fe41fa3b00
Add complete model layer with GIN and GAT encoders, 5 task-specific prediction heads, uncertainty-weighted multi-task loss (Kendall et al. 2018), and config-driven factory functions. New modules: - graph_conv.py: datagen dict -> PyG Data conversion (22D node/edge features) - encoder.py: GINEncoder (3-layer, 128D) and GATEncoder (4-layer, 256D, 8-head) - heads.py: edge classification, graph classification, joint type, DOF regression, per-body DOF tracking heads - assembly_gnn.py: AssemblyGNN wiring encoder + configurable heads - losses.py: MultiTaskLoss with learnable log-variance per task - factory.py: build_model() and build_loss() from YAML configs Supporting changes: - generator.py: serialize anchor_a, anchor_b, pitch in joint dicts - configs: fix joint_type num_classes 12 -> 11 (matches JointType enum) 92 tests covering shapes, gradients, edge cases, and end-to-end datagen-to-model pipeline.
Kindred Solver
Assembly constraint solver for Kindred Create. Combines a numerical multibody dynamics engine (OndselSolver) with a GNN-based constraint prediction layer.
Components
OndselSolver (C++)
Numerical assembly constraint solver using multibody dynamics. Solves joint constraints between rigid bodies using a Newton-Raphson iterative approach. Used by FreeCAD's Assembly workbench as the backend solver.
- Source:
OndselSolver/ - Entry point:
OndselSolverMain/ - Tests:
tests/,testapp/ - Build: CMake
Theory: MbDTheory
Building
cmake -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build
ML Solver Layer (Python)
Graph neural network that predicts constraint independence and per-body degrees of freedom. Trained on synthetic assembly data generated via the pebble game algorithm, with the goal of augmenting or replacing the numerical solver for common assembly patterns.
- Core library:
solver/ - Data generation:
solver/datagen/(pebble game, synthetic assemblies, labeling) - Model architectures:
solver/models/(GIN, GAT, NNConv) - Training:
solver/training/ - Inference:
solver/inference/ - FreeCAD integration:
freecad/ - Configuration:
configs/(Hydra)
Setup
pip install -e ".[train,dev]"
pre-commit install
Usage
make help # show all targets
make dev # install all deps + pre-commit hooks
make test # run tests
make lint # run ruff linter
make check # lint + type-check + test
make data-gen # generate synthetic data
make train # run training
make export # export model
Docker is also supported:
docker compose up train # GPU training
docker compose up test # run tests
docker compose up data-gen # generate synthetic data
Repository structure
kindred-solver/
├── OndselSolver/ # C++ numerical solver library
├── OndselSolverMain/ # C++ solver CLI entry point
├── tests/ # C++ unit tests + Python tests
├── testapp/ # C++ test application
├── solver/ # Python ML solver library
│ ├── datagen/ # Synthetic data generation (pebble game)
│ ├── datasets/ # PyG dataset adapters
│ ├── models/ # GNN architectures
│ ├── training/ # Training loops
│ ├── evaluation/ # Metrics and visualization
│ └── inference/ # Runtime prediction API
├── freecad/ # FreeCAD workbench integration
├── configs/ # Hydra configs (dataset, model, training, export)
├── scripts/ # CLI utilities
├── data/ # Datasets (not committed)
├── export/ # Model packaging
└── docs/ # Documentation
License
OndselSolver: LGPL-2.1-or-later (see LICENSE) ML Solver Layer: Apache-2.0