Next Article in Journal
A Qualitative Analysis of Australian Perspectives Regarding the Cost of Groceries Using the Online Social Media Platform, Reddit
Previous Article in Journal
Quantitative Determination of Nitrogen Fixed by Soybean and Its Uptake by Winter Wheat as Aftercrops Within Sustainable Agricultural Systems
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Optimizing Waste Sorting for Sustainability: An AI-Powered Robotic Solution for Beverage Container Recycling

1
Graduate School of Environment and Energy Engineering, Waseda University, Shinjuku-ku, Tokyo 162-0041, Japan
2
EII, Inc., Chiyoda-ku, Tokyo 101-0054, Japan
3
Environmental Research Institute, Waseda University, Shinjuku-ku, Tokyo 162-0041, Japan
*
Author to whom correspondence should be addressed.
Sustainability 2024, 16(23), 10155; https://doi.org/10.3390/su162310155
Submission received: 29 October 2024 / Revised: 18 November 2024 / Accepted: 18 November 2024 / Published: 21 November 2024
(This article belongs to the Section Waste and Recycling)

Abstract

:
With Japan facing workforce shortages and the need for enhanced recycling systems due to an aging population and increasing environmental challenges, automation in recycling facilities has become a key component for advancing sustainability goals. This study presents the development of an automated sorting robot to replace manual processes in beverage container recycling, aiming to address environmental, social, and economic sustainability by optimizing resource efficiency and reducing labor demands. Using artificial intelligence (AI) for image recognition and high-speed suction-based grippers, the robot effectively sorts various container types, including PET bottles and clear and colored glass bottles, demonstrating a pathway toward more sustainable waste management practices. The findings indicate that stabilizing items on the sorting line may enhance acquisition success, although clear container detection remains an AI challenge. This research supports the United Nation’s 2030 Agenda for Sustainable Development by advancing recycling technology to improve waste processing efficiency, thus contributing to reduced pollution, resource conservation, and a sustainable recycling infrastructure. Further development of gripper designs to handle deformed or liquid-containing containers is required to enhance the system’s overall sustainability impact in the recycling sector.

1. Introduction

Due to the declining birthrate and aging population, Japan’s labor shortages and difficulties in maintaining and passing on technical skills in the waste treatment and resource recycling fields become increasingly urgent. Meanwhile, there is an increase in environmental awareness, the need to reduce infectious risks, and the promotion of recycling [1]. Such situations lead to the increasing expectation of adopting automation technologies such as sorting robots in the recycling industry. A literature review of Japanese academic papers using the keywords “robot”, “waste”, “sorting”, or “recycling” is conducted and shows that automatic sorting robots for waste have emerged since the 1990s and have been actively researched, with studies on this topic being published in large numbers since 2011 [2].
Automatic sorting systems for waste materials mainly use image recognition artificial intelligence (AI) and robot arms to sort objects of mixed types and properties into recyclable materials by category. However, it makes waste sorting with robots difficult due to the large number of features and the heterogeneous nature of waste materials. Take PET bottles as an example; even though both the production and recycling sites contain a sorting process, the sorting conditions and target are quite different. A production line of PET bottles is designed to use specific materials (PET) to produce bottles with the same features such as shape and volume; the sorting process is aimed at removing the objects that do not fulfill the favorable requirement of the line. However, the sorting process of a recycling line is required to separate PET bottles from the mix of materials of beverage containers such as cans, glass, and PET bottles. The waste comprises products from different makers, making the features impossible to be uniformed. Even PET bottles from a common production line cannot be collected with the same feature, since the consumers and collectors may change the features by cutting, squishing, twisting, or compressing the bottles due to their habits or the requirement of collecting efficiency.
Many existing waste sorting robots use image recognition AI to verify the diverse properties of waste. Technicians train the image recognition AI with a favorable image of targets as teaching data, which allows the AI to identify the objects to be acquired. However, the inaccessible discrimination criteria of the image recognition AI make it hard to improve the training efficiency. Also, due to the problem of waste stacking on the sorting line, multiple objects coming into contact with each other, and objects with similar characteristics being mixed, the accuracy of image recognition AI in the recycling field faces technical barriers.
In this study, we aim to develop an automated sorting robot system for beverage containers and to identify the challenges of robot sorting with existing technologies. Unlike the reverse vending machine targeting the collection stage and the PET sorting machine to improve the purity of waste-sourced PET flakes, this study aims to automate the pretreatment stage of recycling by sorting the beverage containers by materials and colors at once [3,4]. By analyzing the challenges of sorting robots on a full scale with actual waste collected from recycling facilities, we aim to inform automatic waste sorting system developers about the key factors that must be considered during robot development, particularly regarding the working environment, the characteristics of targeted waste, and robotic abilities.
The remainder of this study is organized as follows. Section 2 presents literature reviews to clarify how we developed the composition of our automated sorting system and provide theoretical support for waste identification algorithms based on previous studies. Section 3 describes the materials and methods used in this study, such as the composition and systematic algorithm of our robot, the descriptions of our experiment settings, and the methods to test the efficiency of our robot’s final pick-up point decision and color sorting. The composition and creation method of our image recognition AI’s training data and experimental setups are also described here. The results and discussions of the experiments are presented in Section 4. In this study, we analyze the effectiveness and challenges of our approach by detecting the material type of beverage containers, distinguishing the color of glass bottles (brown, clear, and others), and evaluating the acquisition process of targeted containers through an actual equipment scale experiment. Finally, the conclusion and recommendations are presented in Section 5.

