Efficient crowd simulation in complex environment using deep reinforcement learning.

Area of Science:

• Computer Science
• Artificial Intelligence
• Robotics

Background:

• Crowd simulation is vital for applications like film, evacuation planning, and rescue operations.
• Deep reinforcement learning (DRL) has been applied to agent steering but struggles with complex, heterogeneous environments.
• Existing DRL methods often fail to generalize beyond simple scenarios, limiting their practical use.

Purpose of the Study:

• To develop an advanced crowd simulation approach for efficient and reliable autonomous navigation in complex virtual environments.
• To improve the generalization capabilities of deep reinforcement learning in crowd simulation.
• To introduce a novel method for environment construction and performance evaluation in crowd simulation.

Main Methods:

• Combined deep reinforcement learning with anisotropic fields to provide agents with global environmental awareness.
• Developed a parameterized method for constructing complex crowd simulation environments.
• Evaluated the proposed method across three distinct scenario complexity levels.

Main Results:

• The proposed method achieved impressive motion navigation results in complex scenarios.
• Agents demonstrated efficient and reliable autonomous navigation without repeated global path computation.
• The approach significantly enhanced efficiency and efficacy compared to state-of-the-art methodologies.

Conclusions:

• The integration of DRL with anisotropic fields offers a powerful solution for complex crowd simulation.
• The novel approach improves agent navigation performance and generalizability in diverse virtual environments.
• This work provides a robust framework for crowd simulation applicable to various real-world challenges.