Proactive care architecture for care robots

Abstract

Care robots have been developed to address the shortage of human caregivers in aging societies facing declining birth rates. However, currently care robots only passively perform care tasks based on commands, they are unable to provide care services to bedridden patients with communication disorder in individual homes. This paper first proposes the concept of “proactive care”, which expects robots to act as a human caregiver even in a complex and changing environment according to the real-time state of the care recipient, such as proactively determine and complete the optimal care program without any commands. Then, a proactive care architecture (PCA) is proposed, which in turn gives a proactive care model (PCM) that can proactively generate the desires of care recipients. Finally, the feasibility of PCA and the validity of PCM were verified through simulation. PCA opens up an avenue for robots to care for bedridden patients with communication disorder, which will contribute to alleviating the social problem of human caregivers shortage.

Introduction

There is a worldwide trend of declining birth rates and aging populations, particularly in regions such as Japan, China, and Europe^1,2. The shortage of caregivers to look after bedridden patients at home has become a serious social problem^3,4. If robots can be developed to care for bedridden patients as human caregivers do, the shortage of human caregivers will be alleviated^5,6. Currently, although many types of care robots have been developed in the world^7,8, most of them need commands to provide care services. However, bedridden patients with communication disorder are often unable to give explicit commands or any commands at all. The robot may provide incorrect care services due to inaccurate commands or not provide care services at all. In addition, relying solely on the care recipient’s willingness may ignore their actual physical state and fail to ensure scientific care. Specifically, care robots differ from human caregivers in three essential ways:

• (1) They cannot perform care tasks without explicit commands from the care recipient or human caregiver.

• (2) They lack the ability to generate care tasks based on abstract desires(e.g., hunger and thirst) on their own.

• (3) They are unable to handle unpredictable situations independently while performing care tasks.

These limitations highlight the absence of three critical functions in current care robots: task executing without commands, proactively generating abstract desires and reasonably decompose them into several specific care tasks, and the ability to adapt to changing situations during the care delivery. This paper summarizes and defines the three functions as “ proactive care” , aiming to develop a robot that can perform care tasks proactively. The primary contributions of this work are as follows:

• (a) To enhance the intelligence of care robots and enable proactive functions, a proactive care architecture (PCA) is proposed as a comprehensive framework.

• (b) To implement this architecture, we develop a proactive care model (PCM), which anticipates and responds to the evolving desires of care recipients over time.

• (c) By leveraging PCA and PCM, care robots can alleviate caregiver shortages by achieving proactive care and improving the overall quality of care.

Related work

Care robot

In caregiving domains, in order to improve the health status and quality of life of bedridden patients, at least three basic needs of the care recipient must be met: (1) personal hygiene, (2) excretion, (3) nutrient intake. Regarding (1), the development of computer vision and multimodal technology has enabled robots to perform safer and more humanized bathing tasks^9,10. In terms of excretion, a variety of intelligent excretion care systems have also been developed to reduce the burden on human caregivers^11,12. In addition, telepresence robot is also regarded as an effective approach to addressing the above two aspects. By leveraging its remote control capabilities, caregivers can perform complex or physically demanding care tasks without direct contact, thereby effectively reducing physical fatigue and psychological stress^13,14,15. Nutritional intake means providing the care recipient with food, drinks and some necessary medicines. The robotic platform CHARMIE¹⁶, Lio¹⁷, Hobbit¹⁸ have all made progress in this area.

Among these three aspects, nutrient intake is essential for maintaining basic bodily functions of the care recipient. This work is based on a personal care robot named SUT-PCR, which is primarily designed to provide care services including delivering objects and controlling electrical appliances. The purpose of developing care robots is to enable them to effectively replace humans in performing caregiving tasks. Therefore, care robots not only need to have human-like capabilities to perform actions, such as indoor navigation^19,20 and object grasping^21,22, but also need to be able to think and reason like humans. Currently, research on the intelligence of care robots in performing caregiving tasks can generally be divided into four levels.

Level 1: Please give me a bottle of water on the table (the object and its location are specified).

Level 2: Please give me a glass of water (the object is specified, but the location is not).

Level 3: I’m thirsty (expresses a physiological desire without specifying an object or location).

Level 4: Without any command being made, when the care recipient actually needs care services, the care robot proactively inferences the required care items and implements the services.

