (19 planar, 42 cylindrical) vs. AI Candidate: 30–38 faces
- ▸Principal Envelope Drift: +18.2% on X-axis, -22.0% on Y-axis (s_bbox drops to 0.52)
- ▸Closed-Shell Volume Error: +44.8% excess solid volume from un-cut internal trunnion bores
Test verify_step.py Live in Your Browser (<10ms Execution)
Select a reference blueprint below, toggle between Human Ground Truth and Frontier Model candidate .STEP files—or drag-and-drop your own .STEP / .STP file to compute exact B-Rep RLVR rewards in real time.
Tip: Try editing CLOSED_SHELL to OPEN_SHELL or modifying CARTESIAN_POINT coordinates above to watch the B-Rep reward update instantaneously.
R = 0.35·s_bbox + 0.25·s_vol + 0.20·s_top + 0.20·s_surfHeld-Out Enterprise Split: 105 Verified Sessions & 29 Multi-Trajectory Part Families
While 62 episodes are published openly in spacetime-ml/cad-cua, SpacetimeML maintains an air-gapped, continuously expanding private holdout (currently 105 verified human CAD sessions · 72.12 hours · 102,821 frames · 502,556 OS events · 0 duplicate deliverable SHA-256 hashes) featuring 29 multi-engineer part families (part_family_id) where 2–4 distinct mechanical engineers independently model the exact same 2D blueprint.
part_family_id) — Paired Human Trajectories for DPO & GRPO| Part Family ID | Blueprint Spec & Component | Engineers | Session Durations (Min) | Horizon Spread | Frames / Events | Post-Training RL / DPO Signal |
|---|---|---|---|---|---|---|
| pf_model_mania_2011 | [AX-145 / 018]Model Mania 2011 Ribbed Clevis & Boss | 4× Human | 10.4m · 13.6m · 25.2m · 46.0m | 4.4× efficiency spread | 2,309f / 15,316e | DPO Action-Horizon Pruning + Pass@4 GRPO |
| pf_ax_167 | [AX-167 / 174]Multi-Port Hydraulic Valve Block | 4× Human | 15.3m · 18.7m · 32.6m · 64.2m | 4.2× efficiency spread | 3,748f / 5,909e | Feature-Tree Ordering Preference Pairs |
| pf_model_mania_2008 | [AX-023 / 119]Model Mania 2008 Swept Bellcrank Yoke | 3× Human | 20.8m · 68.1m · 101.6m | 4.9× efficiency spread | 3,815f / 42,647e | Cross-Suite Trajectory Alignment (SolidWorks / Creo) |
| pf_handle_phase1 | [AX-123 / 106]Ergonomic Contoured Actuator Handle | 3× Human | 31.2m · 80.1m · 95.7m | 3.1× efficiency spread | 5,204f / 15,229e | Surface Fillet & Shell Recovery Trajectories |
| pf_model_mania_2014 | [AX-118 / 126]Model Mania 2014 Tri-Boss Trunnion | 3× Human | 29.5m · 49.6m · 68.1m | 2.3× efficiency spread | 4,331f / 4,364e | Multi-Datum Sketch Plane Grounding |
| pf_ax_166 | [AX-166 / 176]Slotted Eccentric Cam Housing | 3× Human | 21.4m · 37.6m · 60.6m | 2.8× efficiency spread | 3,245f / 3,504e | Concentric Bore Constraint Satisfaction |
| pf_ax_105 | [AX-105]Counterbored Bearing Pillow Block | 3× Human | 28.9m · 36.5m · 39.9m | 1.4× efficiency spread | 2,839f / 2,849e | Consistent B-Rep Topology Convergence (χ=2) |
| pf_stepped_shaft | [AX-017 / 014]Keyed Multi-Diameter Drive Shaft | 3× Human | 8.2m · 14.1m · 17.2m | 2.1× efficiency spread | 797f / 955e | Revolve vs. Stacked Extrude DPO Pairs |
verify_step.py)| Reward Term | Weight | ISO-10303-21 B-Rep Mathematical Invariant | Model Failure Modes Caught |
|---|---|---|---|
| Principal-Axis Bounding Box IoU (s_bbox) | 35% | Sorted principal extents (ΔX, ΔY, ΔZ) Volumetric Intersection-over-Union | Scale hallucination, aspect-ratio distortion, wrong extrusion depth |
| Closed-Shell Volume & Surface Area (s_vol_area) | 25% | Signed tetrahedral mesh divergence volume & triangulated B-Rep surface area | Missing pockets, un-shelled solid blocks, wall-thickness errors |
| Topological Graph & Euler-Poincaré (s_topology) | 20% | CLOSED_SHELL >= 1, OPEN_SHELL == 0, face/edge/vertex counts & χ = V - E + F | Non-manifold gaps, self-intersecting fillets, dangling surfaces |
| Analytical Surface & Bore Spectrum (s_surface) | 20% | PLANE, CYLINDRICAL, CONICAL, TOROIDAL distributions & sorted bore radii | Missing counterbores, wrong bolt-circle hole radii, omitted chamfers |
The 4 Core Architectural Pillars
Replacing subjective VLM judges with exact B-Rep geometric verifiers and synchronized human CAD telemetry.
