Recent Developments in Technology for Sorting Plastic for Recycling: The Emergence of Artificial Intelligence and the Rise of the Robots

Abstract

Plastics recycling is an important component of the circular economy. In mechanical recycling, the recovery of high-quality plastics for subsequent reprocessing requires plastic waste to be first sorted by type, color, and size. In chemical recycling, certain types of plastics should be removed first as they negatively affect the process. Such sortation of plastic objects at Materials Recovery Facilities (MRFs) relies increasingly on automated technology. Critical for any sorting is the proper identification of the plastic type. Spectroscopy is used to this end, increasingly augmented by machine learning (ML) and artificial intelligence (AI). Recent developments in the application of ML/AI in plastics recycling are highlighted here, and the state of the art in the identification and sortation of plastic is presented. Commercial equipment for sorting plastic recyclables is identified from a survey of publicly available information. Automated sorting equipment, ML/AI-based sorters, and robotic sorters currently available on the market are evaluated regarding their sensors, capability to sort certain types of plastics, primary application, throughput, and accuracy. This information reflects the rapid progress achieved in sorting plastics. However, the sortation of film, dark plastics, and plastics comprising multiple types of polymers remains challenging. Improvements and/or new solutions in the automated sorting of plastics are forthcoming.

1. Introduction

The path to resource utilization and circularity passes through the recycling of post-consumer plastics and the reduction in the amounts of plastics landfilled or incinerated. The existing recycling infrastructure cannot manage well the high volume and complexity of the plastic waste generated. Higher throughput and recycling rates of post-consumer plastics can be achieved by increasing the efficiency of sorting. Recycled plastic that is poorly sorted increases reprocessing costs and decreases the value of reprocessed plastics [1].

The recycling of plastics can be achieved via mechanical or chemical processes [2,3,4,5]. The mechanical recycling of plastic involves identification, sorting, washing, shredding, and reprocessing of desired types of plastic. All these take place while the solid polymer remains intact. Classification and sortation are key in mechanical recycling, as plastics need to be separated by type and color before reprocessing. Chemical recycling, also called advanced recycling, involves breaking down used plastics into raw materials for fuel, new plastics, or other chemicals [4,6] using chemical processes such as liquefaction, pyrolysis, and gasification [4,7]. Chemical recycling also includes chemolysis, which depolymerizes polymers into monomers, and dissolution/precipitation, which is a solvent-based physical separation of different polymers that does not involve breaking polymer chains [4,8,9,10]. Pyrolysis is the most common chemical recycling process, but not all plastics are suitable for pyrolysis. For example, the pyrolysis of poly(vinyl chloride) (PVC) is undesirable due to the production of hydrogen chloride (HCl), which causes corrosion to equipment. Similarly, the use of poly(ethylene terephthalate) (PET) in pyrolysis is limited due to the low yield (~50 wt %) and oxygen content, which may lead to combustion. Hence, plastics need to be sorted out for both mechanical and chemical recycling applications [4,11].

Separation of plastic by type is typically performed at Materials Recovery Facilities (MRFs). For recycling purposes, plastics are classified as (1) PET, (2) high-density polyethylene (HDPE), (3) PVC, (4) low-density polyethylene (LDPE), (5) polypropylene (PP), (6) polystyrene (PS), and (7) “other”, where the numbers 1, 2, …, 7 refer to the Plastic Identification Codes [12]. In principle, all these types of plastics have the potential to be sorted; however, most types of plastic have low or no market value to justify the cost of sorting, the exceptions being PET and HDPE. The “residual” plastic is typically landfilled.

At MRFs, post-consumer plastics are sorted manually by operators and/or mechanically, based on differences in the properties of plastics [13,14]. Separation of plastics by type, color, or shape/size requires specialized equipment such as optical sorters [15,16,17]. The optimal sorting method depends on the plastic type and product(s) of interest. Sorters of different modality can be combined to improve the sorting efficiency and yield of the desired product [17]. Manual sorting can reduce contamination and improve product quality, but can be relatively costly and slow for high volumes of waste, and potentially dangerous to operators [18,19]. Automated sorting, however, can prove more efficient and cost-effective [18].

Challenges that MRFs face in the sortation of plastic recyclables were identified in a study that our team conducted two years ago [20]. MRFs utilizing manual sorting reported a lower throughput compared to MRFs with automated sorting. For automated MRFs, one of the main challenges are tanglers wrapping around sorting equipment. Films are difficult to sort and typically have high contamination rates. The sortation of black plastics was another challenge identified in this study [20]. The same study compiled and compared information on commercially available automated sorting equipment, thus capturing the progress made in the ten years prior to our study, when similar reports were last published.

The state-of-the-art in technology and equipment for the classification and sortation of plastics is analyzed here. Spectroscopy is primarily used to identify plastics, increasingly augmented by machine learning (ML) and artificial intelligence (AI). The previous report on equipment for sorting plastic [20] dates from over two years ago. In the meantime, the demand for plastics recycling has increased and the recycling technology has advanced. These motivate the present updated inventory of established and emerging sorting equipment and their evaluation regarding their sensors, types of plastics they can sort, primary application, throughput, and accuracy. The information compiled here captures the rapid progress made in recent years that holds promise for positive future developments.

2. Spectroscopic Methods for Identification of Plastic Type

Spectroscopy techniques currently used to identify plastic waste in the context of sorting are based on VIS (color analysis), near infrared (NIR), and X-ray fluorescence (XRF) [21,22,23,24,25,26]. Mid-infrared (MIR) spectroscopy, hyperspectral imaging, and shape recognition show potential for classifying plastics but are not yet deployed in large-scale sorting [27,28,29,30]. NIR, XRF, and VIS have different advantages in sorting different types of plastic, as outlined below.

NIR sensors detect variations in the absorption, transmittance, and scattering of light by different materials in infrared wavelengths, which inform on the plastic type [19,31]. NIR intensity can also be influenced by the color, surface texture, and shape of the plastic object [28,32]. Advantages of NIR include high-speed, high penetration depth, and high signal-to-noise ratio [29]. However, NIR is not effective for black plastics, because the black pigment absorbs most light, nor for plastics that incorporate brominated flame retardants (BFRs) [33]. Since the NIR spectra are affected by instrumental noise, baseline drift, and scattered light, preprocessing of spectral data is required for sorting applications [34]. MIR probes the CH₃, methylene (CH₂), and methine (CH) functional groups and can address some of the drawbacks of NIR, but, at present, standoff measurements are not practical [24,35].

XRF shines primary X-rays onto the plastic object under testing, and measures the fluorescent X-rays emitted at a different wavelengths by the elements present in the plastic [36]. XRF sorters are widely used to classify PVC and plastics containing BFRs [19]. However, their application is typically limited to the removal of PVC contaminant from PET [37,38].

Visible spectrometry works by analyzing the total range of the visible spectrum, thus accurately characterizing all colors. VIS sorts plastics by color [14,39] using a prism-coupled color camera [40] which measures colors (red, green, and blue) based on intensity [40].

Some of the challenges highlighted above can be addressed by combining spectroscopy with machine learning or artificial intelligence [41,42,43].

3. Utilization of Machine Learning or Artificial Intelligence in Plastic Type Identification

To improve the identification accuracy and separation efficiency of plastics, optical detection methods combined with ML/AI are developed and increasingly being deployed [44]. ML is designed to emulate human intelligence by using data to learn from the surrounding environment [45]. Plastics are identified, classified, and sorted based on data captured digitally in real time with sensors or cameras, and then applying algorithms [46]. The classification and detection are done using a combination of sensors and ML algorithms.

ML algorithms can be supervised, unsupervised, semi-supervised, and reinforced [47] (Figure 1). Supervised machine learning algorithms predict an outcome based on previously characterized input data [48,49]. For their learning, supervised models need to be trained with tagged or sorted data [48]. In unsupervised learning, the data input into the model is not presorted or tagged, with no guide to a desired output. Unsupervised models are ideal when used to identify unknown relationships in training data [50,51]. Semi-supervised learning is the combination of supervised and unsupervised learning [51]. The approach employs a limited collection of sorted or tagged training data alongside an extensive compilation of untagged data. The models are used to conduct specific computations to reach the correct outcome. Moreover, the semi-supervised models need to perform the learning and data organization, while they are only given small sets of training data. Semi-supervised models can have better accuracy than supervised and unsupervised models [51]. For waste management applications, supervised learning (classification) and neural network models are often used [48,49], as discussed below.

Identification of materials though ML is done in three stages: data processing and feature extraction, selection of machine learning algorithms as classifiers, and testing and performance evaluation [44,59]. The input data are extracted from sensors (e.g., images, spectra), while extraction of features is done through image processing. The spectral data are often pre-processed for baseline corrections and to reduce their dimensionality (e.g., by principal component analysis, PCA), thus helping to reduce the computation time. The classifier transforms the data and, based on these transformations, identifies the optimal boundary between the possible outputs. Performance evaluation selects the best model [59]. Classification models (Classifiers) anticipate or draw conclusions of the input data given for training, and then predicts the class and category for the data. The ML workflow shown in Figure 2 is often utilized for plastic sorting.

3.1. Classifiers

Several algorithms, such as the adaptive network fuzzy inference system (ANFIS), artificial neural networks (ANNs), decision trees (DTs), support vector machines (SVMs), naive Bayes, k-nearest neighbor (KNN), and random forest (RF) have been used in machine learning and deep learning to classify waste [60]. The classification algorithms (models) that have been utilized in sorting plastic waste are discussed below.

3.1.1. Convolutional Neural Networks (CNNs)

Neural networks find many applications in solving a range of problems such as classification and regression [47]. CNN is intended to resemble the human brain. CNN is made up of neurons, which receive input signals and lay out output by measuring the input data with images on many channels. Images go through convolution layers with filters, as indicated in Figure 3 [47,61]. Most calculations are conducted in the convolutional layers. The rectifier function is used by the activation layer to correct the non-linearity of the image, while the pooling layer limits the search of an image on optimal features (e.g., dimension). Afterwards, the assembly is converted into a column by interconnected layers and is transmitted to the neural network for processing. Finally, the activation function sorts the output [61].