For the first level, the state machine is a classic method for solving this problem^23,24. In the research on the second level, Dora can find a book with a specified name in an unfamiliar environment²⁵. CHARMIE can fetch a glass of water for a bedridden care recipient without being told exactly location¹⁶. As for the third level, our team’s previous research could enable the robot to reason about the best care service based on the physiological desires (e.g., hunger and thirst) presented by the care recipient^26,27. This research focuses on the fourth level, how a care robot should perform a care task if the care recipient is unable to make a care request due to physical reasons, or the human caregivers are not with the care recipient and thus unable to give care commands. Human caregivers can think proactively without any requests and provide proper care services to the care recipient at the right time. If care robots can effectively replace human caregivers in care tasks, the robots must also be proactive.

Robotic proactivity

How to make robots proactive has become a hot research topic among scholars in recent years. Harman et al. converted Planning Domain Definition Language (PDDL) defined problems into action graphs to predict human actions and provide proactive assistance²⁸. Pequeño-Zurro et al. proposed a proactive robotic control system that includes three modules: robotic navigation control, visual human detection, and online robot speed adaptation²⁹. This system can predict the walking speed of a human and adjust the robot’s speed accordingly for proactive guidance. Maroger et al. proposed a human gait trajectory prediction model and its coupling method with a Walking Pattern Generator (WPG), which can enable a robot to predict human walking trajectories and proactively accompany humans³⁰. Buyukgoz et al. proposed an integrated system that enables robots to exhibit proactive behaviors by reasoning human intentions and predicting possible future events³¹. Ujjwal et al. integrated the belief–desire–intention (BDI) model into the Robot Operating System (ROS) to enable robots to have proactive behaviors like humans³². Liu et al. trained a robot using examples of human-human interactions. An attention mechanism was used during training to allow the robot to recognize opportunities for proactive behavior³³. Grosinger et al. proposed a proactive robot computation framework that formalizes the actions performed by the robot, the state of the environment, the state of the user, and the user’s intentions to enable the robot to generate its own goals and implement them³⁴. However, these models generally assume that the care recipient is capable of physical activity and can express intentions or interact with the robot through observable behaviors. This presents a challenge when dealing with bedridden patients with communication disorders, as such individuals lack discernible actions or the ability to express their needs. The PCA proposed in this paper is a time driven proactive care framework that enables the robot to provide services, such as administering medication or offering water, even when the care recipient makes no request or displays no behavior. Therefore, PCA is not only distinct from existing intention or recognition based architectures, but also offers a novel approach to proactive care in low-interaction or non-interaction scenarios.

SUT-PCR

We developed SUT-PCR as a humanoid care robot for the purpose of life support, as shown in Fig. 1. It consists of humanoid upper body and omnidirectional mobile platform equipped with omni-directional wheels. It allows free movement on indoor floors as well as complex object manipulation.

Basic specifications

The basic parameters of the robot are shown in Table 1. The two arms of the robot (part H) have six degrees of freedom, the end gripper (part L) has one degree of freedom, and the head (part C) and waist (part E) have three degrees of freedom each. RGB depth (RGB-D) cameras are mounted on the head and chest (part A), and microphones and speakers (part B) are also equipped on the upper body of the robot, which support it to perform tasks such as recognizing as well as performing human robot interaction (HRI). Four omni-directional wheels are mounted on the mobile platform (part I), and they allow the robot to move in any direction without turning. It is also equipped with 2 laser sensors (part F), 6 ultrasonic sensors (part M) and 4 touch sensors (part G). This allows it to perform functions such as path planning and obstacle avoidance in narrow and complex environments (home environments), allowing it to perform tasks while also safeguarding property and personal safety. The mobile platform was also fitted with a torso lifter (part K), which allowed the robot to accomplish elevation in height and increased the flexibility of the robot to perform tasks.

Fig. 1

SUT-PCR.

Table 1 Basic parameters of SUT-PCR.