2. Literature Reviews

2.1. Current State of Waste Classification Systems

Among the existing waste recycling methods, manual sorting is considered unsafe, causing health issues, and unproductive [5]. On the other hand, traditional automated classification methods based on wind selection, screening, and flotation are inefficient and often inaccurate [6]. Our review of the existing literature found that there are limited available papers related to robotic waste sorting that is operating on a fully industrial scale [7]. The literature discussed that the challenges of image recognition AI in waste sorting areas are due to the lack of availability and selection of applicable data, poor reproducibility, and less evidence of applications in real solid waste [8].
Image recognition based on deep learning was widely applied to the classification systems consisting of low implementation cost and the ability to sort dark-colored targets, which cannot be classified with an NIR-based sorting system [9]. There is a study on improving the waste classification process with image recognition that provides a data augmentation method to resolve the challenge of insufficient datasets [10]. Another study introduced deep learning methods to test whether image recognition could classify varied categories of waste [11]. The possibility of applying those methods to portable devices and smart cities’ recycling decisions is also evident from the existing research [12,13]. However, further studies are required with complex situations to achieve automated classification in recycling facilities.
While research on classification robots consists of the challenge of recognition accuracy, they are more focused on applying techniques to improve an automated sorting result. Meanwhile, the on-site classification robot for construction waste needs to be created and tested for its ability in patrolling, localization, and sorting [14]. For example, a study testing a suction-type object manipulator and AI using an R-CNN mask showed that cans, paper nylon, and PET bottles can be successfully identified and picked up [15]. Another study improved this method by applying augmented reality in the construction waste sorting robot system so that the recognition accuracy can be increased by adding the operator’s corrections to the automated classification system [16]. It can be concluded that compared to the waste recognition methods, the stability of the robot arm dominates the performance of the sorting system [17].
It is now clear that AI-based automated waste sorting with robotic arms is superior to other methods. However, there are variations in terms of the sorting targets and sorting facilities in each of the studies. As this study’s interest is to increase the resource recycling rate, particularly from the beverage container waste in Japan, it is therefore necessary to recompose a suitable waste sorting robot system particularly designed for this purpose and can be tested in a full-scale operation.

2.2. Gripper and Sensor Mounted on Automatic Waste Sorting Robot

One challenge of developing a waste sorting robot is to select a gripper that matches the characteristics of the object. The categories of grippers used in waste sorting robots can be classified as follows: two-finger grippers, pneumatic suction grippers, broom grippers, and composite grippers. The gripper currently most widely used in automatic waste sorting robots is the two-finger gripper with easy control, and the second most common gripper is the suction type using a pneumatic suction cup with high speed [2]. Grippers for sorting various types of recyclable waste were developed and evaluated, with experiments showing that both gripping and suction types are equally effective when targeting beverage containers [18].
To automate the waste sorting processes, it is necessary to introduce appropriate sensors to determine the material properties and locate the object. A red, green, blue, depth (RGB-D) camera has been proven to be effective for acquiring three-dimensional (3D) information suitable for sorting, including height direction [14]. In contrast, an infrared sensor and a red, green, blue (RGB) camera are effective for acquiring horizontal coordinate information only [19]. The RGB-D and RGB cameras are also used for material identification by image recognition. However, a sound sensor, which supplements image information, and an infrared hyperspectral camera can sort out the type of plastic but require specific conditions of use to determine the object’s material [20,21]. Moreover, the construction cost of the system is high.

2.3. Algorithms for Waste Identification

Recently, image recognition AI has been employed for waste detection and location by automated sorting systems, but the computational system of the AI and the identification of a mixture of various types of sorted objects have been recognized as challenges [7]. A 15-layer convolutional neural network image classification system was developed to effectively reduce the load on hardware and obtained experimental results showing that a 9:1 ratio of training data to test data provides high accuracy [12]. Referring to the detached links between research and real-life practice caused by over-simplified working conditions in a computer vision model, the training data of this study is collected from real beverage container waste supplied from the recycling facilities [22]. A deep learning-based system was successfully implemented to detect and sort glass bottles by color. However, the potential for workforce replacement is limited, as the system focuses solely on glass bottles, requiring the manual removal of plastic bottles and cans before the robot can be used [23].
Typically, for an image recognition AI model to acquire the ability to perform a detection task with the desired accuracy, it requires a large amount of training data from manual labeling, which costs most of the labor force. A data augmentation method by randomly deleting parts of the image is developed [24]. Meanwhile, a significant improvement in detection performance over the baseline with only 80 original images by expanding the dataset is obtained using a rotation table, an RGB-D camera, and a region adaptation method [25,26]. Referring to these references, we expanded the training dataset to train the image recognition AI in this study.

3. Materials and Methods

3.1. Configuration and Control of the Automatic Beverage Container Sorting Robot

Since this study targets beverage containers collected from emission sites and intermediate treatment facilities that have not been pretreated by manual sorting, a suction-type gripper was selected for lightweight materials and for a high sorting speed, and an RGB camera was mounted on an automatic sorting system mainly for data processing on a flat surface. The sorting target was set at five types of containers: PET bottles, cans, clear glass, brown glass, and other glass. Aluminum and steel cans were excluded from the robot’s sorting target because they can be sorted by machine later. Table 1 shows an overview of the difference between this study and existing automatic beverage container sorting robots [27,28,29] from the perspective of sorting targets, recognition method, and ideal placement location.
Figure 1 shows an overview of our automatic beverage container sorting system. As an introduction to the functions of each unit, the computer performs AI-based computational tasks and controls the camera and relays in the following order: the identification of the object and determination of its material, creation of a segmentation mask, calculation of the acquisition point, and color sorting of the object determined to be a glass bottle. The relay facilitates communication between the robot, camera, and encoder by sending signals when events occur. The robot controller receives signals from the encoder, relays them, and controls the robot to pick up the bottles on the conveyor belt.
Figure 2 shows the actual condition of the automatic sorting system for beverage containers developed for this study, and Table 2 shows the detailed specifications of the components in the systems.
The higher the frequency per second (Fps), the more sensitive the image recognition AI system becomes and the more computationally intensive it becomes. Increased computational load also increases the load on the hardware; the RGB camera used in this system was set to a maximum resolution of 1920 × 1080 pixels and 30 fps based on experience, considering the balance between the performance and cost required for the robot. Considering the lighting conditions in the laboratory and the demand for image recognition, a fixed exposure of 390 was set based on the experience gained during the development of the AI. The camera is mounted on the robot’s frame and has a bird’s-eye view of the conveyor belt. To illustrate the camera’s position and the vision system’s field of view, a sample image taken from the setup camera is shown in Figure 3.
A flat conveyor belt is used with a constant speed of 156 mm/s and equipped with an encoder to track the distance traveled on the conveyor belt.
The robot manipulator is a 4-axis parallel robot (3 arm axes and 1 rotation axis) with suction. We have developed a software control system with the pytorch framework consisting of an image recognition AI and a robot control system, which is provided by the robot manufacturer to execute the robot’s movements, and a hardware system consisting of a relay, an encoder, a camera, and a robot manipulator, which can share information and calibrate and align the movement intervals to ensure that bottles moving on the conveyor belt can be sorted in a timely manner [30]. To enhance the success rate of grip suction, the robot is designed to pause at the acquisition point after making contact with the object to be sorted. Figure 4 represents the systematic algorithm of the software control system.

