OccamoLab — an independent research group. We build neural networks that stay small, exact, and auditable by composing them with known laws.
Research interests
- Composing learned models with exact laws — group structure, geometric deep learning
- Robot-arm calibration and inertial navigation
- Parameter-efficient models for edge and embedded deployment
LawKAN
Calibration without the accuracy floor.
Real robot arms deviate from nominal kinematics: joint compliance, play, and tolerances deform the composition law itself. Classical calibration assumes the nominal law and fits per-joint offsets, so it cannot represent such a deformation — accuracy stalls at a floor that more poses will not remove.
LawKAN puts the exact composition law into the layer and learns only its deformation. On real arms from ABB, KUKA, Franka and Kinova, a correction of a few hundred parameters beats classical Gauss–Newton calibration — small enough to audit, cheap enough to run at the edge.
LawKAN: Group-Law Layers for Composition — When Imposing a Known Law Pays Off
- Tiny models — tens to hundreds of parameters — matching or beating networks tens of times larger.
- Data-efficient: matches classical calibration from as few as 20 poses, where tuned neural baselines need far more.
- A learned correction ψ that localizes where and how the hardware deviates from the nominal law.
- A cheap pre-training span check that predicts win, tie, or loss before you train.
Contact
hello@occamolab.org github.com/occamolab
Commercial licensing of the LawKAN code is available on request.