Zherong Pan has presented PSK (Penalized Single-loop K-level method), a way to optimise robot control through a physics simulation using only gradients and a projection, according to a paper on arXiv. Working back through physics steps usually requires solving large linear systems, the author writes.
PSK treats control over many simulation steps as nested optimisation problems recast as one min-max problem, updates all levels in parallel and replaces each contact constraint with a smooth wall of fixed width. The paper proves how quickly the method converges and where its limits lie.
In simulated navigation experiments with several robots among obstacles, PSK reached every goal with 16 levels and, given the same time, achieved lower loss than tuned unrolling in 17 of the 18 cases; implicit differentiation was lower in 10.



