Safety-aware multi-robot coordination in industrial cyber-physical systems using potential fields and handcrafted communication
TL;DR
The findings demonstrate the limitations of the current heuristic integration and support further research on dynamic task assignment, controlled component evaluation, stronger collision-avoidance mechanisms, learned communication, and end-to-end MARL training.
Abstract
Industrial cyber-physical systems increasingly rely on coordinated multi-robot fleets to perform manufacturing, logistics, and assembly tasks within shared workspaces. Achieving efficient task completion while limiting inter-robot collisions remains challenging as fleet size and interaction density increase. This paper presents a safety-aware heuristic coordination framework evaluated in a simulated industrial CPS environment based on the Multi-Agent Particle Environment. Five coordination policies are examined: Random navigation, Greedy nearest-workstation navigation, Greedy navigation with artificial potential-field repulsion, Optimal Reciprocal Collision Avoidance, and a composite Safety-Augmented Coordination Policy combining proportional navigation, static workstation assignment, artificial potential-field repulsion, and a manually defined four-dimensional communication vector. Experiments conducted over 30 paired episodes with four robots showed that Greedy achieved 100% ever-reached workstation coverage and simultaneous full coverage in 97% of the episodes. SACP achieved 85% ever-reached coverage and a 50% full-coverage success rate but recorded a higher collision frequency than Random and Greedy, indicating that its repulsion and communication mechanisms did not consistently improve collision avoidance under the tested configuration. Greedy+APF and ORCA provided stronger reactive baselines for interpreting the task-performance and collision behavior of SACP. Communication activity varied across navigation, collision-proximity, and workstation-arrival phases. Scalability experiments involving two to six robots showed increasingly negative normalized reward and higher collision exposure in the larger tested configurations, although no formal scaling law was established. The findings demonstrate the limitations of the current heuristic integration and support further research on dynamic task assignment, controlled component evaluation, stronger collision-avoidance mechanisms, learned communication, and end-to-end MARL training.