3.2. Detection of Objects and Determination of Material (PET, Bottles, Cans)

For object detection, we trained the YoloV7 image recognition AI model, which can perform the entire process from image input to generating the final identification result in a single step [31]. The COCO dataset, a large dataset, was used for pre-training, and the results were transferred to our image recognition AI model to reduce the training time until model convergence [32]. The dataset used for training was collected from videos taken in our lab, converted to images with the labeling platform Supervisely, and then processed into training data by manually marking the objects we wanted the AI to learn [33].
As a result, we achieved 0.9967 Mean Average Precision (mAP) with 0.5 Intersection over Union (IOU) on the validation dataset consisting of 409 images. The confusion matrix for each class is shown in Figure 5, with the vertical axis representing AI’s predictions for plastic bottle (PET), glass bottle (GLASS), can (CAN), false negative, or none of the above (background FN) and the horizontal axis representing the predictions for PET, GLASS, CAN, false positive, and the probability of wrong recognition of an object that is not an object (background FP) on the horizontal axis, and the numbers indicate the probabilities. To improve the decision speed of image identification AI, we used TensorRT, a deep learning inference engine designed by NVIDIA to accelerate inference speed, which is effective for real-time applications or when low latency is required and possesses advantages such as model optimization, efficient memory usage, and improved accuracy. The use of TensorRT for both object detection and segmentation models resulted in a reduction in inference time to an average of 50 ms, speeding up the recognition of objects from 5 objects/s to 20 objects/s. The image recognition AI training result is shown in Table 3.
We tried two methods to separate the beverage containers and the objects to be sorted from the background. First, we tried the Gaussian background motion model to extract the foreground region from the background for a conveyor belt that is always in motion [34]. Figure 6 shows a sample image of the results. It can be observed that this method treats a part of the background as a segmentation mask. The segmentation mask results from segmenting the image into objects. Therefore, it can be concluded that this method can remove background information from the image from the computational range of AI. Unfortunately, there are parts of the image data that image recognition AI recognizes as a beverage container. Second, we employed a pre-trained Segment Anything Meta AI (SAM) to generate segmentation masks and segment individual bottles for pick point estimation and color classification algorithms [35]. The results are shown in Figure 7. It can be observed that yellow-colored segmentation masks are likely to present a similar shape to beverage containers with less noise from the background. Comparing the segmentation masks created by the two methods, the SAM method provides a higher accuracy in the calculation for the next step.

3.3. Preliminary Validation to Identify Pick Points and Determine Color Sorting Methods for Glass Bottles

3.3.1. Pick Point Identification

For a robot equipped with a suction gripper to successfully pick up a bottle, it is necessary to estimate the pick point on the bottle and send its coordinates to the robot. To calculate the appropriate pick points, the longest section method and the principal component analysis (PCA) method were attempted, and both methods yielded a pick rate of 80%. The longest section method is an algorithm that creates a segmentation mask with a set of parallel cut lines in the same direction, extracts the longest cut line that maximizes the cut plane of the segmentation mask, and uses its center as the pick point. Figure 8 represents the algorithm for identifying the longest cross-section.
However, the longest section method relies on the accuracy of the segmentation mask. If the segmentation mask can reproduce only part of the object’s shape, the pick point determined by the SAM may deviate from the object’s center of gravity. To illustrate this scenario, the leftmost bottle in Figure 7 shows a case in which the SAM missed part of the segmentation mask on the label.
Principal component analysis (PCA) is a method that standardizes image information as numerical values, computes a covariance matrix to determine the principal axes of an object in terms of major direction and importance, and then locates the center of gravity along the principal axes. However, like the longest section method, this method also has the disadvantage that the center of gravity varies greatly depending on the accuracy of the segmentation mask.

3.3.2. Color Sorting

In order to perform color sorting of glass bottles, it is necessary to know the mask that can accurately indicate the location of the detected bottles and to analyze the colors in the mask. Therefore, the segmentation mask generated in the previous step of the pipeline is used to analyze the colors of the glass bottles within the mask area.
To segment the colors in each mask, global filtering was applied to the thresholding process using color space, and after converting the images acquired by the camera into Hue, Saturation, Lightness (HSL) and Hue, Saturation, Value (HSV) color spaces, the threshold values were determined based on statistics derived from the collected data. We attempted to determine the threshold values based on the statistics derived from the collected data. However, the HSL and HSV color space values obtained due to the lighting conditions in the laboratory could not be matched, and reflected light from objects other than the object to be identified was also reflected in the color space, making color discrimination of the object using this method difficult. Figure 9 shows an image converted to HSL color space. The object on the left of the screen cannot be distinguished because of the uneven illumination on the conveyor belt.
We also attempted direct classification of glass bottles based on their image information. The results of training the bottle color-coded convolutional neural network (CNN) showed that the model could not generalize changes in bottle appearance and lighting conditions, so color-coded sorting using this method has an error rate of about 50%, which is quite high, and almost half of the objects are misidentified. The results of the color algorithm and the color space threshold regions are shown in Figure 10.

