Abstract: To address issues in automatic loading systems for automated guided vehicle (AGV) under unfavorable warehouse environment conditions such as insufficient navigation accuracy, poor environmental adaptation and susceptibility to interferences from moving obstacles, a simultaneous localization and mapping (SLAM) navigation system was designed based on the fusion of multiple sensors including light detection and ranging (LiDAR) and camera and inertial measurement unit (IMU). On the multi-sensor fusion portion, a dynamic weight strategy which is adjustable in real-time dependent on indicators such as signal-to-noise ratio and time delay of sensors, and a multi-head attention mechanism which is capable of encoding LiDAR point clouds, camera images and IMU features separately and performing cross-modal aggregation were employed to achieve intelligent fusion of heterogeneous sensor data. On the SLAM portion, closed loop detection is achieved through local subgraph similarity and temporal continuity, and the global pose graph is updated in real-time through nonlinear optimization in local areas. Application tests of the navigation system showed that the positioning accuracy of the system reached ±1.2 cm under standard warehouse environment condition. The adaptability to dynamic environments where obstacle speed reached 1.2 m/s, and that to harsh working conditions including weak illumination, vibration, electromagnetic interference were superior to those of the single-modal algorithm of A-LOAM and the dual-modal algorithms of LIO-SAM, VINS-Fusion and ORB-SLAM3. The stability on the probability of uninterrupted operation for 24 hours and availability on normal operation time/total time× 100% of the system were 94% and 99.2% respectively. This technology can provide references for the improvement of industrial logistics automation.