1. Introduction
Cooking serves to sterilize and remove harmful components and promotes nutrient absorption. It is an indispensable aspect of daily life [
1]. Whether in domestic or commercial settings, cooking is concentrated in kitchens, making it a fundamental and ubiquitous element of daily life. However, the development of cooking equipment and methods has significantly increased the difficulty and threshold for cooking. This has led to a situation in which young people lack the motivation to learn, whereas elderly individuals frequently encounter operational challenges. Meanwhile, in the era of Industry 4.0, everything is interconnected, and a fast pace has become the main theme of the world. Amid mounting work pressure and diminishing leisure time, the prepared meal and delivery sectors have experienced substantial growth by capitalizing on the demand for convenience. This trend enables consumers to access diverse cuisines without the time investment traditionally required for home cooking [
2]. An urgent need exists to update technology in the kitchen sector in terms of convenience and interactivity. The prepared food and takeaway sectors are more prone to food safety and quality issues. Notably, most consumers still prefer to consume dishes cooked in the kitchen if it is made easier and quicker [
3]. Therefore, the advancement of smart kitchens has effectively addressed contemporary consumer demand and is aligned with the characteristics of the modern era. It has broad market prospects and development potential.
Artificial intelligence (AI) enables robots to generate human-like responses under advanced computer development conditions. Recently, with the updating of software and hardware, the perception, decision-making, execution, interaction, and other AI capabilities have made transformative leaps. AI can understand the external world using physical and chemical sensors, which improve the efficiency and accuracy of perception through optimized algorithms [
4]. Decision-making is a key goal of AI in the core areas of autonomous agents and complex scenario interactions. Currently, numerous in-depth studies have focused on how to make machine decision-making processes exhibit the core characteristics of human decision-making, including the adaptive processing of fuzzy information, a trade-off mechanism based on shared social norms and human ethical values, and the ability to flexibly adjust to dynamic environments [
5]. Stone et al. [
6] investigated the ability of AI in strategic decision-making, achieving full coverage in key domains such as the military, economy, and education. This demonstrates that AI decision-making can withstand tests of time and facts. Simultaneously, the emergence of AI has significantly strengthened the connections between people and machines. Machines are no longer confined to performing rigid mechanical operations but are increasingly capable of providing responsive feedback and even emotional engagement to users. For instance, the modality of interaction and the context of use exert considerable influence on user satisfaction and perceived usability within the television viewing experience [
7]. This connection can provide more emotional value to home experiences and may represent a critical trajectory for future innovation [
8].
AI technology has been employed to conduct research related to modern smart kitchen and to develop smart cooking equipment. This has led to great progress in traditional cooking processes including doneness control, food preservation, and other aspects. However, significant limitations still persist, leading to widespread skepticism regarding AI-empowered modern smart kitchen. In order to strengthen the connection between people, machines, and foods, and promote the deep integration of AI technology and modern smart kitchen, there is an urgent need to summarize the development status and possible problems of this field. However, existing smart kitchen equipment has demonstrated notable advancements in cooking convenience, efficiency, environmental protection, and energy conservation [
9], [
10]. However, devices from different manufacturers lack interoperability and most remain in the conceptual or early prototyping phase [
11]. Most smart kitchen machines have yet to undergo rigorous market testing, and pertinent national regulations remain largely undefined. Consequently, there is persistent societal skepticism regarding the deployment of robotic systems in domestic kitchen environments, particularly during cooking activities [
12]. Therefore, when improving the cooking quality and convenience of modern smart kitchen equipment, it is necessary to consider its safety and stability, and fully utilize the perception and decision-making capabilities of AI. Furthermore, engineers must refine human-computer interaction design to optimize user experience and foster a sense of satisfaction. This will also facilitate greater recognition and acceptance of smart kitchen [
13].
Thus far, the cooking systems of the new era have established the infrastructure shown in
Fig. 1. People, machines, and food are interconnected in virtuous cycles. In traditional cooking systems, only a connection exists between people and food. People rely on food to survive and reproduce, and food provides them with a profound sense of belonging and emotional fulfillment [
14]. During the Industrial Revolution, a connection was formed between machines and food. Machines have dramatically increased the efficiency of food production and facilitated advancements in food ingredient processing. Moreover, the characteristics of these ingredients can guide and constrain the development of machines [
15]. In the existing research, the two types of relationships mentioned above have been reported most frequently, and the studies are more detailed. However, according to traditional perception, machines tend to be cold and rigid. They operate by executing predetermined linear instructions and lack autonomous cognitive capabilities for adaptive responses [
16]. This runs counter to the rich spiritual needs of the people and leads to a long-term lack of connection between people and machines. The emergence of AI has effectively filled this gap and formed a complete closed-loop cooking system. This represents a transformative advancement in the field.
The existing literature on smart kitchen reviews remains remarkably limited and primarily concentrates on narrowly defined thematic domains, such as specific technologies or individual device categories. For instance, Liegeard and Manning [
17] provided a comprehensive overview of the role of smart kitchen equipment in reducing food waste and proposed the introduction of emerging technologies by reviewing the evolution of date labels and related packaging information. Pereira et al. [
18] proposed a two-level taxonomy to guide the operational framework of robots and facilitate advancements in autonomous kitchen robots. Their work classified the possible execution of smart kitchen devices into 16 categories. However, their suggestions remained at the theoretical stage and did not involve the specific development of kitchen robots. Liu et al. [
19] summarized the cutting and cooking processes of ingredients in central kitchens with a focus on fundamental three-dimensional (3D) segmentation techniques. However, these techniques are outdated and lack industrial feasibility in the context of rapidly advancing AI. Given the accelerated progress and broad potential of AI, the absence of systematic reviews in this domain may lead to fragmented research and increased learning costs. Consequently, a comprehensive and systematic review of existing literature is required.
To methodically examine the transformative impact of AI-enabled smart kitchens, as shown in
Fig. 2, this review analyzes innovations in smart kitchen equipment through three fundamental development components: hardware, information transmission, and software. Based on this, mainstream smart kitchen equipment is used as an example to summarize the integration methods and development status of smart kitchen technology. Building on the extant research and critical analysis, this study examines the challenges and potential solutions in smart kitchen development across multiple dimensions, including interoperability among merchants and consumer acceptance, along with other factors. These insights aim to bridge the existing gaps between technological capabilities and practical implementation. This review aims to assist researchers in understanding the composition and module grading of modern smart kitchens and facilitate the in-depth development and integration of AI technology. The reduction in the threshold and the proposal of new research directions can collectively promote the development of AI, enabling smart kitchens and realizing the logical closed loop of the three elements of the cooking system. Specific research methods and literature quality evaluation standards are provided in the Table S1 in Appendix A.