3.4. The Pick Point Identification and Glass Bottle Color Sorting Method Finally Adopted

3.4.1. Image Moment Calculation Method

In order to cope with the shift in the cut line and center axis of the object due to the missing segmentation mask, the image moment calculation method was employed to estimate the moments of the image from the segmentation mask and to find the coordinates of the center of gravity from those moments. The formula for calculating the coordinates of the center of gravity using this method is shown by Equations (1) and (2).
G x =     L · x L
G y =   L · y L
G represents the center of gravity, x and y represent coordinates; and L represents luminance value.
This calculation method can accurately reflect the center of gravity of the identified part because it is a weighted average of the positional information of all pixels in the image, thus mitigating the effect on the calculation results caused by missing or irregular shapes reflected in the segmentation mask. As a result, the pick rate of the object was improved from 80% to 90% by changing the calculation of the center of gravity to this method, and this method was finally selected as the center of gravity calculation method to be installed in the robot.

3.4.2. Color Sorting Method Integrating SAM and CNN

Since the color sorting of glass bottles using only the CNN method could hardly handle brown bottles, we tried to improve the accuracy of the image information computed by the image recognition AI by integrating the SAM method, which can remove background information not needed for color sorting and is already used to identify pick points, into the color sorting CNN. Using the segmentation masks obtained by the SAM method, we achieved a pick success rate of more than 90% for sorting brown bottles, which CNN could not handle alone. The method of integrating SAM and CNN was selected as the method to be used in the robot for color sorting of glass bottles. Algorithm 1 represents the pseudocode of glass bottles detection and color classification.
Algorithm 1: Glass Bottle Detection and Color Classification
1: Input: Image with glass bottle
2: Output: Predicted color label for the bottle (e.g., Transparent, Brown, Other)
3:
4: Step 1: Detect Glass Bottle in Image
5: Load detection model (e.g., YOLO, Faster R-CNN)
6: detections ← model.detect(image)
7: bottle_box ← find_bottle_bounding_box(detections)
8:
9: Step 2: Create Bounding Box
10: Extract bounding box coordinates: (xmin, ymin, xmax, ymax)
11: cropped_image ← crop_image(image, bottle_box)
12:
13: Step 3: Segment Bottle Using SAM within Bounding Box
14: Load SAM segmentation model
15: mask ← sam_model.apply_segmentation(cropped_image)
16: segmented_bottle ← apply_mask(cropped_image, mask)
17: Save segmented_bottle as “segmented_bottle.png”
18:
19: Step 4: Preprocess Segmented Image for Color Classification CNN
20: resized_image ← resize(segmented_bottle, target_size = (224, 224))
21: normalized_image ← normalize(resized_image)
22:
23: Step 5: Classify Bottle Color with CNN
24: Load color classification CNN model
25: color_class ← color_cnn.predict(normalized_image)
26:
27: Step 6: Output Classification Result
28: if color_class = 0 then
29: Label as “Transparent”
30: else if color_class = 1 then
31: Label as “Brown”
32: else
33: Label as “Other”
34: end if
35: Print classification result

3.5. Composition and Creation Methods of Training Data

Image data for training this robot’s image recognition AI was collected by a camera installed in the lab and labeled on a labeling platform for labeling and data management. Approximately 1200 images were collected and labeled with a resolution of 1280 × 720 pixels. Then, the entire dataset was split into a 533:63 ratio for training and validation, and the image recognition AI training method of pre-training and data augmentation resulted in a 99% identification accuracy (training effects usually obtained from tens of thousands of teacher data [12]) as shown in Table 3. The details of the objects to be identified in the image are 8662 plastic bottles, 6369 glass bottles, and 2152 cans.
In order to reduce the amount of labor required for the labeling process, we also employed the semi-automatic labeling provided by Supervisely and the modified labeling based on the judgment results of the image recognition AI model trained in the pre-training process. Figure 10 shows the results of judging the image data acquired for the generation of training data with the pre-trained model and the manually corrected data. Analysis of the samples in Figure 11 shows that the data generated by the pre-training model was mislabeled (30%), overlooked (18%), and miss-sized (24%), while about 52% of the labels did not require any correction.
Semi-automatic labeling provided by Supervisely is a tool that can automatically label objects on all frames of video data until the marked objects are no longer visible on the video screen by manually labeling all objects that appear on a single screen in the video data. Of the labels generated from semi-automatic labeling, about 30–40% of the labels have misalignment and size problems. Therefore, a manual label correction was necessary before the final training dataset could be created to ensure the effectiveness of image recognition AI training.

3.6. Performance Evaluation Experimental Setup

  • To test both the image recognition AI and the entire robot system’s performance and to figure out the challenges within the sorting process, a real-scale robot performance experiment is designed as follows.
  • Only the placement of beverage containers is allowed to be performed manually while the whole robot system is not working.
  • After the placement, operators start the robot sorting and take videos both at the robot site and the image recognition AI site, which represents the identification result on the screen. The sorting result is also collected with the progress of the experiment. The data are collected by counts of acquisition actions, positive solutions, acquisition misses, identification errors of AI, missorting, and the time of experiment by seconds.
  • All experiments are conducted with constant room light conditions in the laboratory, a conveyor belt speed of 156 mm/s, and robot settings.
  • The classification of robot sorting is designed as sorting PET bottles, cans, clear glass, brown glass, and other glass simultaneously.
  • Experimental results are analyzed through the data collected during the experiment, and the counting results from video data.
  • The image recognition AI is tested by counting the miss recognitions, and the properties of the mistakes are analyzed.
  • The robot system is tested by counting the rate of acquisition operation, successful picking, correct sorting, and picking misses. The properties of picking misses are analyzed by calculating the counts and percentage of specific properties among the total number of missed acquisitions.
  • A total number of 495 beverage containers are sorted as inputs.
  • The acquisition operation is set to pick up an object with an interval of 180 to 260 ms from the previous operation. Theoretically, 180 ms is the shortest pick interval, but to protect the robot, the interval is set to 260 ms.
  • To avoid malfunctions due to collisions between the robot and the conveyor belt, the pneumatic suction cup attached to the robot is set to descend to about 4 cm above the surface of the conveyor belt after reaching the pick point.
  • The robot is evaluated by calculating the percentage of each sorting and image recognition result. The sorting efficiency is calculated by dividing the count of acquisitions by experimental time in seconds.

