Abstract: Current control technologies for quadrotors mainly adopt the approach of utilizing positive thrust from the motor, limiting the application potential of bidirectional motors. To improve the maneuverability of quadrotors, expand their action space, and achieve agile flight and rapid control, a bidirectional thrust control method for quadrotors based on deep reinforcement learning is proposed. This method combines deep reinforcement learning with bidirectional thrust control for quadrotors for the first time. Based on the dynamic model of quadrotors with bidirectional thrust, a neural network controller based on deep reinforcement learning is designed to control the expected low-level thrust of 4 motors, realizing rapid hovering under extreme conditions. In addition, simulations of quadrotor stable hovering control under extreme conditions, such as large attitudes, high speeds, and high angular speeds, are conducted. The experimental results show that, compared to existing controllers using positive thrust, the proposed bidirectional thrust controller performs smoother actions, with smaller state fluctuations, shorter control times, and stronger robustness, effectively improving the control performance of the quadrotor.