the shittiest least squares optimizer:
error function penalizes label not to be at center of mass of docking points, and docking points not to be at common center of mass.
we measure loss before and after applying a random gradient from a d-dimensional hypersphere.
if loss is better, just use the new configuration. do this 100 times per frame.
convergence is slow, but it's enough to start modeling the loss function.
#devlog #tukan