2. AI-enabling smart kitchen hardware
2.1. Information acquisition hardware
2.1.1. Spectrometer
Spectrometers are classified according to their light emission sources and wavelengths. These techniques include visible spectroscopy, near-infrared spectroscopy, far-infrared spectroscopy, Raman spectroscopy, and Fourier spectroscopy. A spectrometer can provide a more comprehensive analysis of the material in question. However, the data obtained from spectrometers are often extensive and repetitive because of the measurement of multiple wavelengths. The basic knowledge threshold for statistical analysis is also high. Thus, its development has been limited. The high arithmetic power and procedural features of AI have effectively improved this situation, offering a straightforward and efficacious instrument for spectral information processing [
20]. Therefore, the spectrometer has had an outstanding application performance in the field of food testing and is widely used in food picking, screening, processing, exhaust gas treatment, and other links throughout the entire process [
21]. In addition, it promotes a connection between machines and food and is widely used in smart kitchen information collection hardware. However, spectrometers are often impractical because of the size of their equipment. This problem persists in smart kitchen devices. The development and maturity of small spectrometers may help address this problem. The development of AI has improved the construction accuracy of spectrometers and helped the size of spectrometers reach the chip level [
22]. Mayr et al. [
23] used a handheld near-infrared spectrometer to construct a database and determine the caffeine content in food products. The accuracy and stability were not significantly different between the handheld and the large spectrometers. As seen above, the AI-enabling spectrometer does not have to be plagued by the problems of redundant information that is difficult to analyze, and has achieved good results through increased wavelength coverage and information collection capabilities. Therefore, as an important source of information acquisition, the spectrometer is used in the development of smart kitchen equipment. Even so, despite the support of AI technology, mastery of appropriate band screening and operation methods remains a challenge that needs to be faced. Developers should choose the appropriate spectrometer and detection band according to the task. This is a key aspect that must be considered to perfectly integrate AI into smart kitchen equipment.
2.1.2. Camera
Cameras are the most important information acquisition devices for machine vision. The characteristics of the cameras vary depending on the specific model. These include capabilities such as depth and volume detection, heat detection, fluorescent material detection, and other features used for a given task or application based on specific requirements and environmental conditions. The use of cameras to capture sample image data and perform model recognition and judgment has been widely covered in the entire process of the food field, which is recognized as an excellent non-destructive and rapid measurement method [
24]. In addition, AI models can process large amounts of data within a short period. Considering these developments, new types of equipment have been designed, including high-resolution cameras, arrayed 3D stereoscopic cameras, and binocular cameras capable of acquiring multidimensional information in a single acquisition period. This has led to improvements in the accuracy and adaptability of modeling judgments [
25], [
26].
Even if AI technology enhances the application value of cameras in the field of smart kitchens, the variety of camera styles places higher requirements on researchers. Researchers must fully consider the type of camera, setting position, and other parameters for a specific task. For instance, to reduce cost and facilitate operation, Yang et al. [
27] applied a mobile phone infrared camera to detect the storage temperature of agricultural products and obtained a higher image quality. However, most hyperspectral cameras are large and expensive. Multispectral cameras have been used to replace aflatoxin detection tasks, achieving a processing performance of 20 data sets per second at a lower cost [
28]. Cameras have been used in the development of intelligent devices since the advent of machine vision. When focusing on the new revolution in AI empowerment, we should learn from previous research on camera usage and deployment.
2.1.3. Microphone
Traditional computers are limited to more rigid linear calculations, whereas the communication language often needs to consider more nonlinear influences. Therefore, the development of traditional smart kitchens is limited to the connection between machines and food, which is not conducive to an increase in the market acceptance of smart kitchens [
29]. AI technology has driven a revolution in smart kitchens. Microphones are used to receive voice data from both consumers and food sources. Consumers are only required to open their mouths to speak, which significantly enhances their perception of food-related applications. For instance, Pan [
30] realized a far-field voice interaction system by setting up microphone arrays for intelligent robots to recognize sound information from far away. Furthermore, instances exist where applications utilize auditory data to inform dish assessments. For instance, an in-ear microphone has been developed to capture the sounds of people chewing. The program could determine the crunchiness of chewed food with an accuracy of 95% and could also effectively obtain the moisture content of the food [
31].
The concrete information of sound is its amplitude, frequency, and wavelength. It is difficult to understand its relationship with abstract information such as language semantics and food quality [
32]. Consequently, the information obtained by a microphone is extremely difficult to process and understand, and the accuracy rate of relevant studies is mostly below 80% [
33]. This implies that smart kitchen devices require better-quality microphones and quiet environments. This presents a challenge for the use of hardware such as a microphone.
2.1.4. Multimodal information acquisition hardware
The multimodal judgment model developed by AI enables each field to screen relevant sensors suited to their specific objectives such that multidimensional information can be used to improve accuracy [
34]. In the context of food, sensors include weight, temperature, and humidity sensors. The sensors are frequently unable to ascertain the specific nature of a food product on their own. However, when used in combination, they can have a significant multiplier effect. The current applications of a diverse array of sensors have been extended to the domain of modern smart kitchens. The various types of sensors and their applications are listed in
Table 1 [
35], [
36], [
37], [
38], [
39], [
40].
However, a reasonable selection of multimode sensors is often the focus of experimental design.
Fig. 3 shows the calculation and fusion methods for each mode of information under AI empowerment. The figure and table systematically organize the development mode and status of smart kitchen equipment hardware from two perspectives: methods and examples.
2.2. Information processing hardware
2.2.1. Central processing unit (CPU)
The CPU, as the foundation of a computer, is the final unit of the information processing program operation of all computing devices, and comprises a hosting area, logic operation area, and control area. Each area plays its own role in completing the computational commands issued by the computer. Consequently, the CPU has a powerful logical calculation ability and significantly reduces the possibility of errors. Thus, it is widely used in the development of intelligent devices [
41]. The development of integrated-circuit technology has promoted the miniaturization of the CPU. Thus far, the existing process of manufacturing the CPU can integrate more than ten billion transistors within a space of 50 mm × 50 mm × 5 mm, which can easily meet the needs of current smart device programs. Therefore, the application of CPU is characterized by miniaturization, low power consumption, stability, and ease of control [
42]. However, the current use of CPU in smart kitchen equipment still focuses on primary integrated circuits represented by microcontrollers. For instance, a smart kitchen temperature measurement and control system uses ESP32 as the CPU of the entire system [
43]. Arduino Mega 2560 was applied to a smart kitchen cooking system for the elderly, significantly simplifying the complex traditional cooking process and adding safety devices such as emergency termination [
44]. In the context of the booming development of AI technology, simple microcontrollers cannot satisfy the computational requirements of AI models. A revolution in the use of CPU in smart kitchens is imperative. Currently, an industry-recognized methodology utilizes an external high-computing-power CPU for model output and judgment with the help of information transmission. Lin et al. [
45] utilized a cell phone CPU to quickly identify meat freshness with the assistance of smart fluorescent labels to avoid consuming expired and spoiled food and food waste. In addition to their deployment on smart mobile devices, various large-scale AI smart devices developed in recent years have powerful CPU and better computing environments. The integrated technology of information and decision making has matured [
46]. Therefore, it is not necessary to deploy a CPU individually for every kitchen appliance.
2.2.2. Graphics processing unit (GPU)
The GPU has been developed specifically for processing image data and is different from the CPU in terms of architecture, calculation method, and design purpose. The device contains a considerable number of transistors that are well-suited for data stream processing and auxiliary functions, such as image storage and computational sequencing. This configuration confers a significant advantage for tasks that require a substantial number of parallel computations, such as image processing [
47]. The introduction of AI has led to a large number of parallel nonlinear computations in the training of smart kitchen devices. However, the deployment of GPU in existing smart kitchen devices is rare, and most existing devices use basic CPU cores. Existing devices are unable to meet AI and consumer demands. Consequently, determining a path for GPU deployment in modern smart kitchen devices is a critical step in this field. The successful deployment of GPU in other areas can provide us with lessons. Currently, mature applications exist for GPU deployment on existing smart devices that focus on high-end image AI areas, such as smart driving and drone detection. To develop a more reliable and high-level auto-drive system, a GPU was used as the main computing device with the aid of a field-programmable gate array. This has the potential to inform and enhance GPU deployment approaches in other areas [
48]. In the field of trajectory planning for drones and vehicles, a comparative analysis of the computational models on CPU and GPU revealed that the GPU, with its advanced AI parallel computing capabilities, could achieve a processing speed over 100 times faster than that of the CPU, thereby significantly reducing the computational resources required [
49].