CNN is useful in computer vision to extract features in images (e.g., color, size) [27,41,60]. On the basis of differences in the granularity of images, three types of datasets can be identified, resulting in three different approaches: classification, object detection, and segmentation [41]. In classification, the class of an object in an image is determined without providing its location. In object detection, details (categories) and multilabel locations of objects in an image are identified by drawing boundary boxes around them. In segmentation, a pixelwise mask of each object in the image is provided, which facilitates the identification of the shapes of different items [41]. CNN encompasses several variations in architecture (e.g., feed forward networks, deep feed forward). Different CNN architectures are able to extract the features in images layer-by-layer using the information flow from input to output [41].

3.1.2. Support Vector Machines (SVMs)

SVMs are non-linear ML algorithms that have been applied to classify the type and shape of plastics [27,62,63,64]. SVMs construct an ideal boundary within the covariate space (p-dimension) based on the provided samples (x₁, y₁), …, (x_N, y_N) [65]. The input data in SVMs are gathered as points, and these points are classified in a linear manner based on the hyperplane [61]. Furthermore, the algorithm finds or modifies the variables that best fit the hyperplane, and classifies the item being analyzed (e.g., plastic item) into their respective categories. In the SVM classification process, input vectors that are on the hyperplane of the spatial separation belong to one class, and the positions on the other side of the plane belong to a different class [27,61].

3.1.3. Decision Tree Classifier (DTC)

DTC algorithms employ multiple stages to divide data into smaller and less complex sections according to specific criteria [66]. These algorithms are often based on the “if-then-else decision rules”, where classifications are conducted in a tree-type structure, with complexity being directly proportional to the depth of the tree. The selection of functions is automatically done with qualitative and quantitative data [61,67]. DTC constructs a hierarchical structure in the form of a tree, where every inner node corresponds to a characteristic or property, while each terminal node represents a classification or group [65]. The algorithm chooses the feature that provides the most useful information at each node [65,68]. When there is an item (e.g., plastic) that requires sorting, it progresses through the decision tree, commencing from the initial node. At each internal node, the algorithm assesses the value of the corresponding characteristic for the plastic input and proceeds along the suitable branch based on the value of that characteristic [65,68]. Once the algorithm arrives at a leaf node, it designates the relevant class or category to the input plastic. The plastic is then assigned the anticipated class or category [65,68]. In decision trees, data points that cannot be separated linearly are mapped to higher dimensional spaces by the DTC algorithm with appropriate kernel functions so that they can be separated into these spaces. Decision tree algorithms have been used to develop prediction models for waste generation [69,70].

3.1.4. Random Forest (RF)

Random forest (RF) classifiers and extra tree classifiers are ensembles of decision trees that are interconnected. In RF classifiers, the input data are subsampled with bootstrap replicas, whereas extra tree classifiers use original data to create subsets of each tree [66]. RF classifiers have been successfully used to classify different plastic materials with accuracies over 98% [66].

3.1.5. k-Nearest Neighbor (KNN)

KNN algorithms use distance measurement methods [53]. When sorting new plastic items, the KNN algorithm measures the disparity (distance) between the plastic being categorized (sorted) and all the plastics data in the training set. These algorithms identify the k nearest samples to the test data and assigns the most prevalent class label from the learning samples [61]. This process employs a method called “majority voting”, where the label that garners the highest number of votes is selected as the predicted label for the given plastic input. These classifiers do not make assumptions on how data are distributed, as most data often do not follow a theoretical distribution [53]. KNN algorithms have been used in combination with spectroscopy to classify and sort waste plastics [71,72].

3.1.6. Naive Bayes Classifiers

Naive Bayes classification algorithms utilize the Bayes theorem for probabilistic classification (Equation (1)) [73,74]. The Bayes theorem integrates new evidence (i.e., new data) with previous probabilities of hypotheses to obtain new probabilities for the hypotheses [73]. Through the examination of the input data of a given set parameters or features, denoted as “B” in Equation (1), Naive Bayes classifiers can estimate the probability of the input data associated with a particular class, denoted as “A” [74].

$$P\left(A|B\right)=\frac{P\left(B|A\right) P\left(A\right)}{P\left(B\right)}$$

(1)

Naive Bayes classifiers operate by assuming that classified features are independent of each other, given the class variable [66,73,75]. To perform the classification of input data, an assessment of the probability of it belonging to each of the existing classes is conducted, and the class with the highest probability is then identified as the one to which the input data belongs (Equation (2)).

$$A=argma{x}_{a}P(a|b_1, \dots , b_n)$$

(2)

where b₁ is one of the n features or predictors.

The Naive Bayes classifiers have predetermined structures, and during the training phase of the classifier, the class probabilities and conditional probabilities are computed based on the provided training data. Subsequently, the generated probability values are utilized to categorize new observations [76]. This process allows the classifiers to estimate the likelihood of events or outcomes by utilizing conditional probabilities [66,73]. To sort plastic waste, Naive Bayes classifiers can be used by gathering information on different attributes of plastic objects, such as color, shape, size, and composition [77,78]. The collected data are prepared to ensure their reliability, and the classifier is trained on this dataset, acquiring knowledge of conditional probabilities and class probabilities connected to the attributes [78,79]. Subsequently, appropriate attributes are derived from the plastic waste items and utilized as inputs for the trained classifier. The classifier then computes probabilities and decides the most probable class for each item, facilitating the categorization of plastic waste into various groups based on the classification outcomes [73,77,78,79].

3.1.7. Logistic Regression

Logistic regression algorithms are designed to estimate the likelihood of one of two possible outcomes (classes) and make a definitive prediction based on various input parameters. Test data points are predicted using binary scales that range from zero to one. Points with values exceeding 0.5 are assigned to class 1, while points with values below 0.5 are assigned to class 0 [53,65]. For example, logistic regression can be used to sort clear plastics from colored plastics, given some input parameters. In cases were more than two outcomes or classes are required, multiclass logistic regression can be used [53]. In regression metrics, true targets are compared with their corresponding predictions, where metrics are R²-score, mean absolute error (MAE), and root mean square error (RMSE) [46]. The closer to 1 R²-score is, the more accurate the model is [46].

3.1.8. You Only Look Once (YOLO)

YOLO integrates image sensors and AI detection algorithms (e.g., Neural networks) to detect and locate objects [80]. YOLO works by applying a neural network to an image, breaks down the image into grid cells, and forecasts the grid cell coordinates into bounding boxes [80,81]. In YOLO, each grid has a corresponding vector in the output that determines if the object is located in that grid cell; if yes, it helps determine the class of the object and the estimated boundary region of the object [81]. Finally, the algorithm generates the final result, which includes the remaining bounding boxes along with their corresponding categories and confidence scores that best fits the items being sorted. The YOLO detection and location of an object is done by looking at the object only once, or a process known as one-stage detection. In a one-stage detector, location and classification of objects are performed at the same time, contrary to a two-stage detector used in algorithms such as a CNN [80]. As a result, a one-stage detector can be computationally efficient compared to a two-stage detector, though less accurate. The YOLO algorithm training allows it to recognize and classify each category, such as plastic bottle, plastic bag, etc. YOLO can be useful in classifying plastics that differ in physical characteristics (e.g., transparency, flexibility) but have similar chemical structure (e.g., PET and PET-G, polyethylene terephthalate glycol) and similar spectra [80]. When sorting plastics with similar chemical compositions, YOLO can reportedly reach an accuracy > 91.7% and mean Average Precision (mAP) much better than traditional optical sorters [80,82].

3.2. Performance

ML classification performance can be measured in terms of accuracy, recall, precision, and F1-score [66]. The performance of ML algorithms can be evaluated using these metrics by first splitting the dataset into training and test data, and then comparing the predictions of the trained algorithms for test data to the known target variables of the test dataset [46].

In classification problems, y (true labels or classes of a classification problem) can have two values: “positive” (P) and “negative” (N). True (T) and false (F) predictions can be visualized in a 2 × 2 confusion matrix as shown in Equation (3) [46].

$$\mathrm{Confusion} \mathrm{Matrix}=\left[\begin{array}{cc}TP& FP\\ FN& TN\end{array}\right]$$

(3)

where true positives (TP) are the number of samples correctly predicted as “positive”, false positives (FP) are the number of samples wrongly predicted as “positive”, true negatives (TN) are the number of samples correctly predicted as “negative”, and false negatives (FN) are the number of samples wrongly predicted as “negative”.

Classification metrics such as accuracy, recall, precision and F1-score can be obtained from TP, FP, TN, and FN as discussed below. Accuracy is a measure of correctly predicted observations among the total observations (Equation (4)). Accuracy computes how many times a model made a correct prediction across the entire dataset. Accuracy is often useful in evaluating model performance in a class-balanced dataset, where each class in the dataset has the same number of samples [83]. Recall is the ratio of correctly predicted observations among all observations for each class (Equation (5)). Precision is the ratio of correct predictions among all predictions assigned to a class (Equation (6)). F1-score is the weighted average of precision and recall (Equation (7)) [66]. The observations reported in accuracy, recall, precision, and F-1 score can be translated into purity and yield in the case of plastics.

$$Accuracy=\frac{TP+TN}{TP+TN+FP+FN}$$

(4)

$$Recall=\frac{TP}{TP+FN}$$

(5)

$$Precision=\frac{TP}{TP+FP}$$

(6)

$$F1-Score=\frac{2\times Recall\times Precision}{Recall+Precision\prime }$$

(7)

Accuracy, precision, and recall are often useful in evaluating ML model performance in class-balanced datasets, where each class in the dataset has the same number of samples; however, this can be challenging in unbalanced datasets [83]. Most real-world data are often imbalanced datasets; thus, the F1 score is often used for imbalanced datasets [84]. In imbalanced datasets, for precision and recall, one metric comes at the cost of another. The F1 score combines precision and recall, to better reflect the model’s accuracy.

A compilation of performance in accuracy of different algorithms that have been employed to classify and identify waste, extracted from various published studies, is presented in

Table 1. Performance of different ML models for waste classification and identification.

