ABSTRACT For complex nonlinear systems with modeling difficulties and partial observability (only motion signals available), their control performance is significantly limited by model uncertainties, unmeasurable disturbances, and critically, lack of prior knowledge of system dynamics. To overcome these limitations, this article proposes a widely applicable reinforcement learning (RL)‐based output feedback control approach whose model‐free nature and adaptive learning capability enable deployment across diverse nonlinear systems beyond the initial constraints. An integrated controller‐observer design embeds an actor‐critic architecture to achieve model‐free observation and control. Within this framework, the actor dynamically compensates completely unknown unmodeled disturbances while the critic evaluates control performance. Crucially, an interactive mechanism is established where the actor directs observer‐based state reconstruction, with reconstructed states feeding back to both actor and critic. Rigorous theoretical analysis demonstrates that the proposed controller ensures semi‐global uniform ultimate boundedness (SGUUB) for all closed‐loop signals. Simulations validate the effectiveness of the proposed control method and its superior steady‐state performance under significant disturbances.
Zhou et al. (Tue,) studied this question.