MDGAM framework boosts multi-robot task scheduling under communication limits

Cooperative task scheduling for distributed multi-agent systems gets a major upgrade with MDGAM, a neural framework from researchers Licheng Wang, Mingtao Huang, and Yuan Shen. Published on arXiv (2608.00648), MDGAM targets distributed multi-robot task allocation (MRTA) where agents make decisions from partial, dynamic observations under real-world communication constraints. Unlike conventional heuristics that rely on handcrafted bidding rules and repeated consensus, MDGAM uses an extended graph attention mechanism to jointly update node and edge features, then leverages multiple decoders to generate both task-selection decisions and communication messages. This design lets agents explicitly coordinate without needing global observations.

The training side introduces GRMAPG (group-relative multi-agent policy gradient), a critic-free algorithm that constructs group-relative advantages from equivalent task-planning instances. This replaces the critic network used in traditional multi-agent reinforcement learning (MARL), reducing training difficulty and improving convergence. Experiments across varying problem scales and communication ranges show MDGAM outperforms both heuristic and learning-based baselines in task-completion performance. Ablation studies confirm the contribution of each component, while complexity and generalization tests validate the approach's scalability—making it a promising foundation for real-world swarm robotics, warehouse automation, and autonomous vehicle coordination.