4. Results and Discussions

4.1. The Results of the Experiments

The image identification AI model trained in 2.5 was validated in an actual experiment, which showed that 6% of the beverage containers were incorrectly identified. Table 4 shows the results of the AI recognition.
Table 5 shows the results of an automatic sorting experiment for beverage containers on a real machine scale. The acquisition operation is the operation in which the robot control system removes the object from the conveyor belt after the object is identified by the AI. Acquisition and identification errors mean that both types of errors occur simultaneously. The robot is theoretically capable of sorting at a speed of 45 picks/min, but in experiments, sorting speeds ranging from 16 picks/min to 45 picks/min were obtained.
The automatic sorting robot in this experiment made a total of 173 acquisition errors, of which 16% were due to the shape of the object, 25% to robot control problems, 15% to conveyor belt transportation, and 45% to the image recognition AI not sending the picking signal to the robot.

4.2. Discussion of the AI Experiment Results

Based on the labeling results of the pre-trained model for the beverage containers, about half of the labeling results do not require correction, which would reduce the effort required for frame-by-frame labeling. However, error correction in the pre-training model requires manually changing the wrong part of each frame to the correct label, which cannot be performed in batch processing. On the other hand, semi-automatic labeling by video can generate new labeling results after correcting the frame in which the error occurred. However, there is no guarantee that the new label is correct, so it is necessary to repeat the manual correction and checking process until the end. Continued research is needed to understand the time and effort spent modifying automatic and semi-automatic labeling methods and determine their efficiency.
The experimental results of AI recognition show that the image recognition system developed in this study has a 99% correct rate of identification for beverage containers with labels, which is consistent with the theoretical AI training results. However, about one-fourth of plastic bottles and glass bottles are mistakenly recognized in unlabeled beverage containers, suggesting the need to continue developing a discrimination algorithm for unlabeled plastic bottles and clear bottles, which are similar in color and shape.
The results of an automatic sorting experiment on actual-scale beverage containers show that when the automatic sorting robot system developed in this study was used to sort 500 mL beverage containers, the maximum sorting speed was 45 picks/min (2.4 m3/H), which is slower than the 20 m3/H of the existing product Petris [27]. However, the method seems to be effective in reducing the labor required for manual sorting by successfully sorting PET bottles, cans, clear bottles, brown bottles, and green bottles while maintaining a 93% correct rate for the items sorted by this method.
An analysis of the acquisition errors identified in the actual-scale experiments showed that the problem of the image recognition AI not sending acquisition signals, which accounted for 45% of acquisition errors, occurred when simultaneous judgments were required for multiple objects, as shown in Figure 12. To solve this problem, it would be effective to improve the AI’s ability to discriminate simultaneously or to develop a quantitative supply method that does not allow the image recognition AI to judge multiple objects simultaneously.

4.3. Discussion of the Results of the Robot System Experiment

Figure 13 shows examples of the errors caused by robot control problems, which account for 25% of acquisition errors. “Misalignment” represents examples of grasping errors, accounting for about 13%, which occurs when the robot moves from the previous drop point to the next pick point directly, which causes the suction pad to hit the targeted beverage container before the acquisition starts making the picking point mismatch from the calculated center of gravity. “Overlap” represents the acquisition that does not happen since the robot is considered to have insufficient time to operate the task, accounting for about 3%. “Miss out” represents the robot’s success in picking up the target but failure to hold it and finish the sorting because the suction could not keep a seal at the surface of the target, accounting for about 3%. “Release/Collision” errors represent the robot succeeding in acquiring the object but failing to move and release it solely, collision with other targets on the picking list, and changes in the position of the target making all following acquisitions fail, accounting for 6% of the errors. Adjusting the parameters of the controller could possibly resolve the misalignment errors. However, adjusting the output of the suction machine and adjusting the density of the objects flowing in the sorting line may be necessary to provide enough time and space between the acquisition tasks and resolve the errors of overlap, miss out, and release/collision.
The object shape problem, which accounts for 16% of the acquisition errors in Figure 14, is mainly due to the fact that the acquisition point determined by the image recognition AI is lower than the set drop point of the pneumatic suction cup due to a dent. Thus, the gripper cannot adsorb the object due to the presence of many patterns on the bottle’s surface and the acquisition point determined by the SAM method. To solve this problem, it may be necessary to adjust the pattern and label in cooperation with the bottle manufacturer or to change the gripper.
Figure 15 shows errors caused by conveyor belt transport, which accounts for 15% of the total number of errors. The contact surface between the beverage container and the belt conveyor is usually circular and prone to rolling due to vibration. However, if the position shifts after the point at which the image recognition AI finishes identifying the pick point, an empty swing occurs where the gripper cannot acquire the object. To solve this problem, it is necessary to devise a method for the surface of the conveyor belt to prevent the target container from rolling on the conveyor belt.

5. Conclusions and Further Study Recommendations

5.1. Conclusions

This study contributes to the automation of the beverage container sorting process by proving the potential of sorting the PET, can, glass material, and colors of glass bottles (clear, brown, and green) with image recognition AI at once with favorable accuracy. Based on the experimental results, the challenge that image recognition AI faces in the beverage container waste sorting process is the classification between targets that have similar features, such as clear glass bottles and PET bottles with no labels. Compared to the accuracy of image recognition AI’s classification, the ability of grippers that perform the acquisition action for targets with irregular shapes and surfaces, and the transportation system’s stability that transports sorting targets precisely to acquisition points, limited the performance of the automated sorting robot system. Looking ahead, we anticipate further improvements in the waste sorting algorithm to adapt to changes in the work environment, such as the handling of non-beverage containers mixed in the sorting line, thereby enhancing the future potential of the developed automated sorting robot system.

5.2. Further Study Recommendations

Since sorting lines in recycling facilities are usually crowded and many foreign objects such as cardboard and paper waste are present on the conveyor belts, we expect to develop a module to detect foreign objects on the conveyor belts. In addition, we aim to extend the function to pick up dented bottles, which the current system could be better at. In this experiment, a suction gripper was mounted on the robot. However, as a next step, we plan to try flexible grippers such as a grasping gripper and adapt the pick point algorithm according to the type of gripper and the method of operation. Since grasping grippers tend to be the slower type of gripper for the robot to pick and drop objects, we will work on adjusting the planning and operating algorithm so that the system can maintain the necessary speed to perform sorting tasks at the recycling facility. In this way, the system will be able to pick up the various shapes of beverage containers typically found in recycling facilities.
Based on the findings of this study, the following issues and needs are to be addressed in future studies in the information processing field, which is directly related to the development of automatic sorting robots for beverage containers. First, how the mechanical parts of robots should respond to changes in position after an image recognition AI identification should be addressed. Second, there is a need to improve the accuracy of clear bottles and PET bottle classification using image recognition AI. Third, the improvement of identification of containers with leftover drinks is needed. Fourth, there is a need to improve the adjustability of pick points according to the degree of concavity of the object.

