Event

Code-Space Policy Search for Multi-Agent Path Finding

1 reports · 1 sources

Matthias Eder et al. propose a human-guided workflow that lets coding agents iteratively propose, implement, evaluate, diagnose, and improve multi-agent path finding algorithms in code space, combining simulation feedback with explicit correctness verification. The paper claims that on a challenge map from the 2024 League of Robot Runners competition, a policy found by two days of search outperforms a stronger locally reproduced contestant solution by 35.0% on the densest instances and generalizes to unseen maps; in a simulated warehouse with irregular lanes, continuous robot geometry, and asynchronous motion, one week of search discovered a new class of algorithms with 58.9% higher throughput than a calibrated proprietary solution. The authors also claim that, under stated assumptions, they prove the new algorithm is collision-free and that every accepted goal that has not been retracted or replaced can be completed in finite time.