Open-RMF's congestion-aware planning prevents robot deadlocks in tight industrial corridors

Open-RMF, the open-source robot traffic management system, is exploring congestion-aware traffic planning for dense industrial environments such as semiconductor fabrication facilities. The motivating scenario involves many AMRs (autonomous mobile robots) operating simultaneously in long, narrow bidirectional corridors with bottleneck sections that only fit one or two robots, limited passing space, and dynamic robot entry. The core problem: strongly shortest-path-biased routing causes multiple robots to select the same corridors, leading to repeated head-on conflicts, frequent negotiation, excessive waiting, oscillating route decisions, and a rising deadlock probability as more robots enter the system. The proposed solution is a route evaluation function that combines travel time with predicted lane occupancy, route overlap cost, opposite-direction conflict risk, bottleneck entry cost, expected waiting time, negotiation risk, and deadlock risk—explicitly allowing longer routes if they improve stability.

The architecture to achieve this comprises four cooperating layers. First, schedule-aware route planning generates multiple route candidates and evaluates them against committed and predicted traffic, choosing a slightly longer path when it significantly reduces future conflicts. Second, bottleneck reservation and admission control treats narrow corridors as shared resources—reserving a direction, preventing opposite-direction entry, limiting the number of admitted robots, queuing at holding points, and releasing the reservation once cleared. Third, localized replanning only adjusts the affected robots when conditions change, avoiding global replans. Fourth, negotiation remains a fallback for unexpected spatiotemporal conflicts rather than the primary mechanism for predictable congestion. The forum post also raises implementation questions for the Open-RMF community: whether these responsibilities should live in the path planner, a route-candidate evaluator, or the negotiation system; whether multiple candidate evaluation beats modifying A*'s accumulated cost; and whether existing mutex groups are sufficient for directional corridor control. The legacy rmf_traffic system's limitations and next-generation traffic-management interface extension points are also under discussion, with the ultimate goal of preventing conflicts during planning rather than resolving them reactively.