Model	Model Description	Epoch	Layers	Classification Accuracy (%)	Machine Accuracy	Materials Sorted	Reference
CNN	CNN			99.74			[41]
CNN	ResNet-50	24	50	98.81	89.77	PET plastic, plastic bottles, metal, glass	[85]
	CNN	20	15	87			[86]
	Mask-RCNN			89.6	55.6	Opaque and clear plastic bottle, opaque plastic container, cardboard box, drink can	[87]
	Mask R-CNN			71.9	66	Construction waste, i.e., cotton gloves, wood, ferrous items, plastic pipe, bamboo, paper, steel bar	[88]
	Faster R-CNN			91		Cardboard, plastic, glass, paper, metal, and trash	[89]
	Pre-trained Mobile Net			90		Garbage (tested only on bottles)	[90]
	CNN			95.3		Glass, paper, cardboard, plastic, metal, and trash	[91]
	CNN			83		Plastic, paper and metal	[62]
	CNN			76		Plastic, paper, cardboard, metals	[48]
	Fast R-CNN			88
SVM	SVM			94.8		Plastic, paper and metal
	SVM			78.3		Paper, plastic, metal, and glass	[92]
	SVM			96.5		Metal, paper, glass, PET	[93]
	SVM			95.5		PET, HDPE, LDPE, PVC, PP, and PS	[94]
KNN	KNN			98.8		PET, HDPE, LDPE, PVC, PP, and PS	[94]
Logistic regression	Logistic regression			92.9		PET, HDPE, LDPE, PVC, PP, and PS	[94]
Random Forest	Random Forest			97.3		PET, HDPE, LDPE, PVC, PP, and PS	[94]
Naive Bayes	Naive Bayes			90.2		PET, HDPE, LDPE, PVC, PP, and PS	[94]
YOLO	YOLOv3			94.99			[95]
	YOLOX			94.5			[96]
	YOLOv4			95.16

and Figure 4. As shown there, different algorithms can attain high levels of classifications of plastic waste. However, direct comparison of different algorisms (e.g., CNN vs. SVM) is currently not possible due to different sizes of databases, items identified, number of layers in the models, training sets, etc.

4. ML and AI in Combination with Spectroscopy for Plastic Type Identification

As discussed above, spectroscopy plays a key role in the identification and sorting of plastic waste at MRFs. However, various spectroscopy techniques have their limitations when it comes to sorting plastic. For example, NIR has low resolution and cannot sort black plastics. MIR has slow spectrum acquisition and cannot adequately differentiate between HDPE and LDPE. Raman has low sensitivity and is subject to interference from fluorescence. Laser-induced breakdown spectroscopy (LIBS) does not provide molecular structure information and has difficulty distinguishing polymers with similar chemical formulas [47]. Fluorescence is influenced by the overlap of the molecule’s vibrational level with its excited electronic energy level [97].

The fact that no single spectroscopy technique is suitable for all types of plastics motivates the combination of spectroscopy with ML/AI in order to address the limitations [47,98]. The ML models or algorithms discussed in Section 3 have demonstrated the ability to contribute to this end. For example, ML/AI (e.g., CNN) can identify plastics by color, thus correctly sorting black plastics, or sort a plastic based on color (e.g., clear vs. colored PET) [98]. Thus, the combination of AI or ML with spectroscopy-based techniques can increase the sorting accuracy [66]. Carrera et al. [66] used different ML algorithms (SVM, kNN, Naïve Bayes) applied on IR (NIR and MIR) spectra to develop classification models for plastics (PE and PET in the first experiment, PE, PET, PP, and PS in the second experiment, PE, PP, PS, and PVC in the third experiment, and PE, PET, PP, PS, and PVC in the fourth experiment), and reported model accuracy, precision, recall, and F1-score rates all over 99% [66]. Neo et al. [47] used CNN, residual networks and inception networks in a decision tree structure with IR and Raman spectra dataset containing over 20 different polymers to classify and identify PE, PP, and PET with an accuracy of 94.9 and 96.7% with the Raman and IR datasets, respectively [47]. The use of a CNN in combination with spectroscopy technologies did not necessarily require pre-processing of spectral data due its feature extraction capabilities [64].

ML can be combined with optical spectroscopy techniques such as NIR or Mid-IR to increase plastic sorting efficiency (Figure 5).

Bonifazi et al. [98] combined ML and data from laser-induced fluorescence (LIF) to identify and sort black plastics (EPS, PS, PP, HDPE) from a plastic waste stream. Long et al. [99] used a combination of CNN ML and MIR (collected at the rate of 100 Hz) for a fast and accurate characterization of mixed plastics (PE, PP, PS, PVC), reaching an overall accuracy close to 100% [99]. Here, MIR was upconverted from the band 2.0–5.0 μm to the near-visible region 695–877 nm, eliminating the thermal noise present in the MIR range for better sortation [99]. The combination of MIR and ML (CNN) enabled the sortation of plastics by type and color (i.e., blue PS, black PE, deep blue PP, and white PVC) [99].

Neo et al. [47] combined ML and Raman or IR for the identification of plastic waste (consisting of PE, PP, and PET) from a dataset containing over 20 polymers [47]. The identification was conducted using Polymer Spectra Decision Net (PSDN) architecture, achieving an accuracy of 94.9% for Raman and 96.7% for FTIR. The developed PSDN had two neural network modules, with the first trained to classify spectral data into recyclable and non-recyclable, and the second neural network classified recyclable polymers into their individual classes (i.e., PET, PE, PP, Other). PSDN [47] reported higher identification and sortation accuracy compared to the end-to-end neural network often used in ML [47].

Various other models such as like Bernoulli NB, Gaussian NB, decision tree, ensemble models, KNN, SVM, linear models, PLS-DA, and a neural network (MLP) have been tested on spectral data, and the results obtained indicated that five classifiers had accuracy, precision, recall, and an F1-score over 95%, with the MLP classifier having the best performance with 99.71% accuracy, 99.35% precision, 99.82% recall, and 99.58% F1-score [66]. Gaussian NB, Bernoulli, and PLS-DA were reportedly the least effective classifiers, with accuracies of 29.1%, 31.2%, and 75%, respectively [47,66,100].

The following case studies demonstrated the success of AI-based sorting technologies in improving recycling rates. Wilts et al. [17] analyzed the increase in recycling rates and the purity of recovered materials at an MRF in Spain using an AI-based robot (ZenRobotics, Vantaa, Finland) to supplement or replace manual sorting. The waste input of the study comprised 13 different materials, including aluminum, cardboards, HDPE, and textiles. The accuracy or purity of sorted HDPE was approximately 100%, with a recovery rate between 60 to 80% [17]. Manea et al. evaluated the use of smart bins vs. manual sorting or waste segregation by airport passengers, and reported an accuracy of 62% for waste segregation by airport passengers, whereas the smart bin achieved a 90% classification accuracy [100].

5. Application of Robotics in Plastic Waste Management

AI-informed robots have the ability to replace manual sorting and can segregate plastic waste by analyzing the captured information (e.g., color, composition) from cameras and sensors [101]. The integration of cyber-physical systems, blockchains, ML, and the IoT can bridge physical and computational infrastructures in waste management, improving the efficiency of identifying and sorting waste (plastics) for recycling [101]. The efficiency of ML models is related to the computational complexities, resources, and requirement (e.g., training time) in learning and performing classification tasks [102]. AI-based robots used in waste management vary based on application, materials to identify or sort, sizes of materials, etc. IT and robotics can be used for prediction of generation waste, roadside waste collection, smart bins, waste monitoring and tracking, and end-of-life treatments such as pyrolysis or mechanical recycling [103,104,105,106,107].

A variety of robots have been reported in the literature, ranging from mobile robots that can be used to collect waste in challenging environments (e.g., beaches) to fixed robots that be employed in MFRs to identify and sort waste [86,87]. Mobile robots can be equipped with tracks, track belts surrounding wheels, a conveyor to move collected waste, robotic arms, grippers, RBG cameras, actuators, proximity sensors, etc. (Figure 6). To enhance the robot capabilities, configurable platforms can be introduced [108]. Such platforms can provide additional degrees of freedom, often used in robotic arms designed for pick-and-place operations (e.g., SCARA robots, Multiple DoFs Robot) [109]. Furthermore, robot capabilities can be enhanced by using grippers with the ability to handle various shapes, or combined grippers that integrate both a suction cup and finger-like appendages, or having grippers specifically for the items to be separated (e.g., plastic films) [109].

With AI-informed robots, the emphasis is on high speed and low power consumption to reduce sorting costs, making parallel structures like Delta robots a preferred choice, and robots with high degrees of freedom more able to handle materials [108].

AI or ML in combination with automated equipment (robots) making the use of computer vision, sensors, arms, grippers, and suction systems are now being extensively investigated for waste management applications [109]. Lu et al. [47] discussed the usage of both deep learning and machine learning algorithms with computer vision to identify and sort municipal solid waste [104,110]. Sundaralingam et al. [110] reported a waste segregation system that could segregate paper, plastic, metal, organic waste, glass, and one more additional empty class into appropriate bins, using a TensorFlow object detection model and a microcontroller. The developed system could predict and segregate waste in the appropriate bin, with a mean Average Precision (mAP) of 86.5% and recall = 88.3% [110].

Recent studies on robots in plastic waste management often focus on improving the accuracy and efficiency by developing and integrating better sensors and cameras, and better algorithms to accurately classify and sort different types of waste (plastics) [103,111]. Robots can be used to sort plastics based on texture, identifying worn-out plastics and plastics in great physical conditions. With the integration of IT, machine learning, and deep learning into robotics, robots can characterize the shape, size, texture, and colors of different waste materials, and sort them based on adequate categories [109].

Though AI robots have the advantages listed above, they are subject to various limitations, such as not being able to differentiate plastic bottles from glass bottles of the same shape, or to determine between rigid and rubber bottles. Such challenging waste requires better end-effectors and sensors. The end-effector deals with the ability to grasp and sort different waste materials with dirt or deformations, and simultaneously handle both 2D and 3D shaped plastics, while challenges with sensors involve the ability to characterize the shape, color of wet objects, or objects covered with dirt [109]. Examples of AI-informed robots used to sort plastic waste are shown in Figure 7.

