RCAT：面向拥挤人群导航的机器人中心注意力时空建模方法

RCAT: A Spatial-Temporal Modeling Method for Robot Central Attention in Crowd Navigation

摘要: 在拥挤人群环境中实现安全、高效的自主导航是智能机器人面临的核心挑战。现有基于强化学习和自注意力机制的方法在时空建模方面已有一定成效，但往往缺乏对“机器人中心视角”的关注，导致机器人在复杂场景下难以准确聚焦于关键个体。为此，本文提出了一种机器人中心交叉注意力串行时空Transformer (RCAT)方法。该方法首先通过空间模块和时间模块串行建模行人的全局交互与时序动态，随后引入跨注意力机制，以机器人状态为查询向量，从时空特征中筛选与任务最相关的个体信息，从而实现更加任务导向的人群建模。在二维仿真环境中的实验结果表明，RCAT相比SARL方法，在不同人群密度下均表现出更高的成功率、更低的碰撞率以及更短的平均到达时间，并在累积奖励上取得显著优势。研究结果验证了RCAT在复杂人群导航任务中的安全性、效率和鲁棒性。

Abstract: Realizing safe and efficient autonomous navigation in crowded environments is the core challenge faced by intelligent robots. The existing methods based on reinforcement learning and self attention mechanisms have achieved certain results in spatiotemporal modeling, but often lack attention to the “robot central perspective”, which makes it difficult for robots to accurately focus on key individuals in complex scenes. Therefore, this article proposes a robot center cross attention serial spatiotemporal Transformer (RCAT) method. This method first models the global interaction and temporal dynamics of pedestrians in series through spatial and temporal modules. Then, a cross attention mechanism is introduced, using the robot state as the query vector, to filter individual information most relevant to the task from spatiotemporal features, thereby achieving more task oriented crowd modeling. The experimental results in a two-dimensional simulation environment show that RCAT exhibits higher success rates, lower collision rates, and shorter average arrival times compared to SARL methods under different population densities, and achieves significant advantages in cumulative rewards. The research results validated the safety, efficiency, and robustness of RCAT in complex crowd navigation tasks.