Care robot system

The schematic diagram of the care robot system is shown in Fig. 2. The system consists of hardware layer, data processing layer, decision layer and action server layer. The route of information transfer has been demonstrated with red and purple lines, where the red indicates information from bottom to top, while the purple indicates the opposite direction. The black lines represent specific action executions. Cameras and sensors located in the hardware layer are used to collect data on the physiological state of the care recipient and the environment they are in. The collected data is processed by the data processing layer and passed to the decision layer where PCA proposed in this paper is located. The appropriate care service is reasoned by PCA and the sequence of actions that should be performed is generated. The action sequence is processed through various algorithms capable of performing actions, such as indoor navigation and object manipulation, and then transmitted to the actuators that execute specific operations to provide care services (such as delivering food or changing the care recipient’s environment, etc.). It can be seen that the aim of this research is to enhance the intelligence of care robots, enabling them to acquire the ability of proactive care thereby providing more humanized services.

Fig. 2

Care robot system.

Proactive care architecture

General description of PCA

Our team previously proposed a Desire-Driven Reasoning System (DDRS)^26,27, which can infer the appropriate care service $\ a$ based on the physiological desire $\ d$ expressed by the care recipient. However, this system is not applicable to bedridden patients with communication disorders, as these individuals often face physical limitations that prevent them from accurately or fully expressing their desires.

5000 years ago, humans proposed the concept of “time” based on changes in objects and their relative motion. The duodecimal mechanical clock, invented 800 years ago, quantitatively described time using the periodic motion of celestial bodies. Although time is an unseen and untouchable abstract concept, it possesses uniformity and equality for all things. In other words, all things, including humans, equally experience time. Without time, the existence and nature of everything in the world would be inexplicable. Since the announcement of the 2017 Nobel Prize in Physiology or Medicine for “The discoveries of molecular mechanisms for the circadian rhythm”, many researches have been time-oriented, aiming to improve health and prevent diseases^35,36,37. Meanwhile, “chrononutrition” has been developing rapidly^38,39,40. This indicates that scientifically planned time-based health interventions can significantly enhance human physiological functions and overall health. The PCA proposed in this study is inspired by this concept, using time $\ t$ as the foundation of the entire caregiving framework. PCA is proposed as shown in Fig. 3 and consists of two components, PCM and DDRS. PCM is mainly based on time $\ t$ (e.g., the care recipient needs to drink water at regular intervals) and combines the physiological state of the care recipient $s_{k}$ (e.g., food and water should not be provided when the care recipient is asleep) and changes in the surrounding environment $\varvec{e}$ (e.g., checking whether the room temperature is comfortable or not every once in a while) to generate the care recipient’s desire $\ d$. The physiological state of the care recipient is defined as ${s}_{k}$, $k=1,2, \ldots , p$ represents different states of the care recipient, such as awake and sleep, as shown in Table 2.The surrounding environment of the care recipient is defined as a sequence of different parameters, $\varvec{e}=\left( e_{1}, e_{2}, \ldots , e_{l}\right)$, where $e_{1}, e_{2}, \ldots , e_{l}$ are different environmental parameters such as temperature and humidity. Combined with DDRS, this approach allows the robot to perform care tasks proactively without requiring commands.

Table 2 Definition of the care recipient’s physiological states.

Fig. 3

Proactive care architecture.

Fig. 4

General idea of PCA.

The general idea of PCA is shown in Fig. 4. The care recipient’s desires are set as a time-based sequence $\varvec{d}(t)$,

$$\begin{aligned} \varvec{d}(t)=\left[ d_{1}(t), d_{2}(t), d_{3}(t), \ldots , d_{n}(t)\right] \end{aligned}$$

(1)

here, $d_{1}(t), d_{2}(t), d_{3}(t), \ldots , d_{n}(t)$ are functions of $\ n$ types different desires based on time, the desire is generated when the function value is 1. At 7:00, the value of hunger function $d_{2}(t)$ changes to 1, and hunger is passed to DDRS. The role of DDRS is to select the object with the highest contribution value for the incoming desire. Based on the contribution of each object, DDRS infers that providing a cake is the most appropriate option. At 12:00, the value of $d_{2}(t)$ changes back to 1, but since the care recipient is in a sleep state, the service time is postponed. When the care recipient wakes up at 14:00, hunger is passed to DDRS, which infers that providing bread is the most appropriate service. Subsequently, the service time for hunger is also delayed accordingly. At 19:00, the value of the function $d_{2}(t)$ changes to 1 again, indicating that hunger needs to be passed to DDRS again. DDRS infers that providing biscuits is the most appropriate service object.

