Constrained Value Decomposition for Communication-Gated Task Allocation Under Asynchronous UAV Swarm Communication

Consensus-based distributed task allocation in UAV swarms resolves conflicts through repeated neighbor communication, saturating the ad hoc channel with packet loss and latency; reducing transmission frequency without sacrificing allocation quality is the key challenge. Prior communication-reduction schemes use hand-designed, offline-fixed rules, decoupled from the allocation objective and unable to adapt to time-varying channel contention. We instead learn a per-agent binary send/suppress gating policy; since gating trades communication cost against allocation quality, we cast it as a learning problem in which a swarm-level communication-cost penalty regularizes the shared reward. Real swarms are asynchronous: only a subset of agents decides per step, leaving the valuedecomposition joint action partially observed. We propose the Virtual Synchrony Task Allocation (VSync-TA) framework. Its learning method, Virtual Synchrony Proxy–Constrained QMIX (VSP-CQMIX), addresses communication-gated asynchronous UAV task allocation. First, VSP-CQMIX adopts the standard fixed-slot VSP representation and QMIX value decomposition to organize time-varying decision sets into fixed-dimensional joint training data. Second, a swarm-level communication-cost penalty is incorporated into the QMIX objective through slot-conditioned penalty-value estimates and projected nonnegative multiplier updates. The updated multipliers shape the shared reward and thereby affect centralized value-decomposition training. On a 30-UAV, 150-task simulator, VSP-CQMIX has a mean task-conflict rate of 5.23%, compared with 6.31% for ACBBA, corresponding to a 17.1% relative difference in the observed means. Its mean communication overhead and convergence time are 91.3% and 86.3% lower than those of ACBBA, respectively, while exhibiting a favorable allocation–communication trade-off relative to the constrained multi-agent baselines; a search-and-rescue case study further illustrates the influence of the communication policy on task-allocation results and planned trajectories.