Mahalanobis-Guided Latent OOD Detection for Hybrid ES-DRL Control in Time-Varying Systems

This paper studies latent out-of-distribution detection for deciding when a hybrid controller should use a fast reinforcement-learning policy and when it should switch to robust bounded extremum seeking. A variational autoencoder is trained on in-distribution beam-profile observations, and Mahalanobis distance in the learned latent space provides an interpretable test-time switching signal.

Status: Accepted at the ICML RLxF Workshop 2026.

View the ICML RLxF 2026 poster

Recommended citation: Shaifalee Saxena and Alexander Scheinker. (2026). "Mahalanobis-Guided Latent OOD Detection for Hybrid ES-DRL Control in Time-Varying Systems." ICML RLxF Workshop 2026.
Download Paper