Dynamic obstacle avoidance method for 3D animation group based on multi-agent reinforcement learning

Abstract

At the forefront of 3D animation, virtual reality and robotics, it is difficult to realistically and efficiently simulate the group behavior of a large number of people and objects. Previously, traditional rule-based or optimization-based methods often encountered critical problems such as local collisions, unnatural trajectories, or the inability to escape deadlocks—especially in complex, high-density, and dynamic environments. In addition, these old methods also have limitations in overall coordination and flexibility, as they typically rely on local optimal calculations that fail to account for global group dynamics, inevitably leading to deadlocks in high-density continuous control scenarios. To tackle these specific limitations, our proposed approach integrates a localized attention mechanism. By supplying the transformer encoding into a policy network, each “agent” effectively collects and filters information about nearby agents and dynamic obstacles, proactively anticipating collision threats to maintain global coordination. It is noteworthy that the proposed method utilizes a comprehensive reward function that balances obstacle avoidance, movement efficiency, and path fluency. We have fully tested this method called Attention-MAPPO for populations with low/high density, static/dynamic barriers in the Unity 3D modeling environment. The experimental results show that the average obstacle avoidance success rate of this method in all tasks is 92.6%, and the average movement time is 12.4 s. Compared with the baseline standard MAPPO, its success rate has increased by an average of 4.4%, and the average movement time has been shortened by 6.8%, showing significant performance improvement. Furthermore, ablation studies on the attention mechanism clearly demonstrate its crucial role in achieving optimal performance in highly complex dynamic environments. This study provides a powerful and scalable solution for a real, effective team simulation that has a big impact on interactive virtual environments, game development and intelligent robots.