<p>Robotic arms play a crucial role in modern manufacturing and automation, where trajectory planning requires navigating waypoints, avoiding obstacles, and minimizing energy consumption, making it a constrained optimization problem. However, collision detection is a computationally intensive geometry problem, and the collision-free region can be narrow and discontinuous, which challenges traditional optimization algorithms in reducing computational time and increasing the number of feasible solutions. In this paper, a multi-classifier-assisted particle swarm optimization (MCAPSO) is proposed to address the expensive constrained optimization problem (ECOP) in the obstacle avoidance trajectory planning. Specifically, several CNN classifiers are constructed to predict the feasibility of generated trajectories, thereby accelerating collision detection. Throughout the optimization process, Q-learning is employed to maintain prediction accuracy by dynamically determining when to update the classifiers and controlling the frequency of using real evaluations. Meanwhile, a multi-strategy constraint repair technique is developed to accelerate the optimization of infeasible solutions toward feasible regions. Additionally, the particle swarm optimization algorithm has been extensively enhanced to improve global search capability and escape local optima. Experimental results demonstrate that MCAPSO effectively handles scenes with varying levels of obstacle complexity and exhibits competitive performance compared to some state-of-the-art methods, offering a promising new approach for real-world ECOPs.</p>