CLE: An Integrated Framework of CNN, LSTM, and Enhanced A3C for Addressing Multi-Agent Pathfinding Challenges in Warehousing Systems

In the realms of production logistics and intelligent warehousing systems, Multi-Agent Path Finding (MAPF) is paramount, particularly in crafting optimal robot navigation paths while mitigating spatiotemporal conflicts. This paper delves into MAPF within a three-dimensional warehousing context, aiming to concurrently reduce the duration and length of paths. The author proposes the CLE framework, an innovative distributed MAPF model that melds the capabilities of CNN, LSTM and enhanced A3C with the CBS approach. This framework significantly boosts the efficiency of observational data processing by dividing map data into three distinct layers: a global static map layer, local path trajectory layer, and local robot coordinates layer. Within this structure, the global static map and robot coordinates are processed through a CNN network, capitalizing on its strength in spatial data analysis, while LSTM networks manage the local path trajectory, focusing on extracting historical trajectory patterns and forecasting future movements. Utilizing the A3C algorithm, initial strategies and paths are formulated, followed by the application of the CBS algorithm for conflict identification and pathway refinement. These paths are then re-fed into the A3C system for strategic adaptation. The framework operates on a cycle of ongoing environmental updates and performance assessments, effectively orchestrating conflict-free navigation for multiple robotic agents, thereby demonstrating its efficacy in dynamic and complex warehouse settings.