Should I Replan? Learning to Spot the Right Time in Robust MAPF Execution

Researchers at the Czech Technical University in Prague have developed a neural network that decides the optimal moment to replan multi-agent pathfinding (MAPF) during execution. The study, titled "Should I Replan? Learning to Spot the Right Time in Robust MAPF Execution," addresses a critical gap: MAPF assumes perfect synchronization, but real-world robot fleets face delays from internal or external factors. Traditional robust methods like the Action Dependency Graph (ADG) synchronize risky actions but often increase execution costs by forcing agents to wait. The team's feed-forward neural network estimates the benefit of single replanning based on ADG-derived features, trained on 12,000 labeled experiments. Results show it reduces delay impact by up to 94.6% of achievable reduction, enabling safer and cheaper fleet operations.

The approach uses a fully connected neural network with novel ADG-based features that capture the robust execution state and potential delay impacts. By predicting when replanning—either rescheduling or full path replanning—will lower execution costs, the method avoids unnecessary computational overhead from replanning when benefits are negligible (e.g., when plans are identical). This is particularly valuable for logistics, warehouse automation, and autonomous vehicle coordination, where even small delays compound into significant inefficiencies. The paper, submitted for double-blind review to IEEE, offers a practical solution for making multi-agent systems both robust and cost-effective in real-world conditions.