The work aims to improve the efficiency of the "goods to people" picking system in logistics enterprises during actual production, avoid conflict deadlock between automatic guided vehicles (AGVs), and study the conflict free path planning and collaborative obstacle avoidance problem of large-scale multi AGVs. Firstly, A algorithm was improved considering the empty load, load situation, and path expansion cost of AGV, the cost function was adjusted dynamically and the path expansion method was optimized. Then, a conflict detection and avoidance algorithm was proposed, which scheduled path intersections that might generate local conflicts. Local conflict detection was achieved through reserved lock grids, and priority obstacle avoidance strategies were developed to solve local conflicts and deadlocks generated on AGV dynamic driving paths, to achieve global conflict free path planning. Multiple scenarios with different task volumes and AGV scales were simulated. The experimental results showed that the improved A algorithm considering conflict avoidance could effectively achieve dynamic path planning for multiple AGVs in scenarios with 100 tasks, 90 shelf units, and 7 picking stations. Compared to the traditional A algorithm, the average picking time was optimized by 52.61%. This method can achieve dynamic path planning for multiple AGVs in large-scale scenarios, effectively avoiding local dynamic conflicts while paying less turning costs. This method can provide new ideas and theoretical basis for relevant enterprises to achieve collaborative scheduling of multiple AGVs.