Through the above analysis, it can be seen that the advantages of the GPU in improving AI computing speed and conserving resources have been proven. This also provides the basis for the application of the GPU in the field of smart kitchens. Researchers should be able to effectively address the challenges encountered in the development of smart kitchen equipment through a reasonable migration strategy. In light of these considerations, it is inevitable that as modern smart kitchen equipment continues to evolve, the CPU and GPU must collaborate in their respective roles to facilitate the horizontal and vertical expansion of the model, so as to reinforce the connections between machines and people.
2.3. Command execution hardware
Currently, two types of command execution hardware are available in most smart kitchens. One is the modification of basic kitchen equipment such as smart microwave ovens and smart refrigerators. The other is the new AI products in the field of smart kitchens, represented by various humanoid mechanical devices, such as cooking robots. With the help of AI technology, the first type of execution hardware is primarily improved in terms of information perception and process control. This can help complete tasks based on traditional cooking methods. He et al. [
50] used an intelligent microwave oven with integrated infrared and humidity sensors for thawing control and solved the problem of excessive dish heating during microwave thawing. The second category of hardware aims to replace people with AI. They require the ability to perceive, think, and make decisions. Thus, they are not merely the backbone of hardware; they are a complex whole with various sensors. Khan et al. [
51] developed an automated fruit-clamping robot arm comprising aluminum and mild steel to form the main skeleton. The device was equipped with a depth-sensing camera and ultrasonic sensors, enabling the detection of the object depth and significantly enhancing the success rate to exceed 90%. In addition, a robotic arm with five rotational axes was developed, four of which were designed to emulate the human arms, shoulders, elbows, and wrists. The final axis of rotation controlled the movement of the object that was being held. The arm is capable of controlling errors to within ±1° with satisfactory reproducibility, particularly at low speeds during the cooking process [
52]. However, in current industrial and service sectors, the most mature and widely deployed robotic arm is the six-axis robot. Under the premise of cost control, the stability, compatibility, and accuracy have all reached high levels [
53]. Furthermore, its resemblance to human operations facilitates communication and interaction between people and machines. According to Styler et al. [
54], consumers are more interested in interactive assistive operating systems when they replace cooking processes. This could save labor and provide better emotional feedback, which should be considered seriously in the subsequent development of equipment.
Although the second type of hardware is the focus of research in the smart kitchen field amid the AI revolution, the two types of hardware are not superior to each other. Bionic hardware still requires traditional hardware to complete basic kitchen operations. The optimization of traditional execution hardware can also improve the effectiveness of kitchen tasks. The AI-enabling revolution should be a comprehensive one, combining the old and the new.
3. Smart kitchen information transmission
3.1. Internal transmission of equipment
Traditional information transmission in smart kitchens relies on simple and easy-to-understand physical tags. These have matured and are widely used in the food industry. Traditional kitchen front- and back-end information, such as seasoning functions, source of ingredients, and freshness of dishes, is mainly transmitted in the form of text tags. The introduction of AI technology has effectively addressed the problem of machine recognition of text and strengthened communication between machines and food [
55]. However, textual information has drawbacks such as limited expression, inability to update in real time, and susceptibility to damage during transportation. With the improvement in living standards, the requirements of people for food have shifted from the taste of the dishes themselves to the quality and nutritional combinations of ingredients. Text information is gradually being unable to meet consumers’ refined living needs. Therefore, various forms of physical information transmission media such as color, shape, and quick response (QR) codes have been developed.
Fig. 4(a) summarizes the application of QR codes in the food supply chain, which has the advantages of high information content and fast module loading [
56]. Consequently, QR codes in food science have excellent performance in quality evaluation, anti-counterfeiting, and traceability, thereby helping food blockchain technology flourish. Lin et al. [
45] constructed an identification matrix reflecting the freshness of beef based on nanofluorescent clusters and introduced AI technology to correspond the matrix with freshness information. By combining QR codes with invisible photochromic ink, the smart tag constructed by Popović et al. [
57] completely lost its color after 5-6 months of storage. Although the QR codes provide data, they also convey information on the storage conditions and time of the wine.
This type of information transmission is organically combined with new technologies in various fields. It simplifies the intermediate form of information and achieves high-capacity and high-precision information transmission with less cost, thereby improving the economy and convenience of information transmission. From this, it is clear that traditional text-based communication is no longer adequate to meet the needs of modern smart kitchen. Collaborating with other information transmission and condensation systems is imperative to convey information through entities such as shapes, colors, or information transmission systems such as QR codes and barcodes.
3.2. Home local area networks (LANs)
LANs connect computers and other devices within a small area through a transmission medium. It has the advantages of fast transmission speed and high security. In modern smart kitchen equipment, most smart kitchen devices use wireless LANs as the transmission medium for home LANs. The Internet of Things (IoT) is the latest developmental direction for wireless LANs in the era of Industry 4.0. The demand for smart kitchen integration, intelligence, and convenience has spawned an IoT infrastructure with different complexities. An integrated multimodal model that pursues judgment accuracy requires complex and diverse equipment including a spectrometer and microphone [
58]. For fast and convenient edging computing, a single camera can meet the task requirements [
59]. In general, AI enabling IoT facilities discard dangerous chemicals and complex gas sensors. This is a simplified process. Commonly used IoT media include wireless fidelity Wi-Fi, Zigbee, and Bluetooth. Wi-Fi is the most widely used wireless network technology, and with the introduction of the Institute of Electrical and Electronics Engineers (IEEE) 802.11 standard, Wi-Fi 6 was born, which provided users with guaranteed latency and reliability. Consequently, Wi-Fi is becoming a natural candidate for information transfer in the modern smart kitchen. While maintaining the throughput of a single device constant, devices using Wi-Fi 6 protocol had almost no increase in working latency that was significantly better than that of Wi-Fi 5 protocol. This protocol guarantees efficiency and reliability for home area network users [
60]. Zigbee technology does not have the same transmission range and speed as Wi-Fi; however, its lower power consumption often allows it to be integrated into embedded smart devices. Data transmission with a high information capacity is the main feature of Zigbee technology. An effective collision avoidance mechanism can prevent the generation of unstable transmissions, thereby ensuring the secure transmission of data. Based on this mechanism, a smart home system exhibits enhanced performance [
61]. The Bluetooth transmission distance is shorter, and the transmission rate and power are in the middle of those of Wi-Fi and Zigbee. This makes it suitable for smart kitchen close-range device connection and control. It belongs to a class of LANs and does not have an expanded space with an Internet connection. Therefore, it has better stability and security [
62].
The contemporary smart kitchen LANs incorporate three principal information transmission methods: Wi-Fi, Zigbee, and Bluetooth. Each method possesses distinctive characteristics that are well-suited to varying transmission distances, transmission rates, and transmission environments, in alignment with the specific requirements of smart kitchen equipment. In light of the aforementioned considerations, developers are well-positioned to select the most suitable IoT-enabled local area information transmission method, as outlined in this review. This section discusses LANs in the context of IoT, focusing on how home LANs are integrated into a smart home environment. Importantly, these LANs are not directly connected to the broader IoT via a router, which enhances security and stability. However, they are also unable to communicate with the outside world, limiting data access capabilities.
