Adaptive Deep Reinforcement Learning for Robust Control of Uncertain Dynamic Systems

Abstract

Uncertainty is a defining feature of many control problems: parameters drift, actuators saturate or slow down, and disturbances do not follow a single tidy model. When the mismatch between a nominal model and the real plant grows, classical designs that work well in a narrow envelope can lose tracking quality or violate safety limits. This paper examines adaptive deep reinforcement learning for robust control of uncertain nonlinear systems. The control task is posed as a constrained continuous-control problem. We train an actor–critic policy over a family of randomized dynamics and augment the observa- tion with lightweight identification cues extracted from short histories of state and input. At execution time, a small safety layer enforces hard command bounds. Across several uncertain benchmark systems, the resulting controller shows improved ro- bustness to parameter drift and disturbance bursts, with lower violation rates than fixed-gain baselines. We also report sensitivity studies (randomization width, observation latency, and history length) and summarize engineering lessons that matter for deployment.