Author Contributions

Conceptualization, T.C. and H.H.; methodology, T.C. and H.H.; software, H.H.; validation, T.C., H.H., D.K. and H.O.; formal analysis, T.C. and D.K.; investigation, T.C., D.K. and H.H.; resources, T.C.; data curation, T.C. and D.K.; writing—original draft preparation, T.C., D.K., H.H. and A.H.P.; writing—review and editing, A.H.P. and H.O.; visualization, H.H. and T.C.; supervision, H.O.; project administration, H.O.; funding acquisition, H.O. and A.H.P. All authors have read and agreed to the published version of the manuscript.

Funding

This research is funded by the Japan Society for the Promotion of Science (JSPS) Kakenhi grant project number JP23K11548 titled “Improving the social acceptance of smart technologies: A people-centric approach for sustainable smart waste management and smart mobility technologies”.

Data Availability Statement

The data presented in this study are available on reasonable request from the first author.

Acknowledgments

This study reports part of the results of the project proposal program “Development of a sorting robot for non-contact and automated waste disposal and recycling, and scenario building for social implementation” by Tokyo Metropolitan University researchers. We would like to express our gratitude to all the people involved.

Conflicts of Interest

Hao Hu is employed at Environmental Intelligence and Innovation (EII, Inc.). The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

