Abstract: To address the problems that local interpolation methods rely heavily on key tool orientations and graph-search methods suffer from high computational cost in global tool-orientation planning for complex-surface five-axis machining, a global planning method based on improved deep reinforcement learning is proposed. Based on the discrete sampling of feasible tool-axis vectors at cutter contact points, the planning process is formulated as a Markov decision process (MDP), where the state space, action space, and orientation-variation constraints are defined. A reward function is constructed using the variation rate per unit path length and the second-order difference of the pitch and azimuth angles. Then, an ST-DDQN framework is developed by integrating double deep Q-network (DDQN) with a SumTree-based prioritized experience replay strategy. Comparative simulations are carried out on a free-form surface machining path with local obstacle constraints. Results show that the maximum variation rates of the pitch and azimuth angles obtained by the proposed method are 0.052°/mm and 0.11°/mm, respectively, and their second-order differences are both below 0.01°/mm². The average planning time is 65.43 s, which is 9.5% lower than that of quaternion interpolation and 57.0% lower than that of Dijkstra search. The proposed method provides an effective approach for efficient global planning of tool orientations in complex-surface five-axis machining under feasible-domain constraints.