6. Recent Advances in Commercial Equipment That Sort Plastics

6.1. Methodology

In this study, equipment for sorting plastic recyclables was identified using publicly available information obtained from manufacturers’ websites and scientific literature. A search for sorting equipment and companies was conducted using Google, Google Scholar, Web of Science, Science Direct, and Engineering Village databases using the keywords “sorting equipment manufacturers or companies”, “optical sorters”, “plastic sorters”, “plastic sorting machines”, “sorting equipment”, and “plastic recycling”.

Sorting devices based on cameras or lasers (e.g., NIR, MIR) to sort whole plastics are classified herein as optical sorters (

Table 2. Inventory of commercially available sorters for whole (i.e., bottle) plastic. NI: no information reported.

Manufacturer/Brand	Equipment Name	Sorting Method	Primary Application	Plastic Identified	Sorts Non-Bottle Rigids in Addition to Bottles	Non-Plastics Sorted	Colors Sorted/Black Plastic Sorted	Throughput (Average)	Accuracy	Features
Anhui Zhongke Optic-electronic Color Sorter Machinery Co., Ltd.	AMD G-LPI (Uses deep learning)	NIR, deep learning, and visible light technology	Can sort labeled bottles, off-label bottles, plastic bottles with labels, mixed plastic bottles in bale form, loose plastic bottles, plastic food packaging
Binder + Co.	CLARITY belt sorting systems	VIS, NIR, induction and XRT, 3D Scanner	PET, PE, PP, PVC	PET, PE, PP, PVC	Yes	Paper, metals, municipal solid waste, wood, and cardboard	Yes/NI	Up to 30 ton/h for 1000 mm sorting width system and 60 ton/h for 2000 mm sorting width system	Accuracy up to 99.9+%	Metal detection
Green Machine LLC	Green Eye Hyperspectral Optical Sorters (Uses AI Tech)	Patented hyperspectral vision systems and AI driven neural net software	Sorts all plastics	1–7 grades of plastic including difficult-to-sort black plastics, barrier bottles, #5’s, PVCs, vinyls, thermal forms; sorts most plastic grades, black plastics, rubber grades, HDPE, LDPE ABS plastics, and more	Yes	Fiber, C&D, E-waste, Textiles, carpeting	Yes/Yes	Up to 12 ton/h (depends on the belt width)	95% or more	Can be trained to identify and pick out almost any type of polymer by shape and chemical composition
Hefei Golden Sorter Co., Ltd.	Plastic Bottle Optical Sorter Gép-T LP (Uses deep learning)	NIR, VIS, deep learning technology	Bottle sorting equipment	non-PET bottle materials, such as PP/PE/PC/PS/ABS/PVC/PA, and other non-PET bottles	Yes	Non-plastic bottles	Yes/Yes	Up to 4 ton/h
Hefei Mayson Machinery Co., Ltd.	MAS-B series bottle separator (Uses deep learning)	Fusion modeling technology, deep learning algorithm, vision system, image processing system, and intelligent self-learning system	Sorts different types of plastic bottles	Non-PET bottle materials, such as PP/PE/PC/PS/ABS/PVC/PA and other non-PET bottle sorting	Yes	Non-plastic bottles	No/No	From 1.5–2.0 ton/h to 4–7 ton/h	Up to 99%	Deep learning system helps in improving the sorting quality/efficiency
MEYER Europe s.r.o.	KL Sorter (Uses AI Tech)	AI cameras working in the electromagnetic spectrum: full RGB visible light, infrared standard, infrared HD, InGaAs, and UV light	Identify different color PET bottles	Detect and remove non-PET bottles, such as PVC/PS/PC/PA/PP/PE/ABS	Yes	Glass, non-ferrous metal, and ore sorting	Yes/Yes	Up to 7 ton/h
MSW Sorting	Optical Sorter (Uses AI Tech)	VIS, NIR, High resolution camera, and AI	Plastic, paper, glass, and other recyclable materials	PET bottles, HDPE bottles, and plastic bottles	Yes	Cans, glass, and cardboards	Yes/NI	Maximum belt speed can reach 6.5 m/s	Over 95%
NRT Optical Sorting	ColorPlus with Max-AI (Uses AI Tech)	RGB line-scan sensor combined with Max-AI	All plastics	Capture form-specific PET (ex. Bottle only, blue/green bottle only.). Capture food-grade-only PET and/or HDPE. Identify black plastics, thermoform trays	Yes	Cardboard, metal cans, and fiber	Yes/Yes
	SpydIR-R with Max-AI (Uses AI Tech)	NIR and Multi-layered vision system and neural networks	Plastics, paper, metals	Capture form-specific PET (ex. Bottle only, blue/green bottle only.). Capture food-grade-only PET and/or HDPE, and identify black plastics, thermoform trays	Yes	Paper, metal cans, wood, cardboard, fiber	Yes/Yes			PET Boost technology for detection of thin-wall PET, wet PET, and full-sleeve PET
Pellenc ST	Compact+	AI CNS platform		PET, PE, PP, paper, wood, domestic waste, organic, RDF			Yes/Yes			Compact+
	Xpert	X-ray along with machine learning	Chlorine or brominated plastic removal	Chlorine or brominated plastic removal	NI	WEE, glass, aluminum	NI/NI	Top Speed ready < 4.5 ms
PicVisa	Ecopack—Model EP Optical Plastic Sorting Machine	NIR, VIS, deep learning	PET/PE recycling, Plastic film (PEBD, PP, HDPE/LDPE, etc.)	PET, HDPE, PP, PS, PVC, EPS, ABS) HDPE, PET, Mixed LDPE, Sorting film (HDPE/LDPE)	Yes, sorting of films (PE) from bottles of the same material	Paper, and cardboards, wood recycling, metal recycling, textile, RDF, construction and demolition waste	Yes/NI			Allows separating the always-present silicone cartridges in HDPE flows. Can add AI technology
TOMRA systems ASA	Autosort Sharp Eye	NIR, Sharp Eye technology (Add-on sensors: VIS, Deep Laiser, metal detector, and AI based Cameras)	Sorts all resins	Plastisc, paper		Wood, RDF, mixed paper, cardboard, metals, and organic waste	Yes/Yes			Can add AI deep learning to improve sorting accuracy and can sort glass and black polymers by adding the DEEP LAISER sensor. Remote access

). Sorting devices that use ML/AI with cameras or lasers (e.g., NIR, MIR) to sort whole plastics are classified herein as AI-based optical sorters (

Table 3. Optical Sorters with incorporated AI.

Manufacturer/Brand	Equipment Name	Sorting Method	Primary Application	Plastic Identified	Sorts Non-Bottle Rigids in Addition to Bottles	Non-Plastics Sorted	Colors Sorted/Black Plastic Sorted	Throughput (Average)	Accuracy	Features
Anhui Zhongke Optic-electronic Color Sorter Machinery Co., Ltd.	AMD G-LPI (Uses deep learning)	NIR, deep learning, and visible light technology	Can sort labeled bottles, off-label bottles, plastic bottles with labels, mixed plastic bottles in bale form, loose plastic bottles, plastic food packaging					1.5–2.0 ton/h for G-LPI2. 3.0–4.0 ton/h for G-LPI4 model
Binder + Co.	CLARITY belt sorting systems	VIS, NIR, induction, and XRT, 3D Scanner	PET, PE, PP, PVC	PET, PE, PP, PVC	Yes	Paper, metals, municipal solid waste, wood, and cardboard	Yes/NI	Up to 30 ton/h for 1000 mm sorting width system and 60 ton/h for 2000 mm sorting width system	Accuracy up to 99.9+%	Metal detection
Green Machine LLC	Green Eye Hyperspectral Optical Sorters (Uses AI Tech)	Patented hyperspectral vision systems and AI-driven neural net software	Sorts all plastics	1–7 grades of plastic including difficult-to-sort black plastics, barrier bottles, #5’s, PVCs, vinyls, thermal forms; sorts most plastic grades, black plastics, rubber grades, HDPE, LDPE ABS plastics, and more	Yes	Fiber, C&D, E-waste, textiles, carpeting	Yes/Yes	Up to 12 ton/h (depends on the belt width)	95% or more	Can be trained to identify and pick out almost any type of polymer by shape and chemical composition
Hefei Golden Sorter Co.Ltd	Plastic Bottle Optical Sorter Gép-T LP (Uses deep learning)	NIR, VIS, deep learning technology	Bottle sorting equipment	non-PET bottle materials, such as PP/PE/PC/PS/ABS/PVC/PA, and other non-PET bottles	Yes	Non-plastic bottles	Yes/Yes	Up to 4 ton/h
Hefei Mayson Machinery Co., Ltd.	MAS-B series bottle separator (Uses deep learning)	Fusion modeling technology, deep learning algorithm, vision system, image processing system, and intelligent self-learning system	Sorts different types of plastic bottles	Non-PET bottle materials, such as PP/PE/PC/PS/ABS/PVC/PA and other non-PET bottle sorting	Yes	Non-plastic bottles	No/No	From 1.5–2.0 ton/h to 4–7 ton/h	Up to 99%	Deep learning system helps in improving the sorting quality/efficiency
MEYER Europe s.r.o.	KL Sorter (Uses AI Tech)	AI cameras working in the electromagnetic spectrum Full RGB visible light, Infrared Standard, Infrared, HD, InGaAs, and UV light	Identify different color PET bottles	Detect and remove non-PET bottles, such as PVC/PS/PC/PA/PP/PE/ABS	Yes	Glass, non-ferrous metal, and ore sorting	Yes/Yes	Up to 7 ton/h
MSW Sorting	Optical Sorter (Uses AI Tech)	VIS, NIR, High resolution camera, and AI	Plastic, paper, glass, and other recyclable materials	PET bottles, HDPE bottles, and plastic bottles	Yes	Cans, glass, and cardboards	Yes/NI	Maximum belt speed can reach 6.5 m/s	Over 95%
NRT Optical Sorting	ColorPlus with Max-AI (Uses AI Tech)	RGB line-scan sensor combined with Max-AI	All plastics	Capture form-specific PET (ex. Bottle only, blue/green bottle only). Capture food-grade-only PET and/or HDPE. Identify black plastics, thermoform trays	Yes	Cardboard, metal cans, and fiber	Yes/Yes
	SpydIR-R with Max-AI (Uses AI Tech)	NIR and Multi-layered vision system and neural networks	Plastics, paper, metals	Capture form-specific PET (ex. Bottle only, blue/green bottle only). Capture food-grade-only PET and/or HDPE, and identify black plastics, thermoform trays	Yes	Paper, metal cans, wood, cardboard, fiber	Yes/Yes			PET Boost technology for detection of thin-wall PET, wet PET, and full-sleeve PET
Pellenc ST	Compact+	AI CNS platform		PET, PE, PP, paper, wood, domestic waste, organic, RDF			Yes/Yes			Compact+
	Xpert	X-ray along with machine learning	Chlorine or brominated plastic removal	Chlorine or brominated plastic removal	NI	WEE, glass, aluminum	NI/NI	Top Speed ready < 4.5 ms
PicVisa	Ecopack—Model EP Optical Plastic Sorting Machine	NIR, VIS, deep learning	PET/PE recycling, Plastic film (PEBD, PP, HDPE/LDPE, etc.)	PET, HDPE, PP, PS, PVC, EPS, ABS) HDPE, PET, Mixed LDPE, Sorting film (HDPE/LDPE)	Yes, sorting of films (PE) from bottles of the same material	Paper and cardboards, wood recycling, metal recycling, textile, RDF, construction and demolition waste	Yes/NI			Allows separating the always-present silicone cartridges in HDPE flows. Can add AI technology
TOMRA systems ASA	Autosort Sharp Eye	NIR, Sharp Eye technology (Add-on sensors: VIS, Deep Laiser, metal detector, and AI based Cameras)	Sorst all resins	Plastisc, paper		Wood, RDF, mixed paper, cardboard, metals, and organic waste	Yes/Yes			Can add AI deep learning to improve sorting accuracy and can sort glass and black polymers by adding the DEEP LAISER sensor. Remote access