References

  1. Onoda, H. Prospects for Contactless and Automatic Waste Collection. Mater. Cycles Waste Manag. Res. 2021, 32, 155–162. [Google Scholar] [CrossRef]
  2. Kiyokawa, T.; Takamatsu, J.; Koyanaka, S. Challenges for Future Robotic Sorters of Mixed Industrial Waste: A Survey. IEEE Trans. Autom. Sci. Eng. 2022, 21, 1023–1040. [Google Scholar] [CrossRef]
  3. Kim, D.; Lee, S.; Park, M.; Lee, K.; Kim, D.-Y. Designing of reverse vending machine to improve its sorting efficiency for recyclable materials for its application in convenience stores. J. Air Waste Manag. Assoc. 2021, 71, 1312–1318. [Google Scholar] [CrossRef]
  4. Dong, L.; Dong, C.; Zhi, W.; Li, W.; Gu, B.; Yang, T. Combining the Fine Physical Purification Process with Photoelectric Sorting to Recycle PET Plastics from Waste Beverage Containers. ACS Sustain. Chem. Eng. 2024, 12, 11377–11384. [Google Scholar] [CrossRef]
  5. Madsen, A.M.; Raulf, M.; Duquenne, P.; Graff, P.; Cyprowski, M.; Beswick, A.; Laitinen, S.; Rasmussen, P.U.; Hinker, M.; Kolk, A.; et al. Review of biological risks associated with the collection of municipal wastes. Sci. Total. Environ. 2021, 791, 148287. [Google Scholar] [CrossRef] [PubMed]
  6. Xiao, W.; Yang, J.; Fang, H.; Zhuang, J.; Ku, Y. Classifying construction and demolition waste by combining spatial and spectral features. Proc. Inst. Civ. Eng. Waste Resour. Manag. 2020, 173, 79–90. [Google Scholar] [CrossRef]
  7. Satav, A.G.; Kubade, S.; Amrutkar, C.; Arya, G.; Pawar, A. A state-of-the-art review on robotics in waste sorting: Scope and challenges. Int. J. Interact. Des. Manuf. (IJIDeM) 2023, 17, 2789–2806. [Google Scholar] [CrossRef]
  8. Ihsanullah, I.; Alam, G.; Jamal, A.; Shaik, F. Recent advances in applications of artificial intelligence in solid waste management: A review. Chemosphere 2022, 309, 136631. [Google Scholar] [CrossRef]
  9. Ji, T.; Fang, H.; Zhang, R.; Yang, J.; Fan, L.; Li, J. Automatic sorting of low-value recyclable waste: A comparative experimental study. Clean Technol. Environ. Policy 2022, 25, 949–961. [Google Scholar] [CrossRef]
  10. Mao, W.-L.; Chen, W.-C.; Wang, C.-T.; Lin, Y.-H. Recycling waste classification using optimized convolutional neural network. Resour. Conserv. Recycl. 2020, 164, 105132. [Google Scholar] [CrossRef]
  11. Li, N.; Chen, Y. Municipal solid waste classification and real-time detection using deep learning methods. Urban Clim. 2023, 49, 101462. [Google Scholar] [CrossRef]
  12. Bobulski, J.; Kubanek, M. Deep Learning for Plastic Waste Classification System. Appl. Comput. Intell. Soft Comput. 2021, 2021, 1–7. [Google Scholar] [CrossRef]
  13. Mohammed, M.A.; Abdulhasan, M.J.; Kumar, N.M.; Abdulkareem, K.H.; Mostafa, S.A.; Maashi, M.S.; Khalid, L.S.; Abdulaali, H.S.; Chopra, S.S. Automated waste-sorting and recycling classification using artificial neural network and features fusion: A digital-enabled circular economy vision for smart cities. Multimed. Tools Appl. 2022, 82, 39617–39632. [Google Scholar] [CrossRef]
  14. Chen, X.; Huang, H.; Liu, Y.; Li, J.; Liu, M. Robot for automatic waste sorting on construction sites. Autom. Constr. 2022, 141, 104387. [Google Scholar] [CrossRef]
  15. Koskinopoulou, M.; Raptopoulos, F.; Papadopoulos, G.; Mavrakis, N.; Maniadakis, M. Robotic Waste Sorting Technology: Toward a Vision-Based Categorization System for the Industrial Robotic Separation of Recyclable Waste. IEEE Robot. Autom. Mag. 2021, 28, 50–60. [Google Scholar] [CrossRef]
  16. Chen, J.; Fu, Y.; Lu, W.; Pan, Y. Augmented reality-enabled human-robot collaboration to balance construction waste sorting efficiency and occupational safety and health. J. Environ. Manag. 2023, 348, 119341. [Google Scholar] [CrossRef]
  17. Lin, Y.-H.; Mao, W.-L.; Fathurrahman, H.I.K. Development of intelligent Municipal Solid waste Sorter for recyclables. Waste Manag. 2023, 174, 597–604. [Google Scholar] [CrossRef] [PubMed]
  18. Bonello, D.; Saliba, M.A.; Camilleri, K.P. An Exploratory Study on the Automated Sorting of Commingled Recyclable Domestic Waste. Procedia Manuf. 2017, 11, 686–694. [Google Scholar] [CrossRef]
  19. Gupta, T.; Joshi, R.; Mukhopadhyay, D.; Sachdeva, K.; Jain, N.; Virmani, D.; Garcia-Hernandez, L. A deep learning approach based hardware solution to categorise garbage in environment. Complex Intell. Syst. 2021, 8, 1129–1152. [Google Scholar] [CrossRef]
  20. Inamura, T.; Kojo, N.; Hatao, N.; Tokutsu, S.; Fujimoto, J.; Sonoda, T.; Okada, K.; Inaba, M. Realization of Trash Separation of Bottles and Cans for Humanoids using Eyes, Hands and Ears. J. Robot. Soc. Jpn. 2007, 25, 813–821. [Google Scholar] [CrossRef]
  21. Calvini, R.; Orlandi, G.; Foca, G.; Ulrici, A. Development of a classification algorithm for efficient handling of multiple classes in sorting systems based on hyperspectral imaging. J. Spectr. Imaging 2018, 7, 1–15. [Google Scholar] [CrossRef]
  22. Lu, W.; Chen, J. Computer vision for solid waste sorting: A critical review of academic research. Waste Manag. 2022, 142, 29–43. [Google Scholar] [CrossRef]
  23. Nakano, H.; Kawamoto, N.; Umemoto, T.; Katsuragi, I. Development of a Collaborative Robot-Based Support System for Sorting Recyclable Waste; Japan Society of Waste Management: Tokyo, Japan, 2021. [Google Scholar]
  24. Zhong, Z.; Zheng, L.; Kang, G.; Li, S.; Yang, Y. Random Erasing Data Augmentation. In Proceedings of the AAAI Conference on Artificial Intelligence, New York, NY, USA, 7–12 February 2020. [Google Scholar]
  25. Bochkovskiy, A.; Wang, C.-Y.; Liao, H.-Y.M. YOLOv4: Optimal Speed and Accuracy of Object Detection. arXiv 2020, arXiv:2004.10934. [Google Scholar]
  26. Kiyokawa, T.; Katayama, H.; Tatsuta, Y.; Takamatsu, J.; Ogasawara, T. Robotic Waste Sorter with Agile Manipulation and Quickly Trainable Detector. IEEE Access 2021, 9, 124616–124631. [Google Scholar] [CrossRef]
  27. Nikko Petris (PET Bottle Sorting Machine). Available online: https://www.nikko-net.co.jp/product/environment/petris.html (accessed on 26 November 2023).
  28. N. Craft Rattling: Automatic Beverage Container Sorting Machine. Available online: https://www.n-craft.biz/product12.html (accessed on 26 November 2023).
  29. Nihon Cim Sorting Machine. Available online: https://www.nihon-cim.co.jp/product/sorting-machine/hisen.html (accessed on 13 September 2021).
  30. Paszke, A.; Gross, S.; Massa, F.; Lerer, A.; Bradbury, J.; Chanan, G.; Killeen, T.; Lin, Z.; Gimelshein, N.; Antiga, L.; et al. Pytorch: An imperative style, high-performance deep learning library. arXiv 2019, arXiv:1912.01703. [Google Scholar] [CrossRef]
  31. Wang, C.-Y.; Bochkovskiy, A.; Liao, H.-Y.M. YOLOv7: Trainable Bag-of-Freebies Sets New State-of-the-Art for Real-Time Object Detectors. In Proceedings of the 2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Vancouver, BC, Canada, 17–24 June 2023; pp. 7464–7475. [Google Scholar]
  32. Lin, T.-Y.; Maire, M.; Belongie, S.; Bourdev, L.; Girshick, R.; Hays, J.; Perona, P.; Ramanan, D.; Zitnick, C.L.; Dollár, P. Microsoft COCO: Common Objects in Context. In Computer Vision—ECCV 2014; Fleet, D., Pajdla, T., Schiele, B., Tuytelaars, T., Eds.; Springer: Cham, Switzerland, 2014; Volume 8693, pp. 740–755. ISBN 978-3-319-10601-4. [Google Scholar] [CrossRef]