3.3. Internet
Driven by AI, smart kitchen equipment is required to strengthen the connections between machines and people. The need for increasing human-machine interaction functions and self-thinking abilities has increased [
63]. Consequently, such devices are limited by the use of local area data and the need to access the Internet to realize real-time updating and information learning. Modern smart kitchen devices are further developed based on the IoT by connecting the entire kitchen to the Internet through a router with IoT protocols. Most current smart homes use the more secure and faster transmission speed Internet protocol version 6 (IPv6) protocol, which has achieved better results. However, new IEEE 802.15.4 standard devices face obstacles when interconnecting with traditional devices [
64]. For instance, Aguilar et al. [
65] proposed the first adaptation layer design for bidirectional cross-technology communication using the wake-up radio (WuR-CTC) of the IPv6 protocol, the structure of which is shown in
Fig. 4(b). This design can transmit 127-byte IPv6 packets from two different protocol devices within 69 ms without a gateway, which further improves the information transfer rate and reduces resource consumption. In terms of practical applications, the overall system architecture, topology, and Internet protocol (IP) address-allocation scheme are relatively clear. This system can realize intelligent door and temperature control for a smart home based on a simulation platform [
66]. Considering these developments, new Internet transport protocols have overcome the existing constraints. Wozniak et al. [
67] developed a series of home Internet infrastructures based on the sixth-generation mobile communication technology (6G) network standard, and constructed a new IoT implementation system.
Fig. 4(c) shows that the system enabled interconnections between different homes and communities, which significantly improved the integration of information and efficiency of event processing. This approach serves to safeguard the information security of smart kitchen devices, thereby facilitating the secure proliferation of the Internet in modern smart kitchen [
68].
The integration of modern smart kitchens with the Internet is consistent with the prevailing trend of the accelerated advancement and adoption of AI technology. The existing Internet introduction rules and methods are well established, with the two main protocols being IEEE 802.15.4 and IEEE 802.11. Developers can make informed choices based on their specific needs. Concurrently, with the aim of enhancing the security and availability of modern smart kitchens, several studies have been conducted on the compatibility of dual protocols [
69], Internet access systems, and system security assurance [
70].
4. AI-enabling smart kitchen software
Hardware provides the foundation for communication between people, machines, and foods. Meanwhile, software serves as an integral component, facilitating the seamless integration and optimal performance of the modern smart kitchen system. Existing AI models have made significant advancements. However, the application of models with different characteristics is still in its infancy, and researchers are not yet fully aware of the characteristics and potential applications of each model. Accordingly, this section classifies the existing AI software according to different model architectures so as to elucidate their intrinsic characteristics and potential applications in smart kitchen. Furthermore, the current state of research and prospects for development in the field of AI are presented, aiming to serve as a reference for researchers in this domain and provide insights.
4.1. Machine learning (ML)
Traditional ML involves artificial neural networks, support vector machines, and decision trees, and these methods mainly rely on displaying programming laws to set up programs for computers based on human thoughts. Therefore, they offer the advantages of clear thinking, high interpretability, fast computation, and fast application [
71]. However, the ML model has a single input data form and application scenario, which requires complex data preprocessing. Visualization of the output information is not diverse. Data cleaning, feature transformation, feature dimensionality reduction, and region-of-interest segmentation are common data preprocessing methods that can significantly improve model performance and are widely used in various tasks [
72]. After the concept of the smart kitchen was introduced, machine vision was first utilized in the AI-enabling evolution of smart kitchens to facilitate communication between food, machines, and people. Currently, it is widely used for food quality detection [
73], kitchen waste utilization [
74], kitchen gas detection [
75], and other modern smart kitchen functions. The ability of ML to realize multimodal information fusion operates under these different functional applications. Yamaguchi et al. [
76] collected 604 features from 8183 recipes and used six different machine vision models to fuse and classify the features to achieve the ability to predict the flavor and taste of a dish based on the seasoning in the recipe.
Accordingly, good results have been achieved in the development of model improvements and new software calculation methods that have been applied to smart kitchens. Wang and Wang [
77] proposed a stacked supervised autoencoder combined with a kernel extreme learning machine algorithm for nonlinear and deep feature extraction with better feature extraction and higher accuracy and efficiency in determining fatty acid content. The simulation results indicated that this improvement yielded better results. Subsequently, the ML model was applied to the overall development and evaluation of smart kitchens. For instance, a comparatively optimal random forest was integrated into the Unity engine for automatic generation and presentation of 3D kitchens. A perceptual study confirmed that this program could assist in the construction of a smart kitchen system and the training of students in related design disciplines [
78].
While ML has been utilized in various aspects of smart kitchen, the advent of AI technology, coupled with the surge in demand for smart kitchen equipment and the growing complexity of data obtained by sensors, poses more challenges to the advancement of novel equipment. It is imperative that software should be developed with enhanced capabilities in feature extraction, accelerated computation, and optimized user interface. The advent of deep learning models serves to bridge the aforementioned gap, thereby facilitating the more profound and faster evolution of the contemporary smart kitchen.
4.2. Deep learning (DL)
Unlike ML, DL involves a deeper network architecture associated with self-learning. This allows the elimination of feature extraction tasks from machine vision tasks [
79]. The existing DL database contains images, text, sounds, and other modal information [
80]. The diverse original data expands the visualization form of the model reasoning results. In addition, the attention mechanism, multiscale feature fusion, and image enhancement module in DL are effective in solving noise problems such as object occlusion and background noise in traditional ML [
81], [
82]. Consequently, the recognition accuracy of DL is approximately 10% higher than that of ML models [
83]. However, this is followed by a complex model architecture and various application scenarios. The lack of a systematic study on this topic will lead to contradictory results. Therefore, this section introduces the architecture, training methods, and performance of mainstream DL models, and compares models with different improved types. We hope to promote this revolution in the field of smart kitchens objectively and dialectically.
4.2.1. Recurrent neural networks (RNNs)
RNNs are a type of neural network suitable for processing sequential data. Their hidden layer design allows upper-level information to be retained such that contextual information can be considered. To prevent a gradient explosion and retain important information, a gating mechanism was proposed to selectively retain information. With continuous progress in the field of computing, variants such as long short-term memory networks (LSTMs) and gated recurrent units have aided RNNs in wide applications. In AI technology, RNNs are primarily used for behavioral prediction, language recognition, and quality detection [
84]. However, the feature extraction ability of RNNs is weak, and commonly used datasets are mainly text [
85]. Most of the information provided in a kitchen is complex, such as color and shape, which may limit its application in smart kitchens. The preprocessing of imbalanced data or the use of logical decision trees may be an effective means of enhancing the performance of RNN models, facilitating their development in the context of AI-driven technologies [
86].
Modern smart kitchens have more stringent requirements for equipment anthropomorphization and convenience. These requirements necessitate the reasonable prediction of smart kitchen actions that fit the ability of the RNN. Qi et al. [
87] developed an umami-multi-resolution neural network (MRNN) for fresh-flavor peptide prediction. The model merges multilayer RNN to predict flavor-presenting substances in dishes. This type of end-to-end prediction function is the most basic application of RNNs. However, AI technology has more stringent requirements. The smart kitchen revolution has driven deep information mining for obtaining predictive results. A home management server model was built based on the prediction of household electricity consumption using LSTM. This realized a reasonable distribution of the electricity consumption of different electrical appliances according to the predicted electricity consumption [
88]. Simultaneously, to address the problem of the weak feature extraction ability of RNN model data, Nandhini et al. [
89] introduced convolutional neural networks (CNNs) to the RNN model, endowing it with image recognition ability. The average accuracy of pest detection reached 93.6%, which was better than that of the single CNN or RNN models. RNNs exhibit excellent performance in speech recognition and semantic analyses. Speech recognition accuracy based on the LSTM reached 94%, and data processing could be completed in as little as 4.5 s [
90]. The bidirectional encoder representations from transformers (BERT), improved from LSTM, is a current mainstream software program for the semantic recognition of smart devices. For instance, the software inferred personal information and behavioral habits from the display name and user description in the profile and provided appropriate feedback [
91]. With the aid of appropriate datasets or feature extraction modules, RNNs have unique advantages during the AI revolution. They have been applied in modern smart kitchen behavior prediction, language prediction, and quality detection prediction. Its lightweight characteristics make it suitable for embedded software development, which may be a powerful tool for promoting kitchen intelligence [
92].
