Conflict-Aware Graph-Attention MAPPO for Cooperative Local Navigation of Multiple Mecanum Robots

Abstract

Cooperative local navigation of multiple mecanum robots requires efficient coordination around robot–robot conflicts, pedestrians, and static obstacles while preserving direct waypoint following on clear path segments. This paper presents Conflict-Aware Graph-Attention Multi-Agent Proximal Policy Optimization (CA-GAT-MAPPO), a learning-based residual control and coordination framework. Predicted closest-approach events construct a conflict-conditioned robot-interaction graph, so actor message passing is restricted to the local robot and its predicted conflict neighbors. A residual graph encoder preserves waypoint-conditioned state, while a bounded right-of-way coordinator and an interaction gate regulate longitudinal, lateral, and angular residual authority. State-dependent adaptive scalarization combines multiple reward components into a single training objective. The framework is evaluated within a common A*-based waypoint guide, optimal reciprocal collision avoidance (ORCA)-style prior, command-limiting, and safety-envelope interface shared by the compared controllers. Across three four-robot Robot Operating System 2 (ROS 2)/Gazebo scenarios, eight independently trained checkpoints per method–scenario pair were each evaluated in ten randomized episodes. Across the four learned controllers, all 960 main-comparison episodes were completed without a geometric collision or a recorded robot–robot or robot–pedestrian near-miss at the 0.1 s sampled poses under the shared execution boundary. CA-GAT-MAPPO obtained the lowest reported mean completion time, makespan, and waiting time in all three scenarios. The results support a descriptive efficiency advantage for the integrated intelligent-control stack under the evaluated conditions, without establishing universal superiority or a formal safety guarantee.