The theory and methodology of two-sided games under asymmetric conditions have become a research focus, necessitating the development of intelligent decision-making strategies capable of overcoming disadvantages such as numerical inferiority and resource constraints. Targeting a typical 1v2 asymmetric pursuit–evasion scenario involving a high-speed vehicle and two pursuers, this paper first constructs a multidimensional game model that integrates coordinated configurations between pursuers. To address the issues of slow convergence and sparse rewards in conventional reinforcement learning under such complex conditions, the evasion task is decomposed into subtasks, and a Predefined Curriculum Learning - Soft Actor-Critic (PDC-SAC) algorithm is developed to learn phase-wise maneuvering strategies. Simulation results validate the effectiveness of the proposed framework in enabling the high-speed vehicle to successfully evade both pursuers, providing valuable insights for the engineering application of asymmetric intelligent confrontation.