Abstract: To address the issues of strong nonlinear coupling between axial pressing force and rotational motion, as well as the difficulty of sim-to-real transfer in screwdriving operations using a dexterous hand, a reinforcement learning based arm-hand coordinated dexterous screwdriving strategy and a zero-shot transfer method were studied. Distinguishing from existing dexterous manipulation works that mostly focus on object rotation in free space, this paper addresses the tool screwdriving task with sustained contact constraints and proposes an arm-hand coordinated teacher-student distillation framework. An arm-hand collaborative hardware platform consisting of a 7-DOF force-controlled robotic arm, a three-finger dexterous hand, and tactile sensors was established. The screwdriving task was formulated as a Markov decision process, and a composite reward function was designed. The teacher policy was trained in simulation using the PPO algorithm, and the student policy relying only on proprioception and tactile features was distilled via behavioral cloning. Simulation training results show that the student policy converges within about 100 iterations. In zero-shot real-robot transfer experiments with ±1 mm initial position deviation, the success rate reached 95% over 20 trials. This study provides a feasible technical pathway for zero-shot sim-to-real transfer of dexterous tool manipulation.