Research
Compositional-structure evals
Official LEGO models ship as .mpdfiles carrying the designer's authored sub-assemblies — the wheels, the chassis, the cabin a booklet would have you build separately. These evals strip that structure away, hand a model to an algorithm as a flat part list plus its physical joint graph, and score how well the designer's decomposition is recovered. Ground truth is the designer's own partition.
Assembly benchmark
The agent rebuilds official sets from step-by-step instruction images with a fixed part inventory. Survival curves: per-step BOM F1 and placement accuracy.
Decompose benchmark
Browse scored partitions per set: ground truth vs baselines vs AI, colored by module in 3D, with ARI / modularity / interface-cut.
Graph dataset
500 OMR models as graphs — nodes are parts, edges are physical joints, labels are the designer's sub-assemblies. Stats and per-model viewer.
Technic mechanisms
Gear trains, worm drives and gearboxes authored as LDraw and verified through headless LeoCAD renders.
Elevator (demo)
Motor + winch + pulley lifting an elevator, as a mode-switching lumped model: stall, back-driving, slack rope, snap — the Simulink way.
Physics (demo)
Rapier rigid-body world: the motor torques only the 8t pinion; the 24t turns because teeth push teeth. The 3:1 ratio is measured, not scripted.
Motor sim (demo)
A Powered Up motor spins an 8t→24t gear pair via a live DC-motor model: voltage, load, stall — the actuator model for the physics bridge.
RC readiness (demo)
Can this set drive? Functional-role extraction, wheel/axle detection, the RC checklist, and a kinematic drive simulation on a real 1990s car.
Metrics
ARI Adjusted Rand Index
Chance-corrected agreement with the designer's grouping. 1 = recovered the exact sub-assemblies; ≈0 = random.
Q Newman modularity
Intrinsic quality on the joint graph — dense inside modules, sparse between. Needs no ground truth.
Interface cut lower is better
Fraction of physical joints crossing module boundaries. Clean sub-assemblies attach at few, well-defined points.