In recent years, the trend of aging agricultural labor has become increasingly severe, and difficulties in hiring and high labor costs have become bottlenecks hindering the development of the Tea industry. Manual picking of premium teas accounts for approximately 60% of the total labor involved in Tea Garden management. The shoots of high-quality teas are delicate, with varying positions, postures, and densities, making machine harvesting particularly challenging in unstructured environments with changing breezes and light. Therefore, studying intelligent tea-picking technology is of great significance for promoting the development of China's tea industry.

Researchers calibrating a tea-picking robot

1. Tea Bud Recognition Based on Image Processing

To achieve automated tea picking, the accurate recognition of tea buds is essential. In recent years, with the development and application of computer technology, the accurate recognition of tea buds based on image processing has become a hot topic of research.

1.1 Traditional Image Processing Algorithms Based on Color Space

Due to the obvious color differences between young and old leaves, as well as the tree trunk, color features can be used to extract regions containing young shoots from images. Therefore, early research on segmenting tea shoots mostly relied on color features. Traditional image processing algorithms based on color space primarily involve steps such as image preprocessing, color feature selection, and segmentation.

Original tea tree image

Identification of young shoots using the color factor R-B

Identification of young shoots using chromaticity information I

1.2 Recognition Methods Based on Traditional Machine Learning

To further address issues such as the influence of external environmental factors like old leaves, branches, and soil, as well as mutual occlusion and overlap of tea leaves under natural conditions, subsequent studies introduced machine learning methods. These methods identify and detect by extracting and combining various feature samples for training. Commonly used methods for recognizing young shoots include those based on color, texture, shape, and other features, combined with techniques such as K-means clustering, support vector machines, Bayesian discriminant methods, and cascade classifiers. Identification methods based on traditional machine vision still rely on image preprocessing and data transformation, and unreasonable preprocessing can significantly affect model accuracy.

Original tea shoot image (a) and cluster result from machine algorithm (b)

1.3 Recognition Methods Based on Deep Learning

Algorithms based on deep learning exhibit high precision in complex backgrounds, providing a foundation for the research and development of intelligent tea-picking equipment. These can be categorized into three types: classification algorithms, object detection algorithms, and semantic segmentation algorithms. Classification algorithms based on deep learning classify an image and determine whether it contains young shoots or identify the state of the young shoots, such as their opening condition and whether they are ready for picking. This method not only accurately identifies tea shoots but also distinguishes different states of young shoots, meeting the requirements for recognizing young shoots under natural lighting. However, methods based on deep learning rely on large datasets and have lower detection efficiency. Therefore, further research on detecting tea buds is needed, including increasing the number of bud images and developing faster, more accurate, and stable algorithms.

Structure diagram of the AlexNet model, which can effectively recognize the state of young tea shoots under natural lighting

2. End-Effector for Picking

The target of tea picking is the bud and leaf rather than fruit, so conventional end-effectors are not suitable. Researchers have developed new end-effectors specifically for tea shoots. For example, in 2025, a gripper-type end-effector was designed for picking tea shoots. By controlling this end-effector, tea garden picking can be achieved. Test results show that the missed pick rate for one bud and one leaf is 2.8%, and the intact pick rate is 91%; for one bud and two leaves, the missed pick rate is less than 3%, and the intact pick rate is 94%. Most existing tea-picking end-effectors adopt simple mechanical structures with little error compensation capability, making it difficult to ensure a high success rate and intactness of the shoots. To address this issue, an end-effector for premium tea picking based on negative pressure guidance has been designed. This end-effector uses negative pressure to guide the tea shoots from top to bottom, correcting their posture and spatial position. Test results indicate that the designed end-effector has tolerance performance, improving the pick success rate.

Model of a gripper-type tea shoot picker

Tea garden picking test

3. Intelligent Control System

The primary functions of an intelligent control system include control of the driving system and the picking mechanism. Japan has made some research achievements in the control of smart machinery driving systems in tea gardens. For example, Matsuda Corporation has developed an “unmanned tea-picking machine” that moves and harvests tea using artificial intelligence (AI) and sensors, and it is already available for sale. Regarding the control of the picking mechanism, a ride-on intelligent tea-picking machine based on machine vision has been designed for traditional reciprocating cutting harvesters. It proposes automatic recognition of tea buds and automatic leveling control of the cutting blade, addressing the drawbacks of existing tea-picking machines that cut old leaves and young shoots indiscriminately. Currently, all control systems for end-effectors control a single end-effector, leading to low picking efficiency. Future research should focus on multiple end-effectors and collaborative control systems for multiple robotic arms.

Structural diagram of a ride-on intelligent tea-picking machine based on machine vision (1. Cutter level control unit; 2. Cutter height control unit; 3. Cab; 4. Camera; 5. Arc-shaped cutter; 6. Traveling mechanism control unit)

Cutter blade line and cross-section of tea row

4. Existing Challenges and Prospects

The research on harvesting machines for high-quality teas is still in its infancy, with prototypes undergoing testing. There are still some challenges in practical applications, such as the lack of close integration between agricultural machinery and farming practices, significant impact of lighting on bud recognition, difficulty in segmenting images with backgrounds similar to young shoots, and poor recognition results due to occlusion and overlap of leaves. Compared to traditional machine learning, current methods based on deep learning for bud and leaf recognition show promising prospects, but they require large amounts of labeled samples for training, and upgrading hardware systems becomes an issue as network complexity increases. With the rapid development of machine vision and artificial intelligence technologies, a solid foundation has been laid for the development of smart tea-picking machines. Future smart tea-picking machines will likely follow these trends:

4.1 Increasing Sample Data and Developing Recognition Models to Improve Algorithm Performance

The current challenges in recognizing and locating tea buds lie in the diversity of tea varieties and growth environments, strategies for recognizing young shoots under occlusion and overlap, dynamic