). Sorting devices based on cameras or lasers to sort plastic films are classified here as film sorters (

Table 4. Inventory of commercially available film sorters.

Manufacturer/Brand	Equipment Name	Sorting Method	Primary Application	Plastic Identified	Sorts Non-Bottle Rigids in Addition to Bottles	Non-Plastics Sorted	Colors Sorted/Black Plastic Sorted	Throughput (Average)	Accuracy	Features
Binder + Co.	Clarity Plastic	NIR, Reflection VIS, Inductive metal detection	Lightweight packaging, film sorting, plastic flakes, plastic granules, and hallow plastic sorting				Yes/NI	0.5 ton/h for 700 mm sorting width, 0.7 ton/h for 100 mm, and 1 ton/h for 1400 mm		Metal detection
	Clarity Multiway for Light Packaging	NIR, VIS		PET, PE, PP, PVC		Paper and cardboard		Up to 2.1 ton/h for 1000 mmm sorting width and up to 3 ton/h 2000 mmm sorting width
	CLARITY belt sorting	VIS, NIR, induction and X-ray	Plastics, packaging waste, municipal solid waste, refuse-derived fuels, metals, and wood	PET, PE, PP, PVC		Municipal solid waste, refuse-derived fuels, metals, and wood		Up to 30 ton/h for 1000 mm sorting width system and 60 ton/h for 2000 mm sorting width system	Accuracy up to 99.9+%
CP Group (MMS) Sorting Equipment	FilmMax	NIR, color, and metal sensors	Sorts bags, pouches, bags, foil, and other ultra-light products	LDPE/LLDPE, natural/white films, PET, PVC, PS, colored film	No	foil, and other ultra-light products.	Yes/Yes	0.5–3.0 ton/h	Up to 98%	Metal detector upgrade available
	CIRRUS FiberMAX	NIR and color sensors	Flexible plastics packaging (FPP) such as film, bags, pouches			All metal detector		Belt speeds of 1000 ft/min (5 m/s). Capacity 2.0–12.0 ton/h	Up to 98%
RTT Steinert GmbH	Unisort Film EVO 5.0	NIR, VIS, hyperspectral imaging technology	Agricultural film, bio-based film, biodegradable film, conventional PVC film and papers	Identifies and sorts plastics and materials by type. Plastic film, bags, and paper		Beverage cartons, paper, cardboard, paperboard, and textiles	Yes/NI
Pellenc ST	Mistral + Films	NIR	Used to separate films from other plastics	PE film, PP, PVC, metals, fibrous, PS, HDPE		Papers, cardboards, and metals	Yes/No	Up to 2.5 ton/h	Up to 91%
	Mistral + Connect	NIR/VIS spectrum	provides better detection and sorting of PET bottles versus PET trays or thermoforms, paper versus cardboard in sorting centres	PET, PE, PP, paper, films		Wood, domestic waste, organic, RDF	NI/Yes
	Compact+	AI CNS platform		PET, PE, PP, paper, wood, domestic waste, organic, RDF			Yes/Yes
PicVisa	Ecopack—Model EP Optical Plastic Sorting Machine	NIR, VIS, deep learning	PET/PE recycling, plastic film (PEBD, PP, HDPE/LDPE…)	PET, HDPE, PP, PS, PVC, EPS, ABS) LDPE, film (HDPE/LDPE)	Yes, sorting of films (PE) from bottles of the same material	Paper and cardboard, wood recycling, metal recycling	Yes/NI			Allows separating the always-present silicone cartridges in HDPE flows
	ECOPICK (Uses Robotic, AI, and deep learning)	RGB and/or NIR sensors, 3D, AI based robot, deep learning, and machine vision	PET bottles, HDPE bottles, trays, and film	All types		Cans, Tetra Pak, paper, cardboard, glass, textile, aluminium		1 pick/s	>95%
RTT Steinert GmbH	Unisort Film	NIR, VIS	Agricultural film, bio-based film, biodegradable film, conventional PVC film, and papers	Plastic film		bags and paper
TOMRA systems ASA	Autosort Speedair	NIR, SHARP EYE™ technology, and can add-on DEEP LAISER	Plastic films and lightweight packaging	Film (LDPE, HDPE), papers, and packaging			NI/Yes (with DEEP LAISER)			Available as solutions bundle or as an add-on device to an existing AUTOSORT machine setup.
Bollegraaf Group	Opti-Sort	Optical sorting and mechanical sorting by pressure	Processing lightweight materials such as flexible plastic packaging or single sheets of paper					Speed levels to up to 6.5 m/s
NRT Optical Sorting	SpydIR®-R	NIR, In-Flight Sorting	Film, fiber, PET, HDPE, or mixed plastics	PET container stream with high accuracy including PVC, PS, PETG, PLA, and PC, PE, PP, and other polymer contaminants in any combination	Yes	Cardboard, paper, metals, and other fiber		throughput rates exceeding 8 ton/h
	SDi semi-mobile wind shifters	Mechanical sorting based on weight	Plastic, HDPE, film	Plastic, HDPE, film		Wood, cardboard, paper, rubber		Capacities up to 15 t/h		Semi-mobile

), and sorting devices based on cameras or lasers to sort plastic flakes are classified herein as flake sorters (

Table 5. Inventory of commercially available flake sorters.