  33. Supervisely Supervisely. Available online: https://supervisely.com/ (accessed on 26 November 2023).
  34. Zivkovic, Z.; van der Heijden, F. Efficient adaptive density estimation per image pixel for the task of background subtraction. Pattern Recognit. Lett. 2006, 27, 773–780. [Google Scholar] [CrossRef]
  35. Kirillov, A.; Mintun, E.; Ravi, N.; Mao, H.; Rolland, C.; Gustafson, L.; Xiao, T.; Whitehead, S.; Berg, A.C.; Lo, W.-Y.; et al. Segment Anything. In Proceedings of the IEEE/CVF International Conference on Computer Vision, Paris, France, 1–6 October 2023. [Google Scholar] [CrossRef]
Figure 1. The composition of the beverage container sorting robot system.
Figure 1. The composition of the beverage container sorting robot system.
Sustainability 16 10155 g001
Figure 2. Actual condition of the beverage container sorting system developed for this study.
Figure 2. Actual condition of the beverage container sorting system developed for this study.
Sustainability 16 10155 g002
Figure 3. Image of experimental camera field of view.
Figure 3. Image of experimental camera field of view.
Sustainability 16 10155 g003
Figure 4. Systematic algorithm of software control system.
Figure 4. Systematic algorithm of software control system.
Sustainability 16 10155 g004
Figure 5. Confusion matrix of the trained model. This confusion matrix should be read vertically (with columns representing actual classes and rows representing predicted classes). Colum “PET” means a sum of 0.99 because other prediction values are rounded to 0.
Figure 5. Confusion matrix of the trained model. This confusion matrix should be read vertically (with columns representing actual classes and rows representing predicted classes). Colum “PET” means a sum of 0.99 because other prediction values are rounded to 0.
Sustainability 16 10155 g005
Figure 6. Segmentation mask for Gaussian background motion model.
Figure 6. Segmentation mask for Gaussian background motion model.
Sustainability 16 10155 g006
Figure 7. Segmentation mask for SAM method.
Figure 7. Segmentation mask for SAM method.
Sustainability 16 10155 g007
Figure 8. The algorithm for identifying the longest cross-section.
Figure 8. The algorithm for identifying the longest cross-section.
Sustainability 16 10155 g008
Figure 9. Thresholding by lightness value in our conveyor belt illustrates the lighting condition.
Figure 9. Thresholding by lightness value in our conveyor belt illustrates the lighting condition.
Sustainability 16 10155 g009
Figure 10. The sample image of our color algorithm result.
Figure 10. The sample image of our color algorithm result.
Sustainability 16 10155 g010
Figure 11. The image of image recognition AI training data.
Figure 11. The image of image recognition AI training data.
Sustainability 16 10155 g011
Figure 12. Examples of image recognition AI does not send acquisition signals from the AI view and object view.
Figure 12. Examples of image recognition AI does not send acquisition signals from the AI view and object view.
Sustainability 16 10155 g012
Figure 13. Examples of errors due to robot control problems.
Figure 13. Examples of errors due to robot control problems.
Sustainability 16 10155 g013
Figure 14. Examples of errors due to object shape problem.
Figure 14. Examples of errors due to object shape problem.
Sustainability 16 10155 g014
Figure 15. Example of error due to conveyor belt transport.
Figure 15. Example of error due to conveyor belt transport.
Sustainability 16 10155 g015
Table 1. The overview of this study and existing beverage container sorting robot.
Table 1. The overview of this study and existing beverage container sorting robot.
Automatic Beverage
Container Sorting Robot
Sorting TargetsRecognition MethodIdeal Placement Location
This studyPET bottles, cans, clear glass, brown glass, other glassImage recognitionEmission site and intermediate treatment facility
PETRISPET, other bottlesObject detection/transmission detection sensorIntermediate treatment facility
GARAGARAPONPET, aluminum cans, steel cansAir Knife, Aluminum separatorEmission site
HISENPET, other bottlesSuction BlowerIntermediate treatment facility
Table 2. Detailed specification of the components in the system.
Table 2. Detailed specification of the components in the system.
Robot System ComponentsDetails
CPU12th Gen Intel Core i9-12900K × 24
Intel, Santa Clara, CA, USA.
GPUGTX 3090 Ti
NVIDA, Santa Clara, CA, USA.
RAMCFD Crucia DDR4-320MH2 16GB × 2
Crucial, Boise, ID, USA
StorageWeston Digital SN770 500 GB SSD
Western Digital, San Jose, CA, USA.
OSUbuntu 20.04 64 bits
Canonical, London, UK
CameraReal sense D435
Intel, Santa Clara, CA, USA.
RobotHiwin delta robot
Hiwin, Taiwan, China.
Conveyor beltMMW-H-340-1200-400-1VH-20
Okura Yusoki Co., Ltd., Hyogo, Japan.
relaySONGLE SRD-05VDC.SL-G
National Institute of Advanced Industrial Science and Technology, Ibaraki, Japan.
Table 3. Metric of the trained model.
Table 3. Metric of the trained model.
RecallPrecision mAP 0.5–0.95mAP 0.5
0.99480.9840.89680.9967
Table 4. The results of the AI recognition experiment.
Table 4. The results of the AI recognition experiment.
False Recognition TypeWith LabelWithout LabelTotal Amount
MisrecognitionPET-Glass32326
Glass-PET011
Glass-CAN101
total42428
False positive rate (false positive/total)1%24%6%
Total number of experiments394101495
Table 5. The results of the automatic beverage containers sorting robot experiments on an actual scale.
Table 5. The results of the automatic beverage containers sorting robot experiments on an actual scale.
QuantitiesRateCalculation Method
Acquisition actions41383%Actions/inputs received
Succeed picking32278%Number of acquisitions/activities
Correct sorting29993%Positive solutions/successful picking
Picking miss16834%Number of acquisition misses/inputs
Misrecognition236%Identification error/number of inputs
Missorting51%(identification/acquisition error)/number of inputs
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Cheng, T.; Kojima, D.; Hu, H.; Onoda, H.; Pandyaswargo, A.H. Optimizing Waste Sorting for Sustainability: An AI-Powered Robotic Solution for Beverage Container Recycling. Sustainability 2024, 16, 10155. https://doi.org/10.3390/su162310155

AMA Style

Cheng T, Kojima D, Hu H, Onoda H, Pandyaswargo AH. Optimizing Waste Sorting for Sustainability: An AI-Powered Robotic Solution for Beverage Container Recycling. Sustainability. 2024; 16(23):10155. https://doi.org/10.3390/su162310155

Chicago/Turabian Style

Cheng, Tianhao, Daiki Kojima, Hao Hu, Hiroshi Onoda, and Andante Hadi Pandyaswargo. 2024. "Optimizing Waste Sorting for Sustainability: An AI-Powered Robotic Solution for Beverage Container Recycling" Sustainability 16, no. 23: 10155. https://doi.org/10.3390/su162310155

APA Style

Cheng, T., Kojima, D., Hu, H., Onoda, H., & Pandyaswargo, A. H. (2024). Optimizing Waste Sorting for Sustainability: An AI-Powered Robotic Solution for Beverage Container Recycling. Sustainability, 16(23), 10155. https://doi.org/10.3390/su162310155

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop