Our group aims to build a control framework in which learning-based components such as vision-language models and physics-informed prediction models are combined with control design, where stochastic control theory is responsible for suppressing the effect of plant randomness whose realization is not available to those components in advance. This talk begins with an informal picture of this direction and a brief review of stabilization theory for discrete-time systems whose dynamics are determined by a general stochastic process. The main part is then devoted to a specific class within this setting, descriptor systems described by random polytopes, for which gain-scheduled state feedback controllers can be designed via linear matrix inequalities. This part is based on recent joint work with Dr. Dimitri Peaucelle and colleagues, which was presented at the IFAC World Congress 2026.