Manufacturer/Brand	Equipment Name	Sorting Method	Primary Application	Plastic Identified	Sorts Non-Bottle Rigids in Addition to Bottles	Non-Plastics Sorted	Colors Sorted/Black Plastic Sorted	Throughput (Average)	Accuracy	Features
Best	NIREX	NIR, and vision technology	Sorts e-scrap		Yes		Yes/Yes		Depends on product type
Binder + Co	Clarity Plastic	NIR, reflection VIS, inductive metal detection	Light- weight packaging, film sorting, plastic flakes, plastic granules, and hollow plastic sorting				Yes/NI	0.5 ton/h for 700 mm sorting width, 0.7 ton/h for 100 mm, and 1 ton/h for 1400 mm		Metal detection
Buhler	Sortex Z + Series	Vision-based and high-resolution IR sensors	Sorts PET, PVC flakes, and nylon				Yes/Yes	0.675 to 1.16 ton/h depending on model	99.9% or higher
	Sortex N PolyVision		Sorts PET flakes	PET, PVC, PP. PE, PS, PA, POM, PMMA, SAN			Yes/NI	Up to 6 t/h		Integrated chute feeding system
	Sortex B MultyVision		Commodities, but sorts plastics as well	plastic		Pulse, nut, and coffee	No/No	up to 8 t/h		Remote acess for real-time monitoring
	Sortex A GlowVision		Plastic sorting
	Sortex A		Plastics, commodities	Plastics		Nuts, seeds, grains, coffee, pulses	NI/Yes			Remote acess for real-time monitoring
BT-Wolfgang Binder GmbH (Redwave)	Redwave QXR	XRF	Used for PET and WEEE stream purification	Removes PVC and BFR-containing plastics			No/No	2.5 to 8.0 ton/h	80%
	Redwave XRF-P	X-ray Fluorescent	Segregation of dark PVC and brominated plastics from an infeed of shredded plastics.	BFR and chloride-containing plastics			No/No		Up to 99%, depending on input material
	Redwave CX	NIR, metal sensor			Yes	Glass, metals	Yes/Yes
CP Group (MMS) Sorting Equipment	FlakeMax	NIR	Best suited for PET and PE/PP		Non-metals			3–16 ton/h	Up to 98%
	eMax	NIR, color, and metal always included	Designed for e-scrap recyclers	Sorting of opaque, transparent, and black commodities such as ferrous, non-ferrous, and stainless steel, wires, PCB, as well as durable plastics such as ABS, HIPS, PC, and PMMA				0.5–3.0 ton/h	Up to 98%
Eagle Vizion	Black Sorter		Sorts PE and PP Flakes	PE, PP, and others				Up to 0.55 ton/h		2–12 mm
CP Group (MMS) Sorting Equipment	L-VIS	VIS high-resolution color camera	Color sorting, flakes and pellets. Sorts PET PE, and PP flakes and pellets	Yes, electric scrap			Yes/Yes		98%	Statistics and quality control report, metal detector, remote modern or ethernet access
	E-sort	NIR	Separate different types of plastics (all resin) by composition and color	Useful for flake sorting, shredded plastics (i.e., WEEE)			Yes/Yes	Up to 3 ton/h	92–98%
MEYER Europe s.r.o.	CL-L-Sorter (Uses AI Tech)	AI cameras working in the electromagnetic spectrum: full RGB visible light, infrared standard, infrared HD, InGaAs, and UV light	Detect and remove non-PET materials flakes	PVC/PS/PC/ PA/PP/PE/ABS		Rubber/aluminum	Yes/Yes	up to 6 ton/h
Mogensen GmbH/Allgaier Process Technology GmbH	Msort	IR and X-ray	Sorts all resins of size from 0.5 mm up to 250 mm	Sorts all resins (mostly used to sort PET flakes)		Yes	Yes/Yes	Up to 4.4 tons/h. Detection of up to 25,000 particles/s	Up to 99.9%
	MikroSort AF	CCD Linear Camera	Sorts PET flakes by color				Yes/Yes	1–3 ton/h
NRT Optical Sorting	Flakesort	NIR	Mainly used to remove contaminants from PET streams					Up to 2.5 ton/h	Removal efficiency of flakes up down to 0.1 inch
Pellenc ST	Mistral + Metal Sensor	NIR	Applicable for all resins	Mostly used in shredded e-scrap sorting		Paper, cardboard, and metals/No	No/Yes	Up to 6.5 tons/h
Rhewum GmbH	Datasort	CCD camera system, LED	Sorts all resins				Yes/Yes	4.4 to 8.3 ton/h	Up to 97% accuracy
	RHEWUM DataSort S		Mostly used for ore sorting, but can be used to sort plastic flakes as well						Up to 98%
Satake	Scanmaster IE	High-resolution CCD Camera	Separates plastics by color	PET, PVC			Yes/NI	1–3 ton/h		Remote monitoring
	MikroSort AF	CCD Linear Cameras	Sorts PET flake by color				Yes/Yes	0.25–5 ton/h		Remote monitoring
	Satake RNEZX	NIR, full-color RGB camera.	Sorts PET flakes by color			Yes	Yes/Yes
	Beltuza sorter	NIR, full-color RGB	Sorts plastic flakes by color			Yes	Yes/Yes	Up to 12.5 ton/h
	FMSR-IR Sorter	Full-color RGB, infraRed	Sorts plastic flakes by color			Beans, seeds, corns, nuts	Yes/Yes
	ScanMaster XE	Proprietary inGas/Color camera technology	Removes clear PVC from PET, and other non-contaminants	Sorts all resin		Yes	No/No	Up to 3 ton/h		Remote monitoring
	RGB Full Color Belt Sorter	NIR, full-color Cameras (RGB)	Separates plastics by color	PET, twisted PVC			Yes/Yes	9 to 19 t/h	Up to 99%
	Pellet Scan	High-resolution CCD Cameras	Separates plastics by color	No					Up to 99%	Data Scan
Sesotec GmbH (S + S Separation and Sorting Technology GmbH)	Flake Purifier N	NIR	Purifies resin streams, also sorts e-plastic	PET, HDPE, PLA, PVC, and more			No/No	Up to 10 ton/h depending on how the unit is scaled	90% to 99.8% depending on input
	Flake Purifier C	CCD linear camera	Color sorting	No			Yes/Yes	Up to 10 ton/h depending on how the unit is scaled	90% to 99.8% depending on input	Dual ejection
	Varisort X	X-ray	Identifies BFR-containing plastics	Identifies BFR containing plastics			No/No	Up to 2.5 ton/h depending on how the unit is scaled		Dual ejection
TOMRA Systems ASA	Ixus	X-ray	Useful for sorting shredded e-scrap	Useful for sorting BFR- and chloride-containing plastics (i.e., PVC)			No/No	1 ton/h	Depends on product type
	Innosort Flake	NIR, Visible spectra Sensors	Used for purifying PET flakes, purifying transparent and opaque flakes, sorting of mixed color flakes	PVC, PE, PET, PP, PS, and others, including Tetra Pak and film			Yes/NI
	Autosort Flake	Flying beam, full-color camera	Sorts plastic flakes	PET, PO, PVC flakes		Yes, metal removal	Yes/NI	6 ton/h		Advanced statistics for real-time quality control
Unisensor Sensorsysteme GmbH	PowerSort 200	Ultra-high-speed laser spectroscopy	Useful for bottle-to-bottle recycling	Sorts all resins			Yes/Yes	Up to 3 ton/h	98% or higher
Visys	Spyder	Laser	Separation based on color, structure, shape, and size differences	No			Yes/Yes	1–3 ton/h	Up to 99% depending on input
	Python	Laser and cameras	Separation based on color, structure, shape, and size differences
	Tyrex	X-ray	Separation based on density of materials (i.e., plastic, WEEE, ASR)	Useful for sorting BFR and chloride-containing plastics (i.e., PVC)			No/No		Up to 99% depending on input
Wesort	6SXZ-680	AI deep learning		ABS, PC, PE, PET, PP, PPS, PPU, PVC, bottle plastic, resin, masterbatch, nylon, acrylic			Yes/Yes	1.5–2.5 tonne/h	≥99%
	6SXZ-340	AI deep learning		ABS, PC, PE, PET, PP, PPS, PPU, PVC, bottle plastic, resin, masterbatch, nylon, acrylic			Yes/Yes	0.75–1.15 tonne/h	≥99%
	6SXZ-272	AI deep learning		ABS, PC, PE, PET, PP, PPS, PPU, PVC, bottle plastic, resin, masterbatch, nylon, acrylic			Yes/Yes	0.6–1 tonne/h	≥99%	Multidimentional sorting
	6SXZ-204	AI deep learning		ABS, PC, PE, PET, PP, PPS, PPU, PVC, bottle plastic, resin, masterbatch, nylon, acrylic			Yes/Yes	0.45–0.75 tonne/h	≥99%	Dual camera
	6SXZ-272L	AI deep learning		ABS, PC, PE, PET, PP, PPS, PPU, PVC, bottle plastic, resin, masterbatch, nylon, acrylic			Yes/Yes	0.6–1 tonne/h	≥99%	Shape selection
	6SXZ-136L	AI deep learning		ABS, PC, PE, PET, PP, PPS, PPU, PVC, bottle plastic, resin, masterbatch, nylon, acrylic			Yes/Yes	0.3–0.5 tonne/h	≥99%
	6SXZ-68L	AI deep learning		Same as ebove			Yes/Yes	0.15–0.25 tonne/h	≥99%
	6SXZ-68	AI deep learning		Same as ebove			Yes/Yes	0.15–0.25 tonne/h	≥99%
	6SXZ-90	AI deep learning		Same as ebove			Yes/Yes	0.2–0.4 tonne/h	≥99%
	6SXZ-136	AI deep learning		Same as ebove			Yes/Yes	0.3–0.5 tonne/h	≥99%
	6SXZ-136L	AI deep learning		Same as ebove			Yes/Yes	0.3–0.5 tonne/h	≥99%
AnySort	VDR Series (6 Series)		PE, PET, PVC, PP				Yes/Yes			Sorts based on shape as well
PicVisa	EcoFlake X600	NIR, RGB cameras	Plastics	PET, PE	Yes, metal (i.e., copper, brass, and aluminum) and seeds		Yes/Yes	0.4–1.0 ton/h	99.5%
	EcoFlake X1200	NIR, RGB cameras	Plastics	PET, PE	Yes, metal (i.e., copper, brass, and aluminum) and seeds		Yes/Yes	0.8–2.0 ton/h	99.5%
Eagle Vizion	Micro Flake Sorter		Plastics	PE, PET, PP, and PVC	Wood, glass, paper					Sorts particles from 5 mm down to 0.5 mm

). However, sorting devices that use AI in a way to mimic the human brain to make decisions in sorting plastic waste and have a form of robotic arms or SCARA with grippers to sort plastics are classified herein as AI-based sorters or robotic sorters (

Table 6. Inventory of AI-based robotic sorters.

Manufacturer/Brand	Equipment Name	Sorting Method	Primary Application	Plastic Identified	Sorts Non-Bottle Rigids in Addition to Bottles	Non-Plastics Sorted	Colors Sorted/Black Plastic Sorted	Throughput (Average)	Plants in the US Using Equipment	Accuracy	Features
AMP Robotics	Cortex	MSW, e-waste, and construction and demilition waste			Yes	Yes	NI/Yes	60 picks per min	Alpine Waste and Recycling. Denver Co, and Minnesota	99%	Cortex is continuously learning from experience, becoming better all the time
	Cortex C	MSW, e-waste, and construction and demilition waste			Yes	Yes		65 + picks per min per arm		99%	Ideal for smaller spaces
Bulk Handling Systems (BHS)	Max-AI	Deep learning technology and the sorting process is based on the evaluation of optical data determined by VIS-sensors	Extract recyclable commodities from a specific stream of material	PET, HDPE	Yes	Yes	NI/Yes	65 picks per min	Recology, San Francisco		Continuously learning to improve efficiency
Bollegraaf Recycling Solutions	Analyzer		Used to determine material flow and composition in real time.
	RoBB-AQC		Plastics, from PET, HDPE, LDPE, PS and PP to Tetra Pak, OCC, or paper/cardboard of various shapes and sizes	Paper, cardboard, plastic, and metal containers, cartons, residue			Yes/Yes	Up to 70 picks/min per robot			Up to 4 separate sorts per robot Maximum Object Weight: 4.4 lbs. (2 kg)
BT-Wolfgang Binder GmbH (Redwave)	RedWave 2i	NIR, RGB cameras and all-metal detectors	Sorts all resins	Sorts all resins	Yes	Paper, metals, e-waste, glass, construction waste		Up to 7 ton/h			24/7 remote maintenance access for quick service and support
Machinex	SamurAI	Delta robot with vacuum gripper	Extract recyclable commodities from a specific stream of material (e.g., plastics from a reject line)	PET, colored, and natural HDPE	Yes	Yes	NI/Yes	Up to 70 picks per min	Lakeshore Recycling Systems. Forest View, IL	Up to 95%	There is ongoing evolution and optimization of AI material recognition. It continually improves and learns from operating experience to assure maximum recognition efficiency.
Bulk Handling Systems (BHS)	Max AI AQC		Removes contaminants, recovers recyclables			Yes	Yes	Up to 70 picks per minute Up to 6 separate sorts			Maximum object weight: 1 lbs
	Max AI Cobot		Can sort plastics			Yes	Yes				Designed to work safely alongside people
	Max AI Flex		Can sort plastics		Yes	Yes		Up to 35 picks per minutes per robot arm Up to three separate sorts from a single robot	Mechanical gripper, vacuum gripper		Ideal for heavy and/or non-uniform objects in a variety of pre- and post-sort applications. Able to grasp objects up to 15 lbs, including non-uniform material
OP Teknik	SELMA	Deep learning				Wood, stone, concrete, bricks, metals, cardboard, foam, etc.		Up to 10,800 picks/h with 6 robots. or 30 picks/min per robot arm
TOMRA Systems ASA	AutoSort CyBot		Packaging, beverage cartons, and all thermoplastics			Yes	Yes/Yes
Enerpat	Jet Series		Plastics	PET		Yes	Yes/Yes	Up to 8 tons/h		over 95%	Can quickly identify the color, appearance, shape, size, and even brand characteristics of the waste
PicVisa	EcoPic	NIR, RGB sensors (cameras)	Plastics	PET			Yes/Yes	1 pick per s			Maximum payload of 4 kg
Sortera Alloys	Sortera’s A.I.		Can sort plastics	Mainly used to sort metals			Yes/Yes				Pateneted tech, not commercially available yet
Everestlabs	Everestlabs RecycleOS Robotics Cell						Yes/Yes			90% + success rate	Six-axis robotic arm
	Everestlabs AI	AI (A vision system mounted on top of a conveyor to capture a 3D map of the of objects on the conveyor, and provide the data to the robot platform				Yes	Yes/Yes				Capture high-speed images of the materials in the conveyor using a self-lit, self-contained industrial 3D camera for accurate material characterization that robotics use as input
ZenRobotics	ZenRobotics fast picker		Can sort plastics			Yes	Yes/Yes	Up to 80 picks per minute		up to 99%
	ZenRobotics Heavy picker		Can sort plastics			Yes	Yes/Yes	Up to 2300 picks per hour		Up to 99%	Max. object weight: 30 kg. Up to 3 robotic arms
Waste Robotics	Integrates waste handling processes, computer vision, deep learning and robots to improve sorting efficiency		Ability to differentiate types of plastic					Up to 50 effective picks/ min			Lift up to 1 kg for fast picker
Ishitva Robotic Systems	Suka	Netra machine vision system	Can sort plastics by type, size, and color	PET, PP, HDPE			Yes/Yes	2 to 8 tons per h of Plastic Sorting
	YUTA	Netra machine vision system	Can sort plastics by type, size, and color	PET Polymer-based sorting of PET, PP, HDPE			Yes/Yes			>95% accuracy
Recycleye	Recycleye Robotics		Plastics	HDPE, PET, paper

).

Contact information of sorting equipment suppliers is reported in Appendix A

Table A1. Contact information of optical sorter suppliers.

Amut Ecotech Via San Marco 11/a 31052 Candelù—Maserada sul Piave (TV)—Italy Phone: +39-0422-877-688 Fax +39-0422-877-690 E-mail: [email protected] Website: www.amut.it/amutecotech (Last accessed 30 January 2024)	Anhui Zhongke Optic-electronic Color Sorter Machinery Co., Ltd. No. 43, Yulan Avenue, Baiyan Science Park, Hefei high tech Industrial Development Zone, China Email: [email protected] Phone: +8613655516956 Fax: 0551-66396866 Website: http://english.cn-amd.com (Last accessed 30 January 2024)
Anysort ANYSORT, Schnackenburgallee 179, 22525 Hamburg Phone: +49-40-819768-0 Email: [email protected] Website: https://www.anysort-usa.com (Last accessed 30 January 2024)	Binder + Co. Grazer Straße 19-25 A-8200 Gleisdorf, Austria Phone: +43-3112-800-0 Fax: +43-3112-800-300 Email: [email protected] www.binder-co.com (Last accessed 30 January 2024)
Bollegraaf Group Tweede Industrieweg 1 9902AM Appingedam The Netherlands Email: [email protected] Phone: +31-(0)596-65-43-33 Website: https://www.bollegraaf.com (Last accessed 30 January 2024)	Buhler Gupfenstrasse 5 Uzwil 9240 Switzerland Phone: +41-71-955-19-00 Website: https://www.buhlergroup.com (Last accessed 30 January 2024)
Cimbria Faartoftvej 22 7700, Thisted, Denmark Phone: +45-96-17-90-00 E-mail: [email protected] https://www.cimbria.com (Last accessed 30 January 2024)	CP Group (MSS) Sorting Equipment 6795 Calle de Linea San Diego, CA 92154, USA Phone: +1 619-477-3175 Fax: 619-477-3426 https://www.cpgrp.com (Last accessed 30 January 2024)
Eagle Vizion www.eaglevizion.com (Last accessed 30 January 2024)	Green Machine LLC 8300 State Route 79 Whitney Point, NY 13862, USA Phone: +1 800-639-6306 Email: [email protected] Website: www.greenmachine.com (Last accessed 30 January 2024)
Hefei Mayson Machinery Co., Ltd. Block A, Zhongrui Tech-research Building, No. 9 Hongfeng Road, Hefei City, China Email: [email protected] Phone: +86-199-5659-5855 Website: https://hfm-sorter.com (Last accessed 30 January 2024)	Hefei Golden Sorter Co., Ltd. No.230, Jinxiu Road, Economic and Technological Zone, Luan, Anhui province, China. Email: [email protected] Phone: +86-19965476623 Website: https://goldensorter.com (Last accessed 30 January 2024)
IMRO Landwehrstrasse 2, D-97215 Uffenheim, Germany Phone: +49-(0)-9848-9797-0 Fax: +49-(0)-9848-9797-97 Website: https://www.imro-maschinenbau.de/en/ (Last accessed 30 January 2024)	MachineX 2121, rue Olivier, Plessisville QC, G6L 3G9, Canada Phone: +1-877-362-3281 Website: https://www.machinexrecycling.com (Last accessed 30 January 2024)
MEYER Europe s.r.o. Nam. L. Novomeskeho 1 040 01 Kosice, Slovakia Email: [email protected] Phone: +421 948 209 976 Website: https://meyer-corp.eu (Last accessed 30 January 2024)	Mogensen GmbH/Allgaier Process Technology GmbH Ulmer Straße 75 73066 Uhingen Germany Phone: +49-7161-301-175 E-mail: [email protected] https://www.allgaier-process-technology.com/en (Last accessed 30 January 2024)
MSS, Inc. [A division of CP Group] 300 Oceanside Drive Nashville, TN 37204, USA Phone: +1 615-781-2669 Email: [email protected] https://www.mssoptical.com (Last accessed 30 January 2024)	MSWsorting Zhengzhou high-tech zone, China Email: [email protected] Website: https://www.mswsorting.com/index.html (Last accessed 30 January 2024)
NRT Optical Sorting 1508 Elm Hill Pike Nashville, TN 37210, USA Phone: +1-615-734-6400 Email: [email protected] www.nrtsorters.com (Last accessed 30 January 2024)	Pellenc ST 125 rue François Gernelle BP124 84 124 Pertuis Cedex 4 Phone: +33-4-90-09-47-90 Email: [email protected] www.pellencst.com (Last accessed 30 January 2024)
PicVisa Isaac Newton, 2 Barcelona, Spain Email: [email protected] Phone: +34-938-268-822 Website: www.picvisa.com (Last accessed 30 January 2024)	Redwave (a division of BT-Wolfgang Binder GmbH) Wolfgang Binder Str. 4 8200 Eggersdorf bei Graz, Austria Phone: +43-3117-25152-2200 Fax: +43-3117-25152-2204 Email: [email protected] https://redwave.com/en/ (Last accessed 30 January 2024)
Rhewum GmbH Rosentalstrasse 24 42899 Remscheid, Germany Phone: +1-(888)-474-3986 Email: [email protected] Website: https://www.rhewum.com/en (Last accessed 30 January 2024)	RTT Steinert GmbH 1234 Hardt Circle Bartlett, IL 60103, USA Phone: +49-221-49840 Email: [email protected] Website: https://steinertglobal.com (Last accessed 30 January 2024)
Satake 10900 Cash Road Stafford, Texas 77477 USA Phone: +1-281-276-3600 Website: https://satake-usa.com (Last accessed 30 January 2024)	Sesotec GmbH (S + S Separation and Sorting Technology GmbH) Regener Strabe 130 D-94513 Schonberg, Germany Phone: +1-224-208-1900 Fax: +1-224-208-1909 Email: [email protected] www.sesotec.com (Last accessed 30 January 2024)
Steiner US 285 Shorland Drive KY 41094 Walton Phone: +1-(859)-962-2648 Website: https://steinertglobal.com/us/ (Last accessed 30 January 2024)	TOMRA Systems ASA Drengsrudhagen 2 Asker 1385 Norway Phone: +47-66-79-91-00 https://www.tomra.com/en (Last accessed 30 January 2024)
Unisensor Sensorsysteme GmbH Am Sandfeld 11 76149 Karlsruhe, Germany Phone: +49-(721)-97884-0 Email: [email protected] Website: www.unisensor.de/en/ (Last accessed 30 January 2024)	Visys Birlik Sanayi Sitesi 2. Cadde No:97 PK:34520 Beylikdüzü—İstanbul—Turkey Phone: +90-212-876-90-36 Fax: +90-212-876-90-37 E-mail: [email protected] Website: www.visys.com.tr (Last accessed 30 January 2024)
Wesort Building 29 LongWangMiao industrial area, BaiShiXia Community, FuYong Street, Shenzhen, China Phone: +86-13226817096 Email: [email protected] Website: https://www.wesortcolorsorters.com (Last accessed 30 January 2024)

and Appendix B

Table A2. Contact information of suppliers of AI-based robots and sorters.

AMP Robotics 1500 Cherry Street, Suite A Louisville, CO 80027, USA Phone: +1 (888)-402-1686 Website: www.amprobotics.com (Last accessed 30 January 2024)	Back Handling Systems (BHS) 3592 West 5th Avenue Eugene, OR 97402, USA Phone: +1 541-485-0999 Email: [email protected] Website: https://www.bulkhandlingsystems.com (Last accessed 30 January 2024)
BIN-e Pasjonatów 9 62-069 Dąbrowa, Poland Email: [email protected] Website: https://www.bine.world (Last accessed 30 January 2024)	Bollegraaf Recycling Solutions Tweede Industrieweg 1, 9902 AM Appingedam, The Netherlands Phone: +31-596-654-333 Email: [email protected] Website: https://www.bollegraaf.com (Last accessed 30 January 2024)
CleanRobotics Email: [email protected] Website: https://cleanrobotics.com (Last accessed 30 January 2024)	Enerpat Enerpat Group Uk Ltd. 55 Crown St, Brentwood, Essex CM14 4BD, UK Email: [email protected] Phone: +86-15051237913 Fax: +86-513-8778-2755 Website: https://www.enerpatrecycling.com (Last accessed 30 January 2024)
Everestlabs 48820 Kato Rd Suite 500B, Fremont, CA 94538, USA Email: [email protected] Website: https://www.everestlabs.ai (Last accessed 30 January 2024)	Greyparrot Greyparrot AI Ltd. 100 Drummond Road A401 London, SE16 4DG, UK Email: [email protected] Website: https://www.greyparrot.ai (Last accessed 30 January 2024)
Intuitive AI 1200-555 W Hastings St, Vancouver, BC V6B4N6, Canada Email: [email protected] Website: https://intuitiveai.ca (Last accessed 30 January 2024)	Ishitva Robotic Systems Website: https://ishitva.in (Last accessed 30 January 2024)
Machinex 2121, rue Olivier, Plessisville QC G6L 3G9, Canada Phone: +1-(819)-362-3281 Website: www.machinexrecycling.com (Last accessed 30 January 2024)	OP teknik Lastbilsvägen 2 298 32 Tollarp Sweden Phone: +46-(0)-10-456-82-70 Email: [email protected] Website: https://www.opteknik.se/sorteringssida?lang=en (Last accessed 30 January 2024)
PicVisa Isaac Newton, 2 Barcelona, Spain Email: [email protected] Phone: +34-938-268-822 Website: www.picvisa.com (Last accessed 30 January 2024)	Recycleye 179 Hercules Road, London SE1 7LD, UK Email: [email protected] Website: https://recycleye.com (Last accessed 30 January 2024)
Redwave (a division of BT-Wolfgang Binder GmbH) Wolfgang Binder Str. 4 8200 Eggersdorf bei Graz, Austria Phone: +43-3117-25152-2200 Fax: +43-3117-25152 2204 Email: [email protected] https://redwave.com/en/ http://www.btw-binder.com/en/ (Last accessed 30 January 2024)	Sortera Alloys 49 S 500 E Markle, IN 46770, USA Phone: +1 260-330-7100 Website: https://www.sorteratechnologies.com (Last accessed 30 January 2024)
TOMRA Systems ASA Drengsrudhagen 2 Asker 1385 Norway Phone: +47-66-79-91-00 https://www.tomra.com/en (Last accessed 30 January 2024)	Waste Robotics 3055, rue Tebbutt Trois-Rivières, QC G9A 5E1 Canada Phone: +1-819-201-2525 Website: https://wasterobotic.com (Last accessed 30 January 2024)
ZenRobotics Perintötie 8 C 1 01510 VANTAA Finland Email: [email protected] Phone: +358-50-4363803 Website: https://www.terex.com/zenrobotics/ (Last accessed 30 January 2024)

. The companies listed are based in North America, Europe, and Asia. Our search was conducted in the English language; hence, it may not have captured companies in, e.g., China. The producers (and countries) of optical sorting equipment are summarized in Figure 8.

The information collected here reveals the progress made during the two years since our previous report on plastics sorting equipment was published. This information is further used to assess whether the various challenges reported by MRFs in our previous study [20] can be addressed by currently available technologies or emerging technologies.

6.2. Sorting Equipment for Post-Consumer Plastics

This section describes commercial sorting equipment for sorting mixed plastics. The optical sorters considered here utilize NIR, VIS (light- or camera-based), and XRF. Reported technologies are classified here based on criteria such as plastic identification method (e.g., NIR or XRF), primary application, throughput, whether they sort plastics by color and/or by size, accuracy, and additional features (Table 2).

A total of 37 conventional optical sorting machines produced by 16 different companies have been reported. Out of these, 22 sorters possess the ability to classify plastic based on its color, and 18 among them can effectively separate black plastics from other colors through the utilization of both NIR and VIS technologies. Additionally, a total of 13 optical sorting machines integrated with AI, by 11 different companies, have been identified. Among these, 10 machines are capable of sorting plastic by color, and 8 of them have the capability to separate black plastics from other colors by utilizing ML/AI with NIR, and/or VIS technologies. This brings the total number of whole-plastic optical sorters to 50 (conventional optical sorters and optical sorters with integrated AI). Since our 2022 report [20], there has been an 8.7% growth in the number of optical sorters. Moreover, there are now 13 AI-integrated sorters that were not available in our 2022 study (Table 3). The reported accuracy of these sorters in reclaiming materials can reach an impressive 99.99%, contingent upon the input materials being processed. Furthermore, these sorters offer a broad range of throughput capacities, with the capability to handle up to 10 tons per hour.

Although conventional and AI-based optical sorters have high sorting efficiency, they are primarily intended for sorting 3D/rigid plastic items, and are not as efficient when it comes to sorting plastic films and other two-dimensional materials. There is technology available specifically designed for sorting plastic films or 2D plastics. A total of 16 film-sorting machines by 9 different companies have been identified (Table 4). This marks a substantial 60% increase compared to our 2022 report [20]. These film sorters are said to achieve an accuracy rate of 98%, depending on the materials being processed. Out of the total, six of these machines have the capability to sort films based on their color, while the majority of them (81% of sorters) employ a combination of NIR and VIS technologies.

Before plastic is reprocessed, flake sorting is a crucial stage that helps minimize contamination caused by foreign materials or undesired plastics that may have slipped through previous sorting stages. Flake sorters have the capability to segregate plastics based on their size, typically down to 1 mm, although this may vary depending on the specific types of equipment used. Table 5 presents the existing inventory of plastic flake sorters. A total of 55 flake sorters, produced by 21 different companies, are reported. This represents an increase of 57%, compared to the 35 sorters identified in our 2022 publication. The reported accuracy achieved by flake sorters can reach 99%, contingent upon the materials being processed. The identification of plastic types is accomplished through the use of NIR, XRF, and/or VIS technologies. Additionally, out of the 55 flake sorters, 36 possess the capability to sort flakes based on their color.

The variation in the number of all plastic sorters from our 2021 report to this study is presented in Figure 9.

6.3. AI-Based Robotic Sorting of Plastics

In order to tackle challenges encountered in the field of waste management, new approaches are being developed based on the use of computers and robotic technologies [17,112]. Sorting robots, guided by AI, can either operate as an alternative to traditional optical sorters or can supplement optical sorters by purging incorrectly sorted plastics at the end of the sorting process [17]. Moreover, AI sorters have the ability to improve sorting efficiency over time by using available data to mimic a human brain’s learning and decision-making processes [85,113,114].

A total of 22 AI-based robotic sorters are reported here from 16 different companies, all with the ability to sort plastic by type and color (Table 6). Some MRFs have already integrated AI-based sorters in their processing line, according to manufacturers of AI-based sorters that report in their publicity materials lists of MRFs that have adopted their technologies. The plastic identification method or sorting method involves deep learning and VIS, and in some cases, combine deep learning, VIS, and NIR. There is an increase of 214% in AI-based sorters compared to the seven AI-sorters identified in our 2022 report.

7. Conclusions

The increasingly large amounts of plastic produced and used worldwide necessitate the proper management of plastic waste, which includes various types of recycling processes. Proper classification and sorting of plastic waste are important in recycling, as they increase the quantity and improve the quality and value of the post-consumer plastics that are recovered and sent for reprocessing. Automated sorting promises high throughput and efficient classification of mixed plastic waste. The types and capabilities of commercially available equipment for sorting plastic are reported and analyzed here. To support the content on automated sorting equipment, spectroscopic methods for identification of plastic type are highlighted, and basic principles of ML/AI are presented, with an interest in the combination of spectroscopy and ML/AI, which is at the core of modern equipment employed in plastic type identification.

The inventory of commercial sorting equipment includes 49 optical sorters for whole plastic objects, 55 flake sorters (from 11 companies), 16 film sorters (from 9 companies), and 22 AI-based sorters for mixed plastic recyclables. The recovery accuracy of sorting equipment reported herein can be as high as 99.99%, depending on input materials, with a wide range of throughput capacities (up to 10 ton/h). Growth in available optical sorting technology was about 7%, and a significant shift in incorporating AI in optical sorters was observed, compared to our 2022 study [20]. The observed growth in film sorters, flake sorters, and AI-based sorters signifies the emerging importance of plastics recycling. The potential accuracy of sorting equipment, from past reports and currently available, remains similar when it comes to sorting plastics by chemical composition; however, the introduction of AI has given sorting equipment new capabilities it previously did not have, which is sorting plastics based on physical characteristics such as transparency, morphology, etc., which in turn reduces misclassification of plastics based on physical attributes, thus significantly boosting recycling efficiency. The information presented here can address some of the challenges reported by MFRs in our 2022 article, with an increasing number of available film sorters that can now be integrated in MRFs to sort plastic films for recycling, and an increasing number of sorters for black plastics that could not otherwise be sorted solely by NIR-based optical sorters. Although the available technologies to address those limitations are increasing, economic factors should also be taken into account.