4.2.2. CNN
CNNs rely on their unique convolution operations to dominate two-dimensional (2D) image recognition; however, in recent years, they have become a hotspot and a difficult point of research in the field of computers. With in-depth research in the field of traditional image recognition, its recognition ability has surpassed that of experts. Simultaneously, its horizontal research is booming, and it integrates various ML models to complement each other, bringing the recognition ability of image information to various fields.
As well-researched DL models, CNNs have constantly developed variants suitable for different tasks, lowering the research threshold. Therefore, CNNs have been favored by researchers in the field of food because of their simple operation and excellent results. The implementation of a prediction project includes three steps: establishing a dataset, selecting and improving the model, and training the model. Two mainstream methods for establishing datasets exist: one uses file names as labels, and the other uses specialized image annotation tools such as LabelImg and LabelMe. This type of tool provides a more precise annotation method for polygons [
93]. In terms of model selection, CNNs have improved versions suitable for various scenarios. Transfer learning (TL) can be applied to smart kitchens, such as DenseNet121 for eggshell crack images [
94], AlexNet for tea quality identification [
95], and deep adaptation networks for geographical origins of crops [
96]. In addition, the selection and adjustment of hyperparameters are crucial in the model-training process, because they determine the final results of the model. Automatic parameter-tuning methods can reduce the workload, which has become a current research hotspot. The current automatic hyperparameter tuning model was developed from random parameter trial-and-error to iterative parameter trial-and-error, which saves time and computational resources [
97]. However, we cannot blindly trust such automatic parameter tuning systems. Users must have a basic understanding of DL and the ability to determine the rationality of the hyperparameters. The application of the CNN model is an important part of the revolution of AI technology in smart kitchens. Therefore, we focused on its construction process and method. Simultaneously, this process can be used as reference for other DL models.
CNNs are favored by researchers because of their simple operation and excellent results. Image-recognition functions are utilized in the entire process and field of modern smart kitchens. CNNs are widely used in food quality screening and monitoring [
98], cooking process monitoring, dish preservation, and classification and disposal of kitchen waste [
99]. In addition, modern research in the field of smart kitchens has focused on strengthening the connection between food, machines, and people to provide a greater sense of access and satisfaction for consumers. Chiu and Chuang [
100] added a CNN module to a chatbot to help identify the content, taste, flavor, and other characteristics of the dishes offered by the chef, thereby helping the bot with sales pitches and dish recommendations. Furthermore, by deploying the MTx-Net lightweight CNN architecture on a smartphone, consumers can identify whether the beef or lamb they purchase is adulterated or is of high quality by simply snapping a photo on their phone [
101].
CNNs, the most widely used DL model, have been integrated into various fields to handle complex 2D information and obtain unexpected results. For instance, linkage with spectral information provides a solution to the spectral redundancy of information, which is mutually beneficial. Kang et al. [
102] fitted the CNN model to a handheld Raman spectrometer and detected pathogenic microorganisms and accurately determined bacterial species. In addition to the multilevel and multidirectional applications of the model, the security of the CNN was also tested. In the recognition process, Sousa and Bassani [
103] used GoogLeNet to create a comprehensive representation of the visual features acquired within the area covered by each convolutional kernel, thereby effectively reducing the loss of features in unrecognized areas. This helped the model improve its performance by 1%-3% on different datasets. Concurrently, the CNN feature area fast selection function enables local computing power to satisfy demand without Internet transmission, thereby markedly enhancing the security and privacy of the system [
104]. In addition, CNNs can address the problems of object occlusion and complex backgrounds in image recognition, which expand the adaptive scene and application of DL [
105].
The current application of CNN in modern smart kitchen features comprehensive coverage with a certain depth and extensibility. This aspect is the most worthwhile reference element in the selection process for smart kitchen software. However, it is necessary to have a certain programming foundation and fully consider all factors according to the steps for establishing the CNN model.
4.2.3. Generative network
Generative networks can generate analogous samples through learning and data summarization. This functionality makes the network an invaluable tool in the realms of human-computer interaction and database extension, both of which are integral to modern smart kitchens.
Data training often requires numerous samples, but most databases in smart kitchens comprise samples with large variability or that are difficult to obtain, such as the nature of dishes and character actions. Generative adversarial networks (GANs) can solve this problem, both quantitatively and qualitatively. Zhu et al. [
106] used a data augmentation method combining numerical simulation and GAN to augment samples in the process of traditional Chinese medicine milling and found that the similarity between the augmented and unaugmented data was close to 100%. Most of the data in the smart kitchen database are susceptible to operator state images and are often imperfect; therefore, goal-aware generative adversarial imitation learning was developed. The role of GAN in data augmentation has been demonstrated and recognized in various domain reviews [
107], [
108]. It introduces a second discriminator whose role is to provide guidance to the first discriminator, which is responsible for recognizing the content of the database and discarding data biases caused by instability [
109]. Another role of generative networks is exemplified in human-computer interactions. Language is a large and varied system, and most databases used for training have certain limitations. In this context, a GAN can be used to extend linguistic data and obtain better results for a simple subject-verb-object structure [
110]. Furthermore, the most difficult aspect of machine language analysis is understanding human emotions, as the same text can represent different meanings in different contexts. Yang et al. [
111] integrated gate recursive units with residual networks in a GAN to generate multistyle image sentiment captions and validated the effectiveness of mainstream datasets. Furthermore, the robust predictive capabilities of GAN have been recognized in the domains of personalized shopping experiences and enhanced multisensory immersive customer experiences, thereby fostering brand trust and user stickiness [
112].
The implementation of GAN in contemporary smart kitchen is still in its infancy, but its semantic analysis and database expansion capabilities are well-suited to the AI-enabling smart kitchen revolution. This can serve as a point of reference for developers, allowing them to reduce the time required for data collection and enhance the interaction capabilities of their models.
4.2.4. Transformer
Transformer is a revolutionary architectural innovation in the field of computers that has emerged in recent years. It has revolutionized traditional DL architectures and employs an attention mechanism to model input sequences [
113]. Its most significant advantage is in natural language processing (NLP). This helps in the development of generative pre-trained transformer 4 (GPT-4), BERT, and other language-processing models that exceed human cognitive abilities. In modern smart kitchens with high interaction capabilities, the introduction of the Transformer is an optimal solution. Additionally, the Transformer is an effective method for feature extraction; therefore, it is common for Transformer to be integrated with other models in the preprocessing stage.
