Machine Vision Technology and Its Application in Intelligent AS/RS Picking

As a core facility of modern logistics and warehousing, intelligent AS/RS warehouses improve warehouse operational efficiency through automated storage and picking. However, accuracy and flexibility in the picking process remain persistent industry challenges. Traditional manual picking is prone to errors and low efficiency, while early automated picking relies on fixed programming and struggles to adapt to scenarios with diverse SKUs. The introduction of vision recognition technology equips intelligent AS/RS picking with a perceptive capability, enabling intelligent identification, positioning and handling of goods, and has become a key technology to address the above pain points.

Ⅰ. Core Application Scenarios of Vision Recognition in Picking

1. Cargo Positioning and SKU Identification

Inside AS/RS racks, vision systems capture cargo images via cameras mounted on AGVs, robotic arms or fixed positions, and apply deep learning algorithms such as YOLOv8 and Faster R-CNN for target detection. The system can identify barcodes, QR codes and text labels, or match SKU information through appearance features including shape, color and texture, while calculating the three-dimensional coordinates of goods. For instance, vision systems in e-commerce logistics centers can quickly identify packages of various categories and guide AGVs to designated racks, saving the time cost of manual searching.

2. Pose Guidance for Robotic Arm Gripping

For robotic arm picking, 3D vision technologies including structured light, binocular vision and laser triangulation reconstruct three-dimensional models of goods to acquire their size, dimension and spatial posture such as tilt angle and stacking height. Based on such data, the system calculates optimal gripping points and motion paths to ensure precise clamping and avoid cargo damage or gripping failure. In manufacturing parts warehouses, 3D vision can recognize irregularly shaped metal components and guide robotic arms to complete accurate picking, adapting to production demands of small batches and multiple varieties.

3. Picking Quality Inspection

After picking, the vision system conducts secondary verification: confirming the correctness of target SKUs, packaging integrity and quantity accuracy. In pharmaceutical warehouses, for example, the system identifies batch numbers and validity periods to ensure picked medicines meet compliance standards. In fresh food warehouses, it can judge fruit ripeness and surface damage to enhance overall picking quality.

Ⅱ. Key Technical Support

1. Integration of 2D and 3D Vision

2D vision excels at recognizing planar features such as labels and colors, while 3D vision provides depth information. The combination achieves comprehensive cargo perception. In stacked cargo scenarios, 2D vision identifies commodity types and 3D vision determines stacking layers and picking sequence to prevent collisions of robotic arms.

2. Robust Algorithm Optimization

Aiming at complex warehouse environments such as unstable lighting, surface reflection and cargo occlusion, targeted optimization is required. Image enhancement technologies like histogram equalization are adopted to improve image quality under low light; attention mechanisms focus on visible areas to mitigate occlusion interference; transfer learning enables rapid adaptation to new SKUs and reduces model training costs.

3. Edge Real-time Processing

Vision processing modules are deployed on edge devices such as AGV controllers and local computing units of robotic arms to reduce data transmission delay and realize millisecond-level decision-making. AGVs can identify rack goods in real time during movement without relying on cloud computing, effectively improving picking efficiency.

Ⅲ. Application Value and Effect

The adoption of vision recognition technology greatly enhances the picking efficiency and accuracy of intelligent AS/RS warehouses. Industry data shows that automated picking systems equipped with vision recognition can raise picking efficiency by 3 to 5 times and lower error rates below 0.1%. Meanwhile, flexibility is significantly improved, being able to accommodate picking demands for thousands of SKUs and cutting the programming cost and cycle of conventional automated systems. For example, after introducing a vision-based picking system, a FMCG warehouse increased hourly picking volume from 100 pieces to 400 pieces and reduced labor costs by 60%.

Ⅳ. Challenges and Future Development Trends

Current vision recognition technology still faces limitations. Complex warehouse conditions such as fluctuating light and cargo occlusion may affect recognition precision; the relatively high cost of 3D vision equipment restricts large-scale popularization; frequent SKU updates require continuous iteration of algorithm models. In the future, the integration of large AI models and vision technology will endow systems with stronger semantic understanding to process natural language picking instructions. Multi-sensor fusion combining vision, RFID and infrared will further boost operational reliability. The advancement of low-cost 3D vision technology will also accelerate its deployment in more industrial scenarios.The application of vision recognition technology is reshaping the warehousing picking mode of intelligent AS/RS warehouses, realizing the upgrade from basic automation to comprehensive intelligence. Despite existing challenges, vision recognition will serve as an essential support for intelligent warehousing with technological progress, driving the high-efficiency development of the logistics industry.