Policy-Gradient Transfer under a Shared State Estimator
This work investigates transferring control policies from a simulator to a physical plant when both use the same fixed state estimator. The key finding is that the difference between these two policy gradients is bounded linearly by noise-weighted model errors, requiring no guarantee on estimator accuracy. This allows for deriving an occupancy-weighted bound and establishing conditions for the simulator gradient to be an ascent direction for the plant.