Currently, Transformer has few applications in smart kitchens, but it is one of the most cutting-edge directions for cross-disciplinary research in the field of food. Existing food research has focused on its feature extraction capability. Zheng et al. [
114] used Transformer to extract features and imported them into a multilayer perceptron to recognize strawberry quality. Compared with the traditional Transformer and CNN models, the recognition accuracy increased by more than 3%. In addition, it can handle data that is difficult to handle using traditional methods. The characterization of sesame oil samples by fluorescence spectroscopy enabled the acquisition of 3D images of fluorescence spectra. The remaining data processing techniques were significantly inferior in terms of both the quality of the results and the speed of processing compared with the Transformer, which reached 0.99 for both accuracy and recall [
115]. In the context of modern smart kitchens, the connection among food, machines, and people has been emphasized. The NLP potential of the Transformer requires the focused attention of developers. Mujahid et al. [
116] provided a BERT model with three dense neural network layers for sentiment analysis, with a superior accuracy of 96.49%. In addition to existing models, the textual semantic analysis of Transformer was integrated with visual image analysis in the same semantic space, and contextual actions were generated based on dependencies. This enhanced the linkage and efficiency of various modules of the device and improved the overall accuracy by 4.1% [
117]. However, owing to the large size of the Transformer, most researchers are concerned about its slow response rate in integrated devices. Accordingly, a streaming converter was developed to enable the association of connectionist temporal classification with the Transformer. The model accuracy was optimized while maintaining the corresponding rate, resulting in a 20% improvement in the combined accuracy with comparable latency [
87].
As the latest deep learning model, The Transformer demonstrates superior performance in feature extraction and the NLP field. It is expected to be one of the key development directions of smart kitchen software. Nevertheless, the considerable parameter volume and relatively slow processing speed of the model may present challenges in terms of integration. Even if the model is compressed, its application in the smart kitchen remains challenging. Cloud deployment and the development of integrated chips may serve as viable solutions to this problem. Additionally, it is essential to select an enhanced version of the Transformer model that aligns with the current pace of advancement in the computing field and strikes a balance between response time and accuracy.
4.3. Knowledge graph (KG)
KG is a core concept in network theory that describes the relationship between entities.
Fig. 5(a) illustrates the structural features of representing entities as nodes and the relationships between entities as edges. The construction of a KG is analogous to the process of establishing a database with the primary objective of extracting and integrating requisite information. It is a prevalent tool in the domains of NLP and AI, streamlining information retrieval and acquisition, and enhancing the precision and responsiveness of information-based models. With the help of DL, the construction speed and scale of the KG have significantly improved. The construction process is automated and convenient [
118].
Currently, the most important applications of KG in smart kitchens include semantic analyses and recipe recommendations. Diab and Demiris [
119] constructed a KG with different operations as entities and similarities between different entities as edges. This helped the robot refer to the executed commands while recognizing commands and determining the exact semantics of the new input commands based on possible continuity. In addition, it is used as an auxiliary function to facilitate the effective operation of smart kitchen systems. With the help of the KG, Has et al. [
120] enriched the ability of the robot to find an action that must be performed, allowing it to cancel a previously received command. This occurred more often in the kitchen scenario, where the technique was validated. In addition, compared with traditional regression-planning networks, motion-planning algorithms that use KG as an auxiliary data source exhibit more comprehensive capabilities in kitchen cooking and table settings without significant operational bias [
121]. KG requires a large amount of data to improve its effectiveness, and its complex construction process restricts its development [
122]. Previous KGs required the developer to fill in a predefined template according to the sample properties. As the earliest form of construction, it suffered from the problems of slow data filling and easy limitation of information [
123]. Therefore, ML techniques have been introduced to realize the rapid and comprehensive construction of a KG. The model automatically learns important connections between samples, dramatically improving construction efficiency and data comprehensiveness [
124]. With the efforts of researchers, the existing KG developed a self-learning function. A new protocol for the KG was utilized to construct display rules to check problems in the sensors, which enabled the system to grow by absorbing the problems into the tacit knowledge of the KG while posing them [
125]. The excellent capabilities of KG for information summarization, rapid querying, prediction, and recommendation are in line with the requirements of modern smart kitchens, which have enormous potential for development [
126].
Therefore, the advantages of the KG fit well with those of smart kitchens. The problem of database construction was also solved, indicating that it deserves more focus and in-depth research.
4.4. Digital twin (DT)
DT technology has emerged because of the exponential growth in the development of computer systems. As shown in
Fig. 5(b), the basic concept of the DT is to build a set of computer software programs capable of simulating a real-world system. This process comprises initial model scanning, model building, and aggregation. This enables the monitoring and prediction of real-world environments. Modern smart kitchens require consumers to comprehend and regulate the status of the kitchen in real time. However, this demand is frequently affected by the surrounding environment and hardware conditions. The DT provides another safeguard for this requirement. Its integration can reduce reliance on costly and complex sensors and their associated analysis software. However, its construction is complex and requires relevant knowledge of physics and engineering, which hinders its application in smart kitchens.
Currently, the DT has appeared in the fields of smart kitchens and homes, and can realize the coverage and simulation of most processes. Padmapriya and Srivenkatesh [
127] constructed a DL framework using DT, which aided in constructing DT systems for home states and possible threat predictions. In addition, consumers are placing a higher demand on DT under the AI revolution. Therefore, large volumes of data and rigorous logical reasoning are required. Consequently, the existing DT can be constructed using various types of DL networks, simulation software, and other stable models that have been validated by the industry. For crepes, the DT for the entire cooking process was constructed using a neural network containing more than 400 000 analog data points. In the final model, the difference between the crepe state and its true state at each cooking period was less than 5%, demonstrating the stability and predictive power of the DT [
128]. Sa et al. [
129] constructed the movement trajectory and distribution of grease and smoke during the cooking process based on the 2D of Echarts and the 3D of Unreal Engine 5 (UE5) flow field methods. This helped consumers understand the current kitchen environment and helped developers build smoke exhaust channels.
The development of current DT systems can be approximately divided into three parts: the physical entity, the virtual entity, and the connection between them. Most DT systems are built based on mathematical formulas and rigorous logic with DL and virtual simulation software as aids, which require a high level of mathematical skill from developers [
130]. Therefore, few studies have been conducted on the DT in smart kitchens. In this study, the establishment and connection criteria of the hardware and software in a smart kitchen are clearly determined, which provides the basis for the development of the DT to seamlessly integrate multisensor data and interconnect between components. This can enable the DT to assist in the development of modern smart kitchens more effectively.
4.5. Supplement of AI-enabling smart kitchen software
ML has been extensively integrated into smart homes and kitchen systems. Based on the applications listed in this paper, numerous fusion models or improved models that further enhance the development of modern smart kitchen software exist.
Table 2 ([
124], [
131], [
132], [
133], [
134], [
135], [
136], [
137], [
138]) lists the representative examples of other model applications. Researchers can refer to the table and their own equipment requirements to reasonably select the best model. Relevant research usually evaluates models in terms of accuracy, F1 score, model size, and reasoning speed. For instance, tasks based on mobile devices need to prioritize model size and computing speed [
51], whereas tasks seeking quality need to prioritize accuracy and F1 score [
139].
5. AI-enabling smart kitchen equipment
5.1. Smart refrigerator
Smart refrigerators are rapidly developing, comprehensive, and extensively researched components in modern smart kitchens. Consequently, the field of smart refrigerators has witnessed significant technological advancements and emergence of a wide range of applications. This section presents an overview of the existing technology integration methods and the equipment development status of smart refrigerators, aiming to provide insights and guidance for the advancement of modern smart kitchen equipment in the context of AI.
Detecting the type and freshness of content with the help of images is an important development direction and was the main feature of early smart refrigerators. Accordingly, a basic recognition unit for smart refrigerators was developed. This unit, containing a receptor, a processor, and a detector, enables recognition and analysis of the basic environment. This functionality is now more mature and has been verified in numerous cases [
140]. In addition, the introduction of AI technology accords smart refrigerators more possibilities and presents higher demands for human-machine interactions. Smart refrigerators need to establish a bridge between machines and people while undertaking the task of communicating with food and machines and have the ability of self-thinking and self-adaptation. For instance, the smart refrigerator developed by Chen et al. [
141] was based on an RNN network that can detect food-quality information while recording the types and quantities of purchased dishes. A refrigerator can automatically calculate the types and quantities of dishes to be ordered when there is a shortage of food. In addition, some researchers have also integrated the recipe recommendation function in a smart refrigerator, which can recommend the most suitable dishes according to the existing ingredients and consumer demand and provide the required ingredients and cooking methods for the dishes, which enhances the efficiency of the family cooking process.
