A method for evaluating the comfort of a transfer nursing robot

Abstract

The application discloses a kind of comfort evaluation methods of transfer nursing robot, comprising the following steps: S1, acquisition pressure characteristic data and position posture data, construct training task, and the training task is divided, and support set and query set are obtained;S2, the support set and query set of training task are used to train the comfort evaluation model of transfer nursing robot, and the training completed comfort evaluation model of transfer nursing robot is obtained;S3, the task to be evaluated is evaluated using the training completed comfort evaluation model of transfer nursing robot, and the comfort of transfer nursing robot is obtained;The application measures comfort from position posture and pressure two aspects, so that the moderate evaluation result is more reliable, and is used in daily use, can update model according to input data, can better adapt to the body characteristics of different nursing population, while a small amount of sample can achieve good training effect.

Description

Technical Field

[0001] This invention relates to the field of medical device technology, and specifically to a method for assessing the comfort of a transfer nursing robot. Background Technology

[0002] Transfer care robots are a type of robot that assists medical staff in providing daily care for bedridden patients and other individuals with limited mobility. They can transfer patients between beds, wheelchairs, and operating tables. Since their primary users are disabled elderly people, patients, and other physically vulnerable individuals, even mild discomfort that is tolerable for most people can cause them significant psychological burden. Furthermore, some patients experience communication difficulties and cannot accurately express their feelings during the use of transfer care robots, making it impossible for caregivers to adjust the patient's transfer posture accordingly. To address these issues, this invention proposes a comfort evaluation method for transfer robots. This method not only guides the motion control and planning of the transfer robot but also provides comfort evaluations during use, helping caregivers adjust the patient's transfer posture.

[0003] The following are some existing comfort assessment methods:

[0004] 1. Comfort Assessment Method Based on Weighted Function. A dynamic model of the person-carrying transfer motion is established, and the force factors affecting comfort are analyzed. A weighted function is used to assess comfort, with the weights determined by methods such as the analytic hierarchy process (AHP) and subjective weighting. This type of method analyzes comfort influencing factors using a dynamic model, considering only the force level and treating the contact point between the human body and the robotic arm as point contact, neglecting many pressure characteristics, resulting in poor reliability of the comfort evaluation results.

[0005] 2. Comfort Assessment Method Based on Fuzzy Mathematics. This method divides influencing factors into target, intermediate, and bottom layers using the analytic hierarchy process (AHP). It then calculates the weight matrices and evaluation matrices for each factor in the intermediate and bottom layers and performs fuzzy calculations to complete the final evaluation of the target layer. The final evaluation is represented by a comprehensive evaluation vector obtained through fuzzy calculations of the indicator weight set and the single-factor evaluation matrix. However, this method is cumbersome in its solution process, cannot update the comfort assessment model with data, and is difficult to apply to the daily use of transfer and nursing robots, thus hindering the provision of comfort assessments.

[0006] 3. Neural Network-Based Comfort Assessment Method. This method obtains the characteristic parameters of each comfort-influencing factor and their corresponding comfort ratings through experiments. These parameters are then imported into a neural network model for training, ultimately yielding the mapping relationship between each comfort-influencing factor and the comfort score—the comfort assessment model. While this method can automatically output quantitative comfort assessment results using a computer and easily adapts the assessment model to data updates, it requires a large training sample size. Since those receiving care are typically disabled elderly or patients, obtaining such samples is challenging. Summary of the Invention

[0007] To address the aforementioned shortcomings in the existing technology, the present invention provides a method for evaluating the comfort of a transfer nursing robot, which solves the following technical problems:

[0008] 1. The comfort assessment method based on weighted functions only considers the force level and treats the contact position between the human body and the robotic arm as point contact, ignoring many pressure characteristics, resulting in poor reliability of the comfort evaluation results.

[0009] 2. The comfort assessment method based on fuzzy mathematics cannot update the comfort evaluation model with data, making it difficult to provide comfort evaluation in the daily use of the transfer and care robot.

[0010] 3. The comfort assessment method based on neural networks requires a large number of training samples, and the care recipients are generally disabled elderly people, patients, etc., making it difficult to obtain such samples.

