COLLISION AVOIDANCE PRIORITY FOR TRAFFIC CONGESTION IN MULTI-AGENT NAVIGATION Mathias Della Giustina * , Ubirajara Franco Moreno † , Henrique Simas ‡ * Universidade Federal de Santa Catarina Departamento de Engenharia Mecˆanica Florian´opolis, Santa Catarina, Brasil † Universidade Federal de Santa Catarina Departamento de Automa¸ c˜ao e Sistemas Florian´opolis, Santa Catarina, Brasil ‡ Universidade Federal de Santa Catarina Departamento de Engenharia Mecˆanica Florian´opolis, Santa Catarina, Brasil Emails: mathiasdella@hotmail.com, ubirajara.f.moreno@ufsc.br, henrique.simas@ufsc.br Abstract— In this paper, we present an approach to manage crowds of autonomous agents when navigating around narrow areas of a scenario in a way that ensures every agent is able to reach its destination. In our formulation, agents that find themselves stuck in the crowd receive a priority number which signals nearby agents to steer away from them. This new technique will be tested on a simulated environment and compared to other algorithms that do not contain such crowd priority management. Keywords— Navigation, Agent, Crowd, Pathfinding, Collision Avoidance Resumo— Neste artigo, ser´a apresentado uma abordagem de controle de multid˜ao para agentes autˆonomos que navegam por ´ areas estreitas de um cen´ario de forma a garantir que todos agentes sejam capazes de chegar aos seus destinos. Nessa formula¸c˜ao, agentes que se encontram presos e parados na multid˜ao recebem um n´ umero de prioridade que sinaliza agentes pr´oximos para se afastarem. Esta t´ ecnica ser´a testada em um ambiente simulado e comparada com outros algoritmos que n˜ ao possuem essa gest˜ao de multid˜ao com prioridade. Palavras-chave— Navega¸c˜ ao, Agente, Multid˜ao, Enxame, Desvio de Colis˜ ao 1 Introduction Navigation algorithms for autonomous agents are codes capable of guiding agents step by step from their current position towards their goal position. To achieve this result, these algorithms must plan a path for the agent to follow, and as the agent follows the path, the algorithm must tell the agent how exactly it should move each step and how to avoid collisions with nearby obstacles such as nearby agents or nearby walls and buildings. In order to plan paths for the agents, the al- gorithm must first have a geometrical represen- tation of the environment in which the agents are navigating, so the algorithm needs some sort of a map containing the environment obstacles. These maps are traditionally represented with grids (Yap, 2002), but more recently, they have been represented with polygons (Leonard, 2014), (Pratt, 2014) and triangles (Chew, 1987). Once there is a map representation of the sce- nario available, it is possible to plan paths for the agents using pathfinding algorithms such as (Kallmann, 2010) or (Demyen and Buro, 2006) in case of triangle representations or (Dechter and Pearl, 1985) for grid representations. The final core part of a navigation algorithm is to guide the agent through its planned path while avoiding collisions with nearby obstacles. As these obstacles may be moving and their movement may be unknown or unpredictable, avoiding collisions with nearby moving obstacles is not as easy as it seems. There are many important researches on the subject such as Reynolds steering behaviors (Reynolds, 1999), ClearPath (Guy et al., 2009) and Relative Velocity Obstacles (Van Den Berg et al., 2011). Additionally, navigation algorithms can have extra functions that are able to modify agents behaviors. There are researches on steering the agents with least effort approaches (Guy, Chhugani, Curtis, Dubey, Lin and Manocha, 2010), researches on giving agents group behav- iors so that they stay grouped during collision avoidance (He et al., 2016) and attempts to copy humans movement behaviors (Guy, Lin and Manocha, 2010). More specifically in managing crowds, there are some recent researches that at- tempt to enforce crowd rules (Foudil and Noured- dine, 2006) to solve some situations, as well as some work on long range movement predictions (Golas et al., 2014) that allows agents to predict and start avoiding from collisions a long time be- fore they are a threat. XIII Simp´ osio Brasileiro de Automa¸ c˜ ao Inteligente Porto Alegre – RS, 1 o – 4 de Outubro de 2017 ISSN 2175 8905 912