Deterministic .STEP B-Rep Verifiers
Zero-dependency mathematical reward R ∈ [0, 1] via verify_step.py
VLM-as-a-judge cannot verify whether a 3D mechanical part is manifold or meets ±0.1mm bore tolerances. Our self-contained ISO-10303-21 .STEP parser grades exact B-Rep geometry out-of-band in milliseconds.
- 35% Principal-Axis Bounding Box IoU + 25% Closed-Shell Volume/Area
- 20% Euler-Poincaré topological graph + 20% Cylindrical bore spectrum
- Strict pass gate: Watertight manifold solid (OPEN_SHELL == 0) & R >= 0.85
Tri-Modal Human CAD CUA Trajectories
62 episodes · 11,792 lossless frames · 237,579 desktop telemetry events
Web scraping captures HTML and code, but real hardware engineering happens inside parametric 3D GUIs. We capture synchronized viewport frames, 6-DoF mouse/orbit/scroll telemetry, and native CAD deliverables.
- 6 CAD suites: SolidWorks (51), PTC Creo (9), FreeCAD, AutoCAD, Fusion, Inventor
- 100% paired with 48 ISO-128 4-view Third-Angle orthographic 2D blueprints
- 215 CAD deliverables: 115 native parametric files + 98 ISO-10303-21 .STEP solids
Procedural Parametric CAD Synthesizer
Scaling from human seed models to 50,000+ zero-contamination RLVR tasks
To saturate post-training RL clusters (PPO, GRPO, MCTS), our procedural engine mutates sketch dimensions, hole patterns, chamfers, and tolerances to synthesize held-out 2D blueprints paired with ground-truth .STEP solids.
- Zero pre-training memorization: procedurally parameterized mechanical assemblies
- Automated ISO-128 4-view orthographic + isometric blueprint rendering
- Continuous difficulty curriculum from single-extrude brackets to multi-bore housings
9-Pass OCR Privacy & Cryptographic Provenance
72,070 inspected frames · 0 PII leaks · SHA256SUMS across 401 files
Every public and enterprise trajectory passes a 9-pass CPU Vision OCR redaction pipeline, mandatory Native CAD + .STEP upload gates, and foreground window context verification before entering the training vault.
- 82/82 upstream worker bundles verified with live foreground app telemetry
- Prompt reference blueprints strictly isolated from ground-truth deliverables
- CC BY 4.0 public release on Hugging Face + air-gapped enterprise vaults
The Deterministic 3D CAD RLVR Training Loop
How post-training clusters combine human GUI behavior cloning with high-throughput B-Rep geometric RL rollouts.
ISO-128 Blueprint Prompt
4-view Third-Angle orthographic + isometric engineering drawing with dimension & tolerance callouts.
Human CUA SFT Warmstart
Pre-train on 237,579 timestamped SolidWorks & Creo mouse, keyboard, and 6-DoF viewport orbit actions.
Parallel CAD Rollout
Agent executes GUI actions or parametric CAD synthesis inside headless Linux CAD worker containers.
ISO-10303-21 .STEP Export
Candidate 3D solid is exported to standardized B-Rep .STEP format for out-of-band inspection.
verify_step.py Reward
Exact scalar reward R ∈ [0, 1] across BBox IoU, Volume, Euler-Poincaré topology & bore radii fed to GRPO/PPO.
Inside the Public spacetime-ml/cad-cua Release
Hosted publicly on Hugging Face under CC BY 4.0 across 3 dataset configurations (default, models, and benchmark). Every trajectory is captured from licensed mechanical engineers across SolidWorks (51), PTC Creo (9), FreeCAD, Autodesk Inventor, AutoCAD, and Fusion.
Scale 3D Spatial & CAD Reasoning on Your RL Cluster
License our private vault of multi-hour human SolidWorks & Creo trajectories alongside 50,000+ procedurally synthesized, zero-contamination 2D-blueprint-to-.STEP RLVR environments integrated directly into your Ray, Slurm, or PyTorch post-training cluster.
Custom CAD suite capture & air-gapped cluster deployment available.