Aiming at the traditional attitude control failure in on-orbit service dual-arm space robots under extreme abnormal conditions such as flywheel and engine malfunctions， an emergency attitude control algorithm for dual-arm space robots based on reinforcement learning is proposed. This approach achieves limited attitude recovery of the spacecraft using only the two robotic arms configured on the spacecraft which differs from traditional attitude control algorithms. A physical environment for algorithm training is constructed and a model-free proximal policy optimization （PPO） algorithm is used for attitude control. By incorporating the kinematic constraints of manipulators movements during on-orbit operations， the reward function is designed to optimize the precision of spacecraft attitude control. To validate the effectiveness of the proposed strategy， numerical simulations of the space robot attitude recovery are conducted in the MuJoCo environment. The adaptability of the algorithm is evaluated under various conditions， including various masses of the base， various masses of the end. Results demonstrate that the reinforcement learning method is suitable for spacecraft limited attitude control and show a certain robustness without the need of parameter fine-tuning.