ElmNav: Attention-Guided, Risk-Aware Maskingfor Socially Compliant Dense Crowd Navigation

Recent years have seen a significant surge in the deployment of mobile robots in social environments. Specifically, quadruped, wheeled, and even bipedal systems—primarily used for repetitive tasks such as industrial inspections, warehouse automation, and construction monitoring—have recently become the focus of studies investigating their potential application in fields such as policing in public environments, emergency response, and crowd management during large-scale events. Consequently, these robotic platforms must be capable of navigating crowded and dynamic environments to enable seamless integration into such settings. This paper introduces ElmNav, a reinforcement learning–based algorithm that leverages attention mechanisms and obstacle filtering to achieve collision avoidance in crowded environments. Two variants were developed and tested in simulation under varying degrees of crowd density. The first variant, Elm-DBM, performs potential collision detection and human filtering based on the distance to members of the immediate crowd. The second variant, Elm-CLM, filters humans by assigning a collision likelihood to each individual based on their current trajectory. Following testing in environments containing 5, 10, and 20 humans, Elm-DBM achieved a success rate of 86.7%, while Elm-CLM achieved a success rate of 88.7%. Both variants significantly outperformed the compared state-of-the-art solutions, indicating that the proposed algorithm is a promising approach suitable for deployment on physical robotic systems.