[0011] To achieve the above-mentioned objectives, the technical solution adopted by this invention is: a method for assessing the comfort of a transfer nursing robot, comprising the following steps:

[0012] S1. Collect pressure feature data and position and pose data, construct training tasks, and divide the training tasks to obtain support sets and query sets;

[0013] S2. The support set and query set of the training task are used to train the comfort evaluation model of the transfer nursing robot to obtain the trained comfort evaluation model of the transfer nursing robot.

[0014] S3. The trained comfort assessment model of the transfer nursing robot is used to evaluate the task to be evaluated, and the comfort of the transfer nursing robot is obtained.

[0015] Further, step S1 includes the following sub-steps:

[0016] S11. Acquire pressure characteristic data and position and attitude data through sensors;

[0017] S12. Normalize the pressure characteristic data and position and attitude data to obtain normalized data;

[0018] S13. Label the normalized data with comfort levels to obtain the sample dataset;

[0019] S14. Based on human characteristics, the sample dataset is divided into different task groups to obtain multiple task groups;

[0020] S15. Divide each task group into multiple training tasks, and divide each training task into two disjoint sets, which are called the support set and the query set.

[0021] The beneficial effects of the above-mentioned further solutions are as follows: the present invention combines pressure characteristic data and position and posture data to measure comfort from both position and posture and pressure aspects, making the comfort evaluation results more reliable.

[0022] Furthermore, the comfort evaluation model for the transfer nursing robot in step S2 includes: a task learning network and a parameter optimizer;

[0023] The task learning network comprises an input layer, a hidden layer, and an output layer connected in sequence, wherein the number of hidden layers is greater than or equal to two.

[0024] Furthermore, the loss function of the task learning network is:

[0025]

[0026] Where L is the loss function, y is the comfort label, x is the input of the task learning network, f(x) is the output of the task learning network, and δ is the error threshold parameter.

[0027] The beneficial effect of the above further scheme is that when the output f(x) of the task learning network differs significantly from the comfort label y, by... Calculate the loss value to prevent outliers from having an excessive impact on the model; when the output f(x) of the task learning network is relatively close to the comfort label y, [f(x) is calculated]. Calculating the loss value speeds up model training and makes the loss function differentiable when the prediction error is zero.

[0028] Furthermore, step S2 includes the following sub-steps:

[0029] S21. Randomly initialize the initial values ​​of the network parameters of the task learning network, configure an independent learning rate for each network parameter, and form a vector with the same dimension and corresponding element positions as the initial values ​​of the network parameters, i.e., the task learning rate.

[0030] S22. Take a batch of training tasks, and for each training task, perform the following steps S23-S26;

[0031] S23. Train the task learning network using the support set of the current training task, and calculate the task adaptation gradient;

[0032] S24. Iteratively update the network parameters based on the task adaptation gradient and the task learning rate;

[0033] S25. Pass the updated network parameters to the task learning network to obtain the updated task learning network;

[0034] S26. Train the updated task learning network using the query set of the current training task to obtain the loss function corresponding to the query set;

[0035] S27. Based on the loss function corresponding to the query set of each training task in this batch of tasks, calculate the average gradient of the network parameters and the average gradient of the task learning rate in this batch of tasks.

[0036] S28. Based on the average gradient of the network parameters and the average gradient of the task learning rate on this batch of training tasks, update the initial values ​​of the network parameters and the task learning rate in the parameter optimizer.

[0037] S29. Pass the updated initial values ​​of network parameters and task learning rate to the task learning network, and jump to step S22 until the set number of training times is reached to obtain the training completed comfort evaluation model of the transfer nursing robot.

[0038] The beneficial effects of the above-mentioned further scheme are: by learning from similar tasks, the optimal initial values ​​of network parameters and task learning rate for this type of task can be found, so that the model can start training on new tasks with better initial network parameters and a better learning path, thereby achieving the effect of completing training with a small number of samples.

[0039] Furthermore, the formula for calculating the task adaptation gradient in step S23 is as follows:

[0040]

[0041] Where g1 is the task adaptation gradient, L support To support the loss function corresponding to the set, These are the network parameters corresponding to the m-th training task in the (k-1)-th iteration. For network parameters Calculate the partial derivatives for each element.

