SJRL algorithm beats search-based planners in lifelong warehouse robot pathfinding

Lifelong Multi-Agent Path Finding (LMAPF) is the backbone of automated warehouse systems, where robots continuously receive new goals after dropping off items. Most learning-based planners assume simplified kinematics, ignoring real-world constraints like turning radius and safety distances. To fix this, researchers led by He Jiang proposed SJRL (Search-Aided Joint Reinforcement Learning), built on a new realistic model called LMAPF-R2 that incorporates in-place rotation and robust safety constraints. These constraints dramatically increase coordination difficulty, especially in narrow aisles.

SJRL tackles the challenge through a dual strategy. First, it augments neural policies with Causal-PIBT, a single-step search-based planner that resolves collisions and propagates robot intentions. Second, it uses a unified RL framework that jointly optimizes agent actions and an environment policy, which learns edge costs to provide global movement guidance via backward Dijkstra search. In experiments on high-density maps, SJRL significantly outperformed Causal-PIBT alone. The team also validated SJRL in a mixed-reality warehouse with 8 physical and 248 virtual robots, demonstrating practical feasibility. The paper is available on arXiv (2608.05588) and was submitted to the Robotics track.