Abstract: Addressing the insufficient consideration of agent heterogeneity and the injection of human experience and cognition in path planning research, this paper proposes a method for heterogeneous multi-agent path planning with human-machine collaborative decision-making in mixed reality scenarios. Firstly, we designed a heterogeneous multi-agent deep reinforcement learning method based on danger-guided point selection and local collision avoidance using Reciprocal Velocity Obstacles (RVO). By considering the heterogeneity of agents, we utilized both global guidance and local instructions to set up a reward function suitable for heterogeneous intelligent agents, effectively addressing the sparse reward problem. Secondly, in the context of mixed reality-based virtual and real agent interaction, human experience is integrated to refine paths planned by the intelligent algorithm, enhancing human guidance and achieving human-machine collaborative decision-making in path planning, thereby compensating for the limitations of the intelligent algorithm. Finally, experiments conducted in 2D, 3D, and mixed reality environments demonstrate that the proposed method not only applies to heterogeneous multi-agent path planning but also achieves human-machine collaborative decision-making planning in mixed reality scenarios. The results show that the proposed method outperforms baseline algorithms in terms of success rate, convergence, path length, number of turning points, and other evaluation metrics.