[0042] Furthermore, the formula for iteratively updating the network parameters in step S24 is as follows:

[0043]

[0044]

[0045] in, These are the network parameters corresponding to the m-th training task in the k-th iteration. Let θ represent the network parameters corresponding to the m-th training task in the 0th iteration, and let θ be the initial value of the network parameters. Let be the network parameters corresponding to the m-th training task in the (k-1)-th iteration, α be the task learning rate, ⊙ be the Hadamard product operation, and g1 be the task adaptive gradient.

[0046] Furthermore, the formulas for calculating the average gradient of the network parameters and the average gradient of the task learning rate in step S27 are as follows:

[0047]

[0048]

[0049] Where g2 is the average gradient of the network parameters, M is the number of training tasks in the training task batch taken in step S22, m is the training task number, and L qurry The loss function corresponding to the query set. Let g be the network parameters corresponding to the m-th training task in the k-th iteration, α be the task learning rate, and g3 be the average gradient of the task learning rate α. To obtain the element-wise partial derivative of the task learning rate α, For network parameters Calculate the partial derivatives for each element.

[0050] Furthermore, the formula for updating the initial values ​​of network parameters and the task learning rate in step S28 is as follows:

[0051]

[0052]

[0053] Where θ′ is the updated initial value of the network parameters in step S28, θ is the initial value of the network parameters, and β max β represents the maximum learning rate of the network parameters on the parameter optimizer, where β is the initial value of the network parameters. min Let P be the minimum learning rate of the network parameters with initial values ​​on the parameter optimizer, and let P be the total number of training iterations in the current training cycle on the parameter optimizer. j For the j-th training iteration in the current training cycle on the parameter optimizer, sign() is the sign function, fract() is the decimal function, α′ is the updated task learning rate in step S28, and α is the task learning rate before the update. The maximum learning rate of the task on the parameter optimizer. g1 represents the minimum learning rate of the task learning rate on the parameter optimizer, g2 represents the average gradient of the network parameters, and g3 represents the average gradient of the task learning rate α.

[0054] In summary, the beneficial effects of this invention are as follows:

[0055] 1. This invention combines pressure characteristic data and position and posture data to measure comfort from both position and posture and pressure aspects, making the comfort evaluation results more reliable.

[0056] 2. The network parameters in the comfort assessment model of the transfer nursing robot can be trained and updated based on the input data. When applied in daily use, it can better adapt to the physical characteristics of different nursing populations and provide more accurate comfort evaluation and usage guidance.

[0057] 3. By learning from similar tasks, find the optimal initial values ​​of network parameters and task learning rate for this type of task, so that the model can learn more efficiently on new tasks of this type and achieve good training results with a small number of samples. Attached Figure Description

[0058] Figure 1 A flowchart of a method for evaluating the comfort of a transfer nursing robot;

[0059] Figure 2 A diagram showing the factors affecting comfort.

[0060] Figure 3 This is a diagram illustrating the location of the parameters. Detailed Implementation

[0061] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0062] like Figure 1 As shown, a method for assessing the comfort of a transfer care robot includes the following steps:

[0063] S1. Collect pressure feature data and position and pose data, construct training tasks, and divide the training tasks to obtain support sets and query sets;

[0064] Step S1 includes the following sub-steps:

[0065] S11. Acquire pressure characteristic data and position and attitude data through sensors;

[0066] In this embodiment, pressure sensor arrays are arranged on the two arms of the transfer care robot, and depth vision sensors are arranged on the robot's head or other suitable locations.

[0067] Experimenters of different ages, heights, and weights were selected, and their age, height, and weight information were labeled. The transfer nursing robot was then used to hold the participants, and the robot's arms were adjusted to maintain the correct posture relative to the participants, while sensor data was recorded. Comfort factors were considered from two aspects: pressure characteristics at the pressure points and the influence of human posture. Figure 2 As shown.

[0068] like Figure 3 As shown, the following data is obtained from the sensors: left arm pressure area X1, right arm pressure area X2, average pressure of the left arm X3, average pressure of the right arm X4, maximum pressure of the left arm X5, maximum pressure of the right arm X6, distance from the center of pressure of the left arm to the center of the hip X7, distance from the center of pressure of the right arm to the center of the hip X8, human neck joint angle X9, and human hip joint angle X1. 10 Human knee joint angle X 11 .

