Simulating virtual crowds can bring significant economic benefits to various applications, such as film and television special effects, evacuation planning, and rescue operations. However, the key challenge in crowd simulation is ensuring efficient and reliable autonomous navigation for numerous agents within virtual environments. In recent years, deep reinforcement learning has been used to model agents’ steering strategies, including marching and obstacle avoidance. However, most studies have focused on simple, homogeneous scenarios (e.g., intersections, corridors with basic obstacles), making it difficult to generalize the results to more complex settings. In this study, we introduce a new crowd simulation approach that combines deep reinforcement learning with anisotropic fields. This method gives agents global prior knowledge of the high complexity of their environment, allowing them to achieve impressive motion navigation results in complex scenarios without the need to repeatedly compute global path information. Additionally, we propose a novel parameterized method for constructing crowd simulation environments and evaluating simulation performance. Through evaluations across three different scenario levels, our proposed method exhibits significantly enhanced efficiency and efficacy compared to the latest methodologies. Our code is available at
https://github.com/tomblack2014/DRL_Crowd_Simulation....

Efficient crowd simulation in complex environment using deep reinforcement learning

10.1038/s41598-025-88897-2