Full-duplex (FD) and reconfigurable intelligent surface (RIS) are potential technologies for achieving wireless communication effectively. Therefore, in theory, the RIS-aided FD system is supposed to enhance spectral efficiency significantly for the ubiquitous Internet of Things devices in smart cities. However, this technology additionally induces the loop-interference (LI) of RIS on the residual self-interference (SI) of the FD base station, especially in complicated urban outdoor environments, which will somewhat counterbalance the performance benefit. Inspired by this, we first establish an objective and constraints considering the residual SI and LI in two typical urban outdoor scenarios. Then, we decompose the original problem into two subproblems according to the variable types and jointly design the beamforming matrices and phase shifts vector methods. Specifically, we propose a successive convex approximation algorithm and a soft actor–critic deep reinforcement learning-related scheme to solve the subproblems alternately. To prove the effectiveness of our proposal, we introduce benchmarks of RIS phase shifts design for comparison. The simulation results show that the performance of the low-complexity proposed algorithm is only slightly lower than the exhaustive search method and outperforms the fixed-point iteration scheme. Moreover, the proposal in scenario two is more outstanding, demonstrating the application predominance in urban outdoor environments.