[0069] X1 to X6 represent pressure characteristic data, acquired from an array of pressure sensors mounted on the robot arm; X7 to X... 11 The human body position and pose data are acquired by a depth vision sensor deployed on the robot's head.

[0070] S12. Normalize the pressure characteristic data and position and attitude data to obtain normalized data;

[0071] The normalization formula is:

[0072]

[0073] Where, x i For the i-th normalized data, X i X1 to X2 obtained by the above sensors 11 X Max X is the maximum value. Min For the pressure characteristic data X1 to X6, the maximum and minimum values ​​among all pressure characteristic data are taken as the minimum. For the distance data X7 and X8, the minimum value is 0, meaning that the center of the pressure point coincides with the center of the hip. The maximum value of X7 is the distance from the center of the experimenter's neck to the center of the hip, and the maximum value of X8 is the distance from the center of the experimenter's ankle joint to the center of the hip. For X9 to X... 11 The joint angles of the human body are determined with reference to the range of motion of each joint and the specific circumstances of the hugging exercise.

[0074] S13. Label the normalized data for comfort levels to obtain the sample dataset;

[0075] S14. Based on human characteristics, the sample dataset is divided into different task groups to obtain multiple task groups;

[0076] The human characteristics mentioned include age, height, and weight.

[0077] In this embodiment, to increase the discriminative power of different task groups and enhance the model's ability to learn new task features, the sample dataset needs to be divided into different task groups based on features such as age, height, and weight. Using the experimenter's age, height, and weight as references, the sample dataset is divided into 6 task groups: a1, a2, b1, b2, c1, and c2.

[0078] First, the data samples of the participants were randomly divided into three groups: a, b, and c. Group a was based on age; participants under 40 years old were grouped into a1, and those over 40 years old into a2. Group b was based on height; participants under 160cm were grouped into b1, and those over 160cm into b2. Group c was based on weight; participants under 60kg were grouped into c1, and those over 60kg into c2.

[0079] S15. Divide each task group into multiple training tasks, and divide each training task into two disjoint sets, which are called the support set and the query set.

[0080] Specifically, based on the task group, the data samples of each task group are divided into K equal parts, and each part is used as a training task for an independent training model.

[0081] S2. The support set and query set of the training task are used to train the comfort evaluation model of the transfer nursing robot to obtain the trained comfort evaluation model of the transfer nursing robot.

[0082] The comfort evaluation model for the transfer nursing robot in step S2 includes: a task learning network and a parameter optimizer;

[0083] The task learning network comprises an input layer, a hidden layer, and an output layer connected in sequence, wherein the number of hidden layers is greater than or equal to two.

[0084] The loss function of the task learning network is:

[0085]

[0086] Where L is the loss function, y is the comfort label, x is the input of the task learning network, f(x) is the output of the task learning network, and δ is the error threshold parameter.

[0087] In this embodiment, the input layer has 11 neurons, corresponding to 11 normalized x-types. i The number of output layer units is set to 1, representing the comfort score. To increase the model's nonlinear representation capability, ReLU activation functions are configured for both hidden layers A and B.

[0088] S21. Randomly initialize the initial values ​​of the network parameters of the task learning network, configure an independent learning rate for each network parameter, and form a vector with the same dimension and corresponding element positions as the initial values ​​of the network parameters, i.e., the task learning rate.

[0089] S22. Take a batch of training tasks, and for each training task, perform the following steps S23-S26;

[0090] S23. Train the task learning network using the support set of the current training task, and calculate the task adaptation gradient;

[0091] The formula for calculating the task adaptation gradient in step S23 is as follows:

[0092]

[0093] Where g1 is the task adaptation gradient, L support To support the loss function corresponding to the set, These are the network parameters corresponding to the m-th training task in the (k-1)-th iteration. For network parameters Calculate the partial derivatives for each element.

[0094] S24. Iteratively update the network parameters based on the task adaptation gradient and the task learning rate;

[0095] The formula for iteratively updating the network parameters in step S24 is as follows:

[0096]

[0097]

[0098] in, These are the network parameters corresponding to the m-th training task in the k-th iteration. Let θ represent the network parameters corresponding to the m-th training task in the 0th iteration, and let θ be the initial value of the network parameters. Let be the network parameters corresponding to the m-th training task in the (k-1)-th iteration, α be the task learning rate, ⊙ be the Hadamard product operation, and g1 be the task adaptive gradient.