The current generation of smart refrigerators has reached a relatively mature stage in terms of functionality and equipment. Nevertheless, it would be prudent for developers to diverge from the domain of smart manufacturing and revert to certain technologies within the food industry. This approach could facilitate the streamlining of specific processes while developing smart kitchen equipment, thereby capitalizing on the existing trajectory. For instance, nondestructive testing technologies in the food sector, such as colorimetric sensing array (CSA) labelling, have been introduced to alleviate the pressure on model development and equipment integration, providing new ideas for the development of smart kitchen equipment [
142].
Fig. 6(a) illustrates a fluorescent label array for pathogenic microorganism detection in a smart refrigerator, where the rapid output of judgment results can be achieved using a machine vision model in a smartphone.
5.2. Smart cooking integrated stove
Compared with smart refrigerators, smart integrated stoves require more interconnected devices and have a later development time, less related research, and relatively lower technological maturity. Nevertheless, the act of cooking is ultimately responsible for the output of a modern smart kitchen system. A smart integrated stove is an indispensable component of a smart kitchen. Accordingly, this section presents an overview of existing smart integrated cookers, with a particular focus on identifying the research outcomes and implementation strategies that can potentially accelerate the evolution of this device.
A smart integrated stove comprises several devices that require open flames. Therefore, the first concern in modern smart kitchens is improving safety. By leveraging the highly interconnected nature of the IoT and the strong information-processing capabilities of DL technology, a smoke detection device is embedded in the range hood. When the amount and type of smoke detected are abnormal during CPU processing, the stove immediately issues a power-off command to cut off the energy supply and prevent further fire expansion [
143]. Sensitive risk detection capabilities not only avert risk but also avoid excessive security redundancy. Using electromagnetic waves, the system established by Tamaki and Premachandra [
144] detected whether a flame and cooking staff were present simultaneously. The stove did not stop heating immediately when unsupervised but sent out a powerful blast to the user and then stopped heating if the user did not return to the hob within a preset period of time. In addition, smart integrated cookers have been studied for environmental monitoring. As shown in
Fig. 6(b), low-cost air pollutant sensors were integrated into a range hood to construct a small smart kitchen system including a range hood, air purifier, and kitchen exhaust fan, which reduced particulate matter with an aerodynamic diameter ≤ 2.5 μm (PM2.5) concentration by 80% at a cost of 5.6 USD [
145]. In summary, the introduction of DL provided image information for the models, which effectively improved the sensitivity of the cooker to anomalies, and expanded the monitoring range of the models.
In conclusion, the development of smart integrated stoves does not appear to be a disruptive change in device architecture, with existing research resting on the federation of multiple-device or single-device improvements. The introduction of AI technology enhances the need for human-computer interaction, which is not reflected in existing development examples. Developers can provide voice assistants to the system to enable the voice control of the device. Modern integrated stoves provide consumers with more assistance and information during cooking. Smart integrated stoves in modern smart kitchens should fully consider the cooking experience of consumers, meet the needs of the traditional cooking process, and utilize AI and food technologies as tools for designing and realizing functions.
5.3. Smart cooking robot
In the context of smart kitchen systems, the design of other equipment types is informed by the established forms and functions of traditional kitchen equipment. By contrast, cooking robots are novel additions to kitchen systems and represent an emerging product category enabled by AI. Concurrently, they are significant and regarded as the most pivotal execution unit in modern smart kitchen systems, serving to supplement human conduct in culinary pursuits within kitchens. Consequently, most resources and energy are allocated to smart kitchens. Currently, cooking robots can be classified into two main categories: stir-fry and robotic arm robots. Numerous stir-fry robots primarily comprise stir-fry buckets that are incongruous with the smart and personified demands of the AI-enabling revolution. Robotic arms have emerged as a solution to this problem. This section focuses on the essential technologies and implementation techniques associated with robotic arms.
Cooking robots inherit the flexibility of various robots, and the kitchen utensils they use are visually similar to traditional kitchen utensils. The use of DL has the potential to enhance the accuracy and reduce the cost of equipment required for image recognition tasks. Furthermore, DL models acquire and process information, and the cooking robot executes the requisite operations in accordance with the CPU execution command [
146]. However, the environment undergoes real-time changes during the cooking process, and the commands issued by the CPU have a certain degree of delay, which results in suboptimal completion of the cooking task. The recognition and movement-planning functions of robotic arms have been the subject of several studies and deployment in cooking robots.
Fig. 6(c) shows the principles and effects of using a mixed-integer programming method to coordinate and assign cooking tasks. It generates optimal motion trajectories of the robotic arms based on the characteristics of the number and correlation of robotic arms. This technique avoids the problems of collision and interference between two robotic arms [
147]. In addition, Azurmendi et al. [
148] investigated the recognition efficiency of the inception module in a cooking robot and found that it could recognize and classify the different cooking states of the robot based on simple images; they then used this classification to plan the next steps in operations such as robot movement routes. Empowered by AI, cooking robots are not regarded solely as operational execution units. In modern smart cooking kitchens, a cooking robot processor is used as the central CPU to control the cooking process. Therefore, it is necessary to focus on the design and guidance abilities of the entire cooking process of a cooking robot. Based on existing cooking robots on the market, a control model with autonomous recognition capability was developed using a set of processes and criteria for determining the merits of online recipes. The model can also realize recipe recommendations and robot motion path planning [
149].
Fig. 6(d) illustrates how the model split the instruction “refrigerate the apples” into four basic actions and gave the order in which the actions are executed. This result significantly improved the analysis and learning ability of the cooking robot control system for web-based recipe information [
150].
Furthermore, cooking robots can integrate human-machine interaction, nutritional analysis, maturity judgment, and other functions with IoT devices, according to the needs of the consumer. This integration allows for a reasonable combination of functions based on the condition of the device, ensuring a satisfactory consumer experience while avoiding design redundancy and increased costs.
6. Challenges and prospects of AI-enabling smart kitchen
Following a comprehensive analysis of the components of modern smart kitchen equipment and associated equipment development, it became evident that the existing development is still in its infancy. There is a dearth of technological innovations, equipment development, and standards. The system is not yet sufficiently systematic, and relevant research is somewhat blind to certain factors and limited in scope. Moreover, the prevailing consumer concept of a smart kitchen remains relatively conservative. There is still some skepticism about using AI in the kitchen to prepare daily meals. These issues pose significant challenges to the advancement of modern smart kitchens. However, smart kitchens meet people’s needs and adapt to the rhythm of the times. Against the backdrop of frequent problems in the takeaway and pre-prepared dish industry, modern kitchens offer a cooking option that is fast and of high quality. They have significant market and developmental potential. Therefore, as shown in
Fig. 7, this section summarizes the existing problems of smart kitchens in terms of targeted technology, equipment industrialization, interoperability challenges, and AI bias and risk.
