Adaptive optics: a reinforcement learning controller capable of starting up directly under the open sky

Researchers propose pre-calibrating the PO4AO controller using an internal light source, in order to correct for turbulence without the need for a linearly stable controller.

High-contrast imaging from the ground, which is useful in particular for observing exoplanets, requires the correction of the wavefront with extreme precision. The most sensitive wavefront analyzers are also the most non-linear, which puts the classic linear reconstructors and controllers of extreme adaptive optics in difficulty in difficulty.

Doing without the integrator to start learning

Reinforcement learning, and in particular the PO4AO (Policy Optimization for Adaptive Optics) method, has already shown that it can compensate for time delays, alignment errors, and moderate nonlinearity. But PO4AO learns its initial strategy from data collected in a closed loop by a conventional integrator controller, which becomes impossible when the analyzer is highly nonlinear or the conditions are harsh.

The authors propose to pre-calibrate PO4AO using an internal light source and the deformable mirror, in order to close the loop directly on the sky, without a stable integrator controller to prime the first models.