Proactive care model

The following describes how PCM generates the desires of care recipients. The process of desire generation is formalized as a time-based characteristic function $d_{i}(t)$, as shown in formula (2). When the function value is 1, the corresponding desire is generated.

$$\begin{aligned} d_{i}(t)=\left\{ \begin{array}{ll} 1 & t \in \left( t_{i 1}^{j}, t_{i 2}^{j}\right) \\ 0 & t \notin \left( t_{i 1}^{j}, t_{i 2}^{j}\right) \end{array}\right. \end{aligned}$$

(2)

where $\ t$ is time. $i=1,2, \ldots , n$, represents $\ n$ different desires of the care recipient. $j=1,2, \ldots , m$, represents $\ j$th generation of the $\ i$th desire within a day. $t_{i 1}^{j}$ is the start time of the $\ j$th generation of the $\ i$th desire, $t_{i 2}^{j}$ is the end time.

The relationship between the generation time and the end time of the desire can be expressed by following formula,

$$\begin{aligned} t_{i 2}^{j}=t_{i 1}^{j}+t_{a}^{i} \end{aligned}$$

(3)

where $t_{a}^{i}$ is the time that meet the $\ i$th desire once. The time should be different for different desires. For example, it takes 30 minutes to meet hunger, but only 10 minutes for thirst.

Considering that when a desire generates, the care recipient may be in a state which care services cannot be provided, resulting in a delay in service, the delay time for the $\ i$th desire is defined as $t_{d}^{i}$.

We then formally describe PCM. The function of PCM is to dynamically generate the care recipient’s desires $d_{i}$ based on the time t and combining $\varvec{e}$, $s_{k}$ and $t_{d}^{i}$. As shown in Fig. 5, PCM consists of two parts: generator and evaluator. The role of generator is to generate $\varvec{d}(t)$ by $\ t$, $t_{d}^{i}$ and $\varvec{e}$. The role of the evaluator is to assess whether the desire which value is 1 should be met immediately based on the physiological state $s_{k}$ of the care recipient. If there is a desire $d_{i}$ that can be met immediately, it will be passed to DDRS. If the desire cannot be met immediately, the delay time $t_{d}^{i}$ is sent back to the generator. The generator and evaluator will be discussed next.

Fig. 5

Details of PCM.

Fig. 6

Generator.

Generator

Good care services usually need to meet at least the internal desire (nutritional intake) and external desire (comfortable environment) of the care recipient. Desires of different types should be generated in different ways. Internal desires are usually met regularly over time. Take thirst as an example. Usually, the care recipient feels thirsty at regular intervals and need to be offered a drink. As for external desires, such as higher temperature, they usually generate due to changes in the surrounding environment of the care recipient. So, generator consists of internal desire generator (IDG) and external desire generator (EDG) as shown in Fig. 6.

Internal desire

The time interval for the generation of the same desire is described by following formula,

$$\begin{aligned} t_{i 1}^{j+1}=t_{i 1}^{j}+t_{b}^{i} \end{aligned}$$

(4)

where $t_{b}^{i}$ is the time interval. That is, the time of the next generation of desires is only related to the time of the last generation of the same type. The time interval varies depending on the type of desire. For example, for hunger, the interval is 5 h, and for thirst, the interval is 2 h.

The delay of desires that can not be immediately met due to the physiological state of the care recipient is as follows,

$$\begin{aligned} t_{i 1}^{j}{\phantom {A}}^{\prime }=t_{i 1}^{j}+t_{d}^{i} \end{aligned}$$

(5)

where $t_{d}^{i}$ is the delay time. It also causes a time adjustment for the subsequent generation of such desires, so that care services can be provided in a more reasonable manner. For example, the care recipient is supposed to have lunch at 12:00 and dinner at 18:00. However, since the care recipient is taking a nap at 12:00, resulting in lunch not being served until 13:00, according to this method, the time for serving dinner afterwards would also be adjusted to 19:00.

