Deterministic 3D CAD CUA & B-Rep RLVR Solutions
From post-training multimodal reasoning models with continuous .STEP geometric rewards to training 6-DoF desktop Computer-Using Agents across SolidWorks and PTC Creo.
Frontier Post-Training & Reasoning Teams
Scaling test-time compute and GRPO/PPO reinforcement learning with dense, uncheatable B-Rep geometric rewards R ∈ [0, 1].
Current frontier models fail 85.0%+ of the time when reconstructing 3D mechanical parts from 2D engineering drawings (Claude 4.5: 15.0% Pass@1, GPT-5.2: 10.0%, Gemini 3: 10.0%). VLM-as-a-judge rewards hallucinate correctness on broken meshes and miss internal counterbores. SpacetimeML's verify_step.py parses ISO-10303-21 .STEP Boundary Representations directly—delivering exact mathematical reward signals across bounding box IoU (35%), volume/surface area (25%), Euler-Poincaré topology (20%), and analytical surface/bore spectra (20%).
Dense Continuous Reward R ∈ [0, 1]
Eliminate sparse-reward collapse during GRPO/PPO rollouts with 4 continuous geometric sub-scores plus a strict watertight manifold pass gate (R ≥ 0.85).
2D-Blueprint-to-3D B-Rep Synthesis
Train multimodal reasoning models on 48 public ISO-128 4-view orthographic + isometric blueprint pairs (and 50,000+ private synthetic tasks) paired with ground-truth .STEP solids.
Test-Time Search & Best-of-N Verification
Score thousands of candidate CadQuery, OpenCASCADE, or feature-tree rollouts per second in <10ms on CPU without spinning up expensive VLM judges.
Zero Reward Hacking or Prompt Injection
Because verify_step.py inspects MANIFOLD_SOLID_BREP, CLOSED_SHELL, and ADVANCED_FACE entities out-of-band, models cannot game the reward without producing valid 3D geometry.
Computer-Using Agent (CUA) & Multimodal Vision Teams
Moving desktop agents beyond 2D web DOM clicking into high-precision 6-DoF commercial 3D CAD workstations.
Standard OSWorld and web-browsing CUA datasets only teach flat 2D form-filling. Operating professional engineering software requires continuous 3D spatial reasoning: orbiting the viewport camera to inspect occluded faces, selecting tangent edges, applying geometric sketch constraints, and navigating hierarchical feature trees. SpacetimeML provides 237,579 timestamped desktop telemetry events paired with 11,792 privacy-masked lossless PNG frames across 6 CAD suites.
237,579 Real 3D GUI Actions
High-frequency mouse coordinates, keyboard shortcuts, and 3D viewport scroll/orbit/pan/zoom telemetry across SolidWorks (51), PTC Creo (9), AutoCAD, FreeCAD, Fusion, and Inventor.
11,792 Lossless Vision-OCR Verified Frames
Curated from 72,070 raw high-resolution frames with 9-pass CPU Vision OCR verification (0 PII leaks), plus 62 cursor-annotated H.264 MP4s and 62 1-page storyboard PDFs.
Live Foreground Context Tracking
Every trajectory captures real-time window and dialog state transitions (app_context_updated_at), teaching agents when sketch solvers, extrude dialogs, or assembly mates are active.
100% Deliverable-Grounded Episodes
All 48 model-producing episodes are paired 1:1 with archived native parametric files and exported .STEP solids (models/cad-XXXXXX.zip), enabling end-to-end outcome evaluation.
Embodied AI, Robotics & Digital Manufacturing
Bridging foundation models from text and code to watertight ISO-10303-21 .STEP solids ready for CNC machining, FEA, and robotic assembly.
Text-to-3D mesh generators (NeRFs, Gaussian Splats, and triangulated STL diffusion) produce noisy, non-manifold surfaces that cannot be CNC-machined, simulated in FEA, or mated in robotic digital twins. Physical AI and industrial automation require exact analytical surfaces—true cylinders, planes, cones, and toroids defined in ISO-10303-21 .STEP format with editable native parametric history.
215 Production Mechanical CAD Deliverables
Includes 115 native parametric files (.sldprt, .sldasm, .prt, .asm, .fcstd, .ipt, .dwg, .f3d), 98 ISO-10303-21 .step/.stp B-Rep solids, and 2 exchange files.
Watertight Manifold Guarantee
Enforces CLOSED_SHELL ≥ 1 and OPEN_SHELL == 0 alongside exact Euler-Poincaré topology so generated parts load directly into CAM toolpath generators and MuJoCo/Isaac Sim.
Analytical Bore & Mating Surface Precision
Verifies exact cylindrical bore radii, counterbore depths, and planar mounting datums required for real-world fasteners, bearings, and robotic end-effector tooling.
Custom Enterprise CAD Domains
Commission bespoke trajectory and B-Rep RLVR datasets across aerospace brackets, automotive powertrains, microfluidic manifolds, and sheet-metal enclosures.
Why 3D Spatial AI Demands Deterministic B-Rep RLVR
A technical comparison between 2D web CUA datasets, VLM-as-a-Judge rewards, and SpacetimeML's ISO-10303-21 .STEP RLVR infrastructure.
| Evaluation Dimension | 2D Web DOM CUA Datasets | VLM-as-a-Judge / Mesh IoU | SpacetimeML CAD-CUA & B-Rep RLVR |
|---|---|---|---|
| Action & Observation Space | 2D DOM clicks & text inputs | Static rendered images | 237,579 6-DoF GUI actions + 11,792 PNGs |
| Geometric Ground Truth | None (no 3D representation) | Approximate triangle meshes (.stl) | 215 CAD files (115 native + 98 ISO-10303-21 .STEP) |
| Reward Signal & Hallucination | DOM state regex matching | High hallucination on occluded bores | Exact B-Rep R ∈ [0, 1] via verify_step.py (<10ms) |
| Topological & Manifold Check | N/A | Passes self-intersecting / open shells | Strict CLOSED_SHELL ≥ 1 & Euler V - E + F |
| Frontier Model Discriminative Power | Saturating on 2D browser tasks | Noisy ranking across runs | Reveals 85.0%+ frontier failure rate on 20 tasks |
Train Spatial Agents on Verifiable 3D Geometry
Download the public 401-file spacetime-ml/cad-cua release on Hugging Face, or partner with our engineering team to access 50,000+ private 2D-blueprint-to-3D-B-Rep RLVR tasks.