A Deep Reinforcement Learning Algorithm for Warehousing Multi-AGV Path Planning

The problem of AGV path planning has become a key technical problem in the fields of cargo transportation and rapid sorting. Due to the continuous expansion of the scale of application scenarios, more AGV cooperation is required. Traditional planning models are difficult to coordinate path planning between multiple AGVs. Aiming at the warehouse layout scenario of the Kiva system, the article proposes an A3C algorithm combined with the Attention mechanism of the MAA3C algorithm. The advantage function and entropy are used for training to better solve the path planning problem of the unequal number of picking shelves in the storage multi-pick station, MAAC (Uniform), MADDPG, MADDPG+SAC, and COMA+SAC algorithms for comparison. The simulation results show that the MAA3C algorithm has a better convergence effect than other algorithms, the convergence time is 1/3 of other algorithms, and the average reward is increased by about 25%. Therefore, this algorithm can effectively solve the optimization problem of picking path planning under the "goods to people" mode, and improve the collaboration and scalability of multiple AGVs in the warehouse.