External desire

The generation of external desires is based on assessing the environment $\varvec{e}$ over time $\ t$. The environment $\varvec{e}$ should be kept within a comfortable range for the care recipient, and the comfortable environment should change over time. For example, the temperature in the house should be raised appropriately when the care recipient is sleeping. As shown in Fig. 7, [x(t), y(t)] is the comfort zone of the care recipient for a certain environmental parameter. We define the comfort domain of the entire environment as $\varvec{C}(t)$, as shown in the following formula,

$$\begin{aligned} \varvec{C}(t)=\left\{ \left[ x_{1}(t), y_{1}(t)\right] ,\left[ x_{2}(t), y_{2}(t)\right] , \ldots ,\left[ x_{l}(t), y_{l}(t)\right] \right\} \end{aligned}$$

(6)

where $[x_{1}(t), y_{1}(t)],[x_{2}(t), y_{2}(t)], \ldots ,[x_{l}(t), y_{l}(t)]$ represent different environmental parameters. If the corresponding parameters in $\varvec{e}$ are all within the corresponding [x(t), y(t)] range, the care recipient is considered to be in the comfort zone. If there are parameters that are out of [x(t), y(t)] range, a corresponding desire should be generated. x(t) and y(t) are set to functions that change over time as shown in the following formula,

$$\begin{aligned} \begin{array}{l} x(t)=\left\{ \begin{array}{ll} x=a & t \in \left( t_{1}, t_{2}\right) \\ x=b & t \notin \left( t_{1}, t_{2}\right) \end{array}\right. \\ y(t)=\left\{ \begin{array}{ll} y=c & t \in \left( t_{1}, t_{2}\right) \\ y=d & t \notin \left( t_{1}, t_{2}\right) \end{array}\right. \end{array} \end{aligned}$$

(7)

where $(t_{1}, t_{2})$ is the set time range. Take temperature as an example, $a=23\,^{\circ } \text{C}$, $b=26\,^{\circ } \text{C}$, $c=25\,^{\circ } \text{C}$, $d=28\,^{\circ } \text{C}$, indicates that the care recipient’s comfortable temperature within $(t_{1}, t_{2})$ is 23–26 $\,^{\circ } \text{C}$, outside $(t_{1}, t_{2})$ is 25–28$\,^{\circ } \text{C}$. If the temperature is outside the comfort zone, there should be generated a desire to change the temperature. The specific workflow of the external desires module is shown in Fig. 8. Using a 24-h period as a cycle, the environment is checked every minutes based on time $\ t$, and the previously mentioned method is used to determine whether there is a desire that needs to be met based on $\varvec{e}$, $\ t$, and $\varvec{C}(t)$. If not, continue the cycle. If so, update the start and end time of the corresponding desire function $d_{i}(t)$, and continue the loop until the 24-h period ends.

Fig. 7

Comfort zone of environment.

Fig. 8

Flowchart of external desire.

Fig. 9

Flowchart for the evaluator.

Evaluator

The role of the evaluator is to determine whether the desire should be passed to DDRS based on the combination of the care recipient’s state $s_{k}$ and the desire $\varvec{d}(t)$. First, define what kind of desires can be passed when the care recipient is in what kind of state, as shown in Table 3. The flowchart for the Evaluate is shown in Fig. 9. When $\varvec{d}(t)$ is passed from generator to evaluator, first check to see if $d_{i}(t)=1$ exists. If it doesn’t, the evaluation ends. If so, the care recipient’s state ${s}_{k}$ is combined to assess whether desire $d_{i}(t)$ is passed according to Table 3. If the $d_{i}(t)=1$ can be met immediately, it is passed to DDRS. If not, the delay time $t_{d}^{i}$ for the $d_{i}(t)$ returning to the generator. For example, when $d_{1}(t)=1$, the care recipient in state $s_{1}$, then $d_{1}$ can be passed according to Table 3. But when $d_{1}(t)=1$, the care recipient in state $s_{2}$, the delay time $t_{d}^{1}$ of the $d_{1}(t)$ returning to the generator.

Table 3 State-desire evaluation system.

Simulation

For the performance evaluation of care robot systems in caring for bedridden patients, the following challenges may arise:

• (1) Since a care robot needs complete functions to perform care tasks (such as perception of the environment and object manipulation), and the system is complex, conducting experiments directly in a real home or hospital environment would be costly and unsafe.

• (2) If the proposed method has certain defects, providing services to care recipients without prior evaluation and fine-tuning may lead to a poor user experience, further increasing the burden on both caregivers and care recipients²⁷.

