Specialized Deep Residual Policy Reinforcement Learning Framework for Safe and Adaptive Continuous Control

ABSTRACT Traditional controllers have limitations as they rely on prior knowledge about the physics of the problem, require modelling of dynamics, and struggle to adapt to abnormal situations. Deep reinforcement learning (DRL) offers a promising alternative by learning policies through exploration, but its black‐box nature and reliance on random exploration pose challenges in safety‐critical environments. Recognizing that conventional controllers and DRL have complementary strengths, we propose a novel hybrid framework to overcome challenges in both conventional control systems and DRL. This framework integrates residual policy learning, a cycle of learning approach, and a specialized reinforcement learning agent for safety‐critical, continuous control. Residual policy learning enables collaboration between DRL and conventional controllers, the cycle of learning improves learning efficiency by leveraging expert trajectories, and a specialized reinforcement learning agent optimizes policy learning in critical states using an input–output hidden Markov model. The framework is validated on the Tennessee Eastman process through experiments that analyse synchronization, activation mechanisms and an ablation study.