Hybrid Centralized Distributed Control for Lifelong MAPF over Wireless Connections

TL;DR

This work addresses this joint control--communication problem and proposes a hybrid centralized--distributed scheme: a centralized cloud policy sends small residual corrections only when selected, while a lightweight on-board Gated recurrent unit (GRU) policy provides a safe default fallback when wireless connection is not available.

Abstract

In lifelong multi-agent path finding (MAPF) with many robots, unreliable wireless links and stochastic executions are the norm. Existing approaches typically either rely on centralized planning under idealized communication, or run fully distributed local controllers with fixed communication patterns; they rarely couple communication scheduling with policy learning, and thus struggle when bandwidth is scarce or packets are frequently dropped. We address this joint control--communication problem and propose a hybrid centralized--distributed scheme: a centralized cloud policy sends small residual corrections only when selected, while a lightweight on-board Gated recurrent unit (GRU) policy provides a safe default fallback when wireless connection is not available.

Figures (4)

• Figure 1: Sample Obstacle map and its Radio Map.
• Figure 2: Hybrid cloud--edge architecture. The local GRU policy (red) generates masked action logits from $3{\times}W_{\text{FOV}}{\times}H_{\text{FOV}}$ observations and guarantees autonomy during dropouts. The cloud (blue) builds a belief map, predicts event risks with EEM, and uses a shared encoder plus a downlink allocator (W/P expert heads with gating) to compute residual corrections that are injected into on-board logits when the downlink is available, producing the final policy.
• Figure 3: TNCT (128-step horizon) across MovingAI maps with $N\!\in\!\{32,64,128,256\}$ and channel ratio $r_{ch} \!\in\!\{0,25,50,75,100\}$, under a unified execution model: egocentric $7{\times}7$ FOV, a 0.90-centered stochastic transition kernel, cloud overriding local upon packet reception, and safety via attempt-arbitration.
• Figure 4: Performance and training of the hybrid controller. (a) TNCT versus channel ratio on random MovingAI maps with 256 agents; the hybrid policy matches or exceeds the communication-aware ODrM+A* baseline at all bandwidths. (b) TNCT during cloud training on a $64{\times}64$ random map; adding EEM+belief yields the largest gain over local-only and ablated variants.