Therefore, a simulation system based on a virtual human has been developed to evaluate and fine-tune the proposed method before applying them in a real care domain.

Simulation system

In order to simulate a realistic care environment, care robot, care recipient, and environment are modeled as “virtual robot,” “virtual human,” and “virtual environment”. The simulation system consists of these three components. Since DDRS^26,27 has been proven effective, we only need to verify that the proposed method can proactively and rationally generate the physiological desires of the care recipient. Virtual robot is set up to sense the specific parameters of the virtual environment, the physiological state of the virtual human and to proactively generate the physiological desires of the virtual human. Virtual human only simulates the daily life and physiology state, and does not issue any requests for care. Specific attributes are configured according to the following three dimensions:

• (1) Basic information: name, gender, height, weight.

• (2) Life habits: Life habits are described as the following: (1) time of waking up, (2) time of napping, (3) duration of napping, (4) time of taking medication, (5) time of eating, (6) time of drinking, (7) comfortable environment, and (8) time of sleeping at night.

• (3) Characteristic: Characteristics are mainly used to describe the physiological state of the care recipient.

Regarding virtual environment, in order to present the simulation results more clearly, only the temperature is chosen in this paper. At the same time, in order to simulate the environment of a real care recipient and to test the proposed method, the values of the environmental parameters are randomly varied within a given reasonable range.

Simulation settings

For the method proposed in this paper, it is mainly tested in two aspects, (1) whether the method can make the robot proactive in performing the care task, and (2) whether the method can deal with unexpected situations (e.g., whether the robot can promptly adjust care tasks when the care recipient needs water but is in sleep state, or when the care recipient experiences special state). The settings for the virtual human are shown in Table 4, where the characteristic “hypertension” is a permanent trait of the virtual human, requiring long-term medication, and is not considered a special state. On the other hand, the characteristic “diarrhea” is considered an unexpected state. If it occurs, the care recipient’s state $s_{k}$ will change to “diarrhea” and the virtual robot needs to handle promptly. The traits which are temporary traits or permanent traits have been preset in the system. The possible states of virtual human are shown in Table 5. As shown in Table 6, virtual robot generates five different physiological desires according to the proposed method. The environmental temperature is set to fluctuate between $22{-}28\,^{\circ } \text{C}$. The desires of the care recipients that can be met in different physiological states are shown in Table 7.

Table 4 Virtual human configuration.

Table 5 Definition of the state.

Table 6 Desires of the care recipient.

Table 7 State-desire evaluation system.

Simulation results

A week of care is simulated in the simulation system, and one of the purposes of the simulation is the ability of the virtual robot to deal with an unexpected state (the appearance of diarrhea in the virtual human), so the timing of the appearance of diarrhea is not a focus. The time when virtual human appeared diarrhea is set for the weekend. The simulation results are shown in Figs. 10 and 10a–g show the simulation results for each day of the week.

As shown in Fig. 10a–e, during the first 5 days of the simulation, virtual human is in a normal state (awake or sleep). Even if virtual human does not issue any command, the proposed method can proactively generate desires. Also, this method can dynamically adjust the service time based on the state of the virtual human. As shown in Fig. 10a, thirst should have been generated at 13:10, but since the virtual human was asleep at that time, it waited until virtual human wake up. Also, when the virtual human is in sleep state, Only temperature-related desires are generated, which proves the rationality of this approach. As shown in Fig. 10f,g, when the virtual human is in diarrhea (an expected state occur), hunger and thirst are no longer generated, but other desires can still be generated.The simulation is conducted on a PC equipped with a 2.2 GHz Intel Core i9-13900HX. The average reasoning time for each desire is approximately 1 $\mu s$, and the reasoning success rate is $100\%$. Thus, it is demonstrated that the proposed method can deal with the unexpected state of the care recipient. By Combining with DDRS we proposed,proactive care can be achieved.

Discussion

With the intelligence development of caregiving robots, the manner in which robots perform caregiving tasks will also change. Initially, robots autonomously executed tasks based on complete instructions, and later evolved to handle incomplete instructions and interpret abstract physiological desires. However, due to the shortage of caregivers, they are often not present by the patient’s side, making it impossible to give care commands. Bedridden patients with communication disorder are also unable to request care from the care robot. These would result in the robot being unable to perform care tasks due to not receiving commands. Care robots are developed to effectively replace human caregivers in care tasks, which means that care robots need human-like thinking ability. That is, they can think proactively and perform care tasks without any care request, yet current care robots do not have this kind of intelligence.

