This study constructs an interdisciplinary project-based learning model based on artificial intelligence, with deep reinforcement learning as the core algorithm, to realize learner dynamic perception of state, intelligent task allocation and real-time optimization of feedback. The model drives adaptive decision-making through joint state space, weighted reward function and Q-learning iterative update. Empirical analysis shows that the proposed model has a task completion rate of 92%, which is 35.3% higher than traditional project-based learning. The degree of interdisciplinary knowledge integration and collaboration efficiency are significantly improved, and the variance between groups is reduced from 17.1 to 5.6. This model effectively alleviates the problems of knowledge silos and uneven distribution, and provides a technical path for the reform of engineering education.