6.1. Targeted technology
Powered by AI technology, modern smart kitchens require greater convenience and interactivity. For this reason, a major technological innovation has occurred in which computer software represented by DL has flourished and better meets the development needs of related appliances. However, the development of existing software is still stuck in the computer field, and the required computing power far exceeds the capabilities of current smart kitchen appliances. Programs are also becoming more complex, and the threshold for developers is increasing, leading to the application of AI technology by existing smart kitchen device developers that focus on TL and lack model adaptation in the food domain. The development of targeted technologies based on smart kitchen requirements should be the basis for promoting virtuous cycles in this field. Most current research focuses on the stages before ingredients are cooked or after they are fully cooked [
151], whereas research on the process from raw to cooked is rarely reported. Therefore, the development of modern smart kitchen equipment software should focus on the applicability of the developed technology in the field. In addition, to reduce the hardware requirements of the model and improve the practicality of the device, developers can use distillation, pruning, and other lightweight operations to improve the model performance, thereby increasing model depth and volume [
152]. During device development, it is important to avoid blindly relying on TL and to adapt the model to specific application scenarios during the application process. For instance, during the cooking process, water mist and the shape of the cooking dish should be considered when developing doneness assessment software. Consequently, a water mist removal module or context attention mechanism should be added to the model. On this basis, the process from raw to ripe, model pruning, and the development of adaptive technology may be the points that require emphasis in targeted technological innovation.
6.2. Equipment industrialization
The development of modern smart kitchen appliances is still in its early stages. Most of the appliances appear in the form of patents or creative ideas. Therefore, minimal focus has been placed on the implementation of related products and their economic benefits. To achieve the scale and commercialization of modern smart kitchens, developers must consider the cost of the developed equipment and establish an automated supply chain. For instance, developers have improved software and adopted more advanced software to compensate for the problem of insufficient information, thereby reducing the use of sensors to lower costs and improve efficiency, ultimately reducing the cost of smart refrigerators to below 300 USD [
153]. Apart from equipment design, a complete supply chain and reasonable adjustment strategy are top priorities for equipment industrialization and cost reduction. A positive association may exist between the generative AI usage depth and supply chain performance [
154]. Accurate supply and demand adjustments play an important role in streamlining costs and improving efficiency. Empowered by AI, the amount of data and computational needs of most software programs have increased geometrically. Devices commonly used in smart kitchens, such as single-chip CPUs and single-nature sensors, can no longer satisfy emerging software requirements [
155]. In addition to more advanced devices, external devices such as smartphones and smart robots can also be considered for assistance [
156]. This will also avoid the duplication of settings in some devices to reduce costs. With improvements in living standards and privacy awareness, safety has become an important indicator for equipment to be approved for industrialization. Both software and hardware can be used to improve the security of devices, such as the establishment of a firewall between a smart kitchen and the Internet using a router, and develop special sampling methods [
157], automatic encoders [
158], and other aspects of information encryption.
6.3. Interoperability challenge
Research on smart kitchens has not established uniform standards, leading to poor product interoperability between manufacturers. The integration of various elements to complete a smart kitchen ecosystem successfully is an obstacle to industrial development.
IoT is a representative AI technology for smart connectivity and smart homes. It can connect devices with common potentials and needs into a cohesive system, thus realizing collaborative work to improve efficiency and bring happiness and interaction to people. However, this aspect is poorly realized in the field of modern homes [
159]. Existing appliances have been designed individually by researchers, and there is a significant lack of underlying logic in the design of different appliances. In addition, there is a lack of standardization in the development of similar appliances, and developers often ignore the potential problems of subsequent applications when retrofitting existing appliances according to their innovations. To address this problem, we can use the more mature Internet protocol IEEE 802.11, which is a common standard for wireless LANs today and is recognized by various industries. In addition, researchers could reach a consensus on the basic framework of the devices. The inclusion of innovation points can be used to develop new features and improve device performance without altering the underlying logic, which contributes to the standardization and connectivity of devices. Currently, the hotel industry for smart homes has a more mature LAN construction program, device interconnection architecture, and intelligent room control assistance [
160]. The establishment of a unified standard for smart kitchen appliances could be used as a reference. On this basis, it is necessary for researchers in the same field to collaborate to establish higher-level and more instructive national standards, and even world standards. This can promote standardized operations and sustainable development of the modern smart kitchen industry. Simultaneously, ethical industry standards represented by moral management and social pressure must be established. White supply chain management, launched by Suksanchananun et al. [
161], effectively improved the transparency of enterprises, achieved industry standardization, and promoted a healthy and virtuous cycle. Although ethical standards are not mandatory, they are important for promoting sustainable development in the smart kitchen industry.
6.4. AI bias and risk
Modern research in the field of smart kitchens is practical and is empowered by AI technology to realize effective communication between food, machines, and people. We can infer that the feelings of consumers are a top priority in the new development direction of smart kitchens. Therefore, positive consumer opinions are the cornerstone of development in this field. Smart kitchen equipment is closely related to human food products. With the improvement in living standards, the demand for food safety and food quality is also increasing. The importance of diet and AI ethical concerns have been combined to trigger concerns about smart kitchens. Consequently, objects made by machines are assumed to destroy sentimental value and carry negative emotions associated with machine contact [
162]. Risk communication from the government or media can reduce the perceived benefits and risk tolerance of smart devices [
163]. In addition, cultural values have been shown to be significantly correlated with the consumption of high value-added foods, such as organic foods [
164]. Therefore, the development of smart kitchens should focus on humanistic care and reasonably guide consumers to change their perceptions and accept smart kitchens. Meanwhile, consumer trust in devices and food quality has also been proven to be key to sustainable consumption behavior [
165]. Usability, flexibility, and reliability are the most important factors affecting user-perceived value [
166]. Only when consumers verify that the device can solve practical problems can they begin to change their mindset. On this basis, the user interaction experience should be strengthened to enhance awareness of device control, and a good interaction experience can increase consumer acceptance [
167]. Changing consumer attitudes should be accompanied by appropriate promotional strategies based on equipment quality excellence. In the early stages, when functional parameters and actual effects are unknown, increasing consumer experience through promotional strategies, such as actual experience quotas or the establishment of demonstration kitchens, is far superior to advertising placement or price reduction [
168]. In brief, to ensure the quality of one's own equipment, adapting development and promotion according to consumer needs and psychological laws is the key to changing consumer attitudes.
7. Conclusions
This review provides a comprehensive overview of the existing research and the latest technology applicable to the development of modern smart kitchens. It is structured around three fundamental components of device development and expands upon research on mainstream devices. It addresses four key aspects: hardware, information transmission, software, and existing equipment. The findings and limitations of this study are discussed in each section. This study provides valuable references for future research in this field. It also analyzes the challenges of smart kitchens and proposes solutions from four aspects: targeted technology, equipment industrialization, interoperability challenges, and AI bias and risks. This review addresses the lack of a systematic review in the field of smart kitchens and provides developers with complete development processes, complex foundational knowledge, cutting-edge advancements, and emerging research directions. In the development process, developers can focus on human-computer interaction, IoT, food safety, application of new technologies, and other research directions. Under the influence of the continuous development of AI technology, an accelerated social pace, and mounting work pressures, smart kitchens will inevitably change cooking habits and traditional concepts, providing convenience to consumers in a practical way. In addition, they offer consumers a sense of psychological reassurance and satisfaction during use, demonstrating significant growth potential and substantial market demand. The development of this field is expected to strengthen the connections between people and machines, thereby forming a closed loop between people, machines, and food. This relationship forms a virtuous cycle, thereby enriching modern cooking systems. However, smart kitchens are in the early stages of development, with limited existing research and minimal real-world implementation. Therefore, this study adopts a broad thematic scope, aimed at constructing a framework of smart kitchen research for readers and performing dialectical analysis within each module. Its technical depth and content integrity have limitations. However, it is believed that the reform of AI-enabling smart kitchens is in line with the trends of the times, and further research and achievements will be realized. Relevant issues will then be discussed and criticized in more detail.
PDF
(4215KB)