In this work, PCA can generate the physiological desires of the care recipient and proactively provide the right care service to the care recipient. When human caregivers perform care tasks, they do not rely solely on the requests or intentions of the care recipient, but also think independently about the best care service for the care recipient. Compared with existing methods of performing care tasks, this method does not require any commands. PCA provides care services to care recipients based on time, taking into account their physiological state and the environment which they located in. Human caregivers will not provide food or drinks while the care recipient is sleeping. PCA can similarly avoid sleep time by waiting for the care recipient to wake up and then providing food and drinks. Caregivers can adjust the comfort level of the environment in which the care recipient is situated, and PCA also has this function. In addition, human caregivers can adapt care services to the special condition of the care recipient, and PCA can select the correct care service based on the different physiological states of the care recipient. Therefore, PCA proposed in this paper enables the robot to possess intelligence similar to human caregivers, equipping it with proactive care capabilities.

Although the PCA architecture proposed in this study has been validated in the simulated environment with promising preliminary results, we are fully aware of the inherent limitations of simulation systems. The scenarios, sensory data, and interaction complexity within simulations are often idealized and cannot fully replicate the dynamic changes and multi-source disturbances present in real-world environments. For instance, in actual caregiving scenarios, robots must contend with navigation obstacles in complex spatial layouts, accurate object delivery in unstructured settings and sensor noise, these challenges are not adequately reflected in the current simulation platform. Therefore, our future research will focus on advancing the application of PCA in more complex and realistic environment, with particular attention to improving the robustness and adaptability of the robot in executing specific tasks. In addition, to comprehensively evaluate the effectiveness and practicality of the proposed system in real-world caregiving tasks, we plan to conduct experiments involving actual patients. These studies will be conducted under strict medical ethical guidelines to ensure informed consent, data privacy, and experimental safety. We will adhere to established ethical standards and progressively promote the deployment and implementation of PCA in real clinical and home care settings.

Integration challenges

The proposed PCA is essentially a time-driven proactive care architecture designed to enhance the intelligence of caregiving robots, rather than being limited to any specific robotic platform. Owing to its clear logic and concise mechanism, this method possesses high generalizability. However, several challenges may arise when integrating it with other caregiving robot platforms.First, hardware heterogeneity is one of the primary obstacles. Different robots are equipped with varying types of sensors and actuators, which directly affects the accuracy of perception and the effectiveness of care task execution. Second, inconsistencies at the software level also introduce integration complexity. Lastly, cross-platform deployment involves ethical and data governance issues–especially in multi-institutional or cross-border scenarios, where privacy policies, data ownership, and access control standards may differ across platforms, potentially impacting data acquisition and usage.To address these challenges, our future work will focus on developing a modular, platform-independent API layer. In addition, we plan to conduct field deployments and tests across multiple heterogeneous platforms to evaluate the adaptability and performance of the system in diverse application environments.

Conclusion

In this paper, we have proposed PCA, enabling care robots to perform appropriate care tasks without explicit requests from the care recipient. Additionally, we have introduced PCM, which dynamically generates physiological desires of the care recipient. This approach aligns more closely with the intuitive thinking of human caregivers. The effectiveness of the proposed architecture and model was validated through simulation experiments, demonstrating their potential to enable care robots to perform tasks akin to human caregivers. If the architecture is implemented in a care robot, the robot can perform care tasks like human caregivers. In future work, we aim to extend the proactive care approach to more complex scenarios, such as accounting for the recipient’s emotional state (e.g., offering their favorite food or drink when they are feeling down) or adapting care tasks to external conditions (e.g., avoiding window ventilation during rainy weather).

Fig. 10

Simulation results. Dimgray indicates the sleep state of the care recipient, gold indicates diarrhea, tomato indicates awake. Orange indicates hunger, dodgerblue indicates thirst, mediumorchid indicates medicine, red indicates higher temperature, blue indicates lower temperature. Mediumseagreen indicates a comfortable environment, orangered indicates a high temperature environment, lightskyblue indicates a low temperature environment.