From Discrete Plans to Real-World Execution: A World-Model-Driven Framework for Execution-Aware Multi-Agent Path Finding

Researchers present ExecTimeNet, a learned world model that bridges the gap between discrete multi-agent path finding (MAPF) planning and real-world robot execution by predicting how planned paths perform on physical systems with realistic dynamics and delays. The framework includes REMAP, which integrates execution-time estimation into planning, and ESADG, a post-planning optimizer that achieves up to 40% improvement in execution efficiency while maintaining path feasibility.

This research addresses a critical challenge in robotics: the persistent gap between what algorithms plan in simulation and what actually happens in deployed systems. MAPF has advanced significantly in computing collision-free paths for hundreds of agents, yet these solutions rely on simplified models that ignore real-world complexities like heterogeneous robot dynamics, asynchronous interactions, and communication delays. The proposed ExecTimeNet model learns to predict execution outcomes by mapping discrete planning actions to their actual wall-clock completion times and kinodynamic states, essentially creating a bridge between the discrete planning world and continuous physical reality.

The significance extends beyond academic novelty. Automated warehousing and manufacturing systems represent massive economic value, and inefficiencies from planning-execution mismatches directly translate to productivity losses. REMAP's integration of execution-time awareness into the planning phase itself represents a fundamental shift from post-hoc adjustments to proactive, execution-aware optimization. ESADG's ability to optimize scheduling while preserving feasibility offers a practical tool for existing deployed systems.

The evaluation methodology demonstrates maturity: validation spans high-fidelity simulation with up to 300 agents, transfer learning to unseen scenarios, and crucially, physical robot deployment showing 15.3% real-world improvement. This sim-to-real validation gap has historically been where many robotics advances falter, making the demonstrated transfer particularly valuable. For stakeholders in warehouse automation and manufacturing technology, this work suggests that next-generation MAPF solvers will increasingly account for execution realities, potentially enabling higher utilization rates and faster throughput in existing infrastructure.

• →ExecTimeNet learns to predict how discrete MAPF plans execute on physical robots with realistic dynamics and delays
• →REMAP reduces execution delays by up to 21% by incorporating execution-time estimation directly into path planning
• →ESADG achieves up to 40% normalized improvement in execution schedules while maintaining path feasibility
• →Physical robot experiments validate 15.3% real-world execution time reduction, confirming sim-to-real transfer effectiveness
• →The framework addresses a critical challenge in scaling multi-agent robotics for warehouse automation and manufacturing