Supervisor(s): Pavel Shchekaturov, Bogdan Jeram, Marcus Kasper (TBC) , Academic supervisor in control or machine learning (TBC) 

Abstract:

Reinforcement-learning (RL) control for adaptive optics (AO) now works on sky, but no observatory runs it in operations with a guaranteed safe fallback. This PhD builds a control engine in which an RL policy drives the AO loop while a runtime supervisor checks residuals, actuator stroke, inference deadlines and output validity, and hands control back bumplessly to a linear controller when limits are exceeded. 

The engine is a portable hard-real-time core with thin adapters. It is evaluated on sky on the SPARTA system of ERIS, whose adaptive secondary mirror is the closest VLT analogue of ELT M4 (the SPARTA system of SPHERE is a second option), through staged commissioning from replay to full authority. The thesis compares RL, neural and linear predictors on the same Hard Real-Time Core (HRTC) hardware, analyses switching stability, and measures the latency and jitter the engine adds compared with a well-tuned linear integrator. 

The key contribution is the supervised, safe RL core: whenever the RL output does not meet the safety or DM-stroke criteria, the system switches to the linear controller. After such a fallback, the RL engine resets its internal state and recalibrates its network weights, so that it can take over again once the safety and DM-command limits are met and the recalibrated policy has passed validation in shadow mode. 

GHOST bench integration is a stretch goal. 

Deliverables: a reusable engine and supervisor layer for VLT SPARTA systems, portable to the ELT environment as a pathfinder for ELT instrument RTCs. 

Time constraints: Ideally starting in 2027 (ELT scientific first light is planned for December 2030); ERIS technical time is needed in months 23–30.