Multi-Agent Path Finding (MAPF) is critical for warehouse automation. Classical solvers like CBS provide optimal guarantees but cannot handle dynamic environments. Fully learned approaches suffer from cold-start. We present CBS-Bootstrapped MAPPO: a three-phase curriculum using CBS as a decaying reward signal to bootstrap MARL. The trained policy surpasses CBS at every difficulty level, maintains 79–95% success under dynamic obstacles (vs 0–14% for CBS+dynamic), and accepts natural language commands via a local LLM (qwen:4b via Ollama) for zero-shot zone assignment across a 20×20 warehouse grid with 12 agents.
Kyaw Linn Khant (Mon,) studied this question.