[0099] S25. Pass the updated network parameters to the task learning network to obtain the updated task learning network;

[0100] S26. Train the updated task learning network using the query set of the current training task to obtain the loss function corresponding to the query set;

[0101] S27. Based on the loss function corresponding to the query set of each training task in this batch of tasks, calculate the average gradient of the network parameters and the average gradient of the task learning rate in this batch of tasks.

[0102] The formulas for calculating the average gradient of the network parameters and the average gradient of the task learning rate in step S27 are as follows:

[0103]

[0104]

[0105] Where g2 is the average gradient of the network parameters, M is the number of training tasks in the training task batch taken in step S22, m is the training task number, and L qurry The loss function corresponding to the query set. Let g be the network parameters corresponding to the m-th training task in the k-th iteration, α be the task learning rate, and g3 be the average gradient of the task learning rate α. To obtain the element-wise partial derivative of the task learning rate α, For network parameters Calculate the partial derivatives for each element.

[0106] S28. Based on the average gradient of the network parameters and the average gradient of the task learning rate on this batch of training tasks, update the initial values ​​of the network parameters and the task learning rate in the parameter optimizer.

[0107] The formula for updating the initial values ​​of network parameters and the task learning rate in step S28 is as follows:

[0108]

[0109]

[0110] Where θ′ is the updated initial value of the network parameters in step S28, θ is the initial value of the network parameters, and β max β represents the maximum learning rate of the network parameters on the parameter optimizer, where β is the initial value of the network parameters. min Let P be the minimum learning rate of the network parameters with initial values ​​on the parameter optimizer, and let P be the total number of training iterations in the current training cycle on the parameter optimizer. j For the j-th training iteration in the current training cycle on the parameter optimizer, sign() is the sign function, fract() is the decimal function, α′ is the updated task learning rate in step S28, and α is the task learning rate before the update. The maximum learning rate of the task on the parameter optimizer. g1 represents the minimum learning rate of the task learning rate on the parameter optimizer, g2 represents the average gradient of the network parameters, and g3 represents the average gradient of the task learning rate α.

[0111] In this embodiment, when P j <P, for each complete parameter update of θ and α, let P j+1 =P j +1; when P j =P, then P j Reset to 0, let P = ε·P, and start the learning rate decay for the next cycle. The growth coefficient ε is an integer greater than 1.

[0112] During the training and updating of the initial values ​​of network parameters and the task learning rate, the learning rate is periodically decayed and restarted by taking advantage of the alternating pattern of slow and fast decay. This helps the model get rid of local optima during training and achieve better training results.

[0113] S29. Pass the updated initial values ​​of network parameters and task learning rate to the task learning network, and jump to step S22 until the set number of training times is reached to obtain the training completed comfort evaluation model of the transfer nursing robot.

[0114] In this embodiment, the initial values ​​of the total number of training iterations P, the growth coefficient ε, and the maximum learning rate β of the network parameters on the parameter optimizer for the current training cycle are: max The minimum learning rate β of the initial network parameters on the parameter optimizer. min The maximum learning rate of the task on the parameter optimizer. The minimum learning rate of the task learning rate on the parameter optimizer. This is set according to the needs of the project.

[0115] In the loops S22 to S29, after each loop, the latest values ​​of the task learning rate α and the initial values ​​of the network parameters θ are adopted. For example, after the first loop, the initial value of the network parameters θ is θ′ and the task learning rate α is α′. That is, the latest initial values ​​of the network parameters and the task learning rate are adopted in each loop.

[0116] S3. The trained comfort assessment model of the transfer nursing robot is used to evaluate the task to be evaluated, and the comfort of the transfer nursing robot is obtained.

[0117] It should be noted that, in this invention, the term "comprising" or any other variation thereof is intended to cover a non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0118] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

Claims

1. A method for evaluating the comfort of a transfer nursing robot, characterized in that, Includes the following steps: S1. Collect pressure feature data and position and pose data, construct training tasks, and divide the training tasks to obtain support sets and query sets; S2. The support set and query set of the training task are used to train the comfort evaluation model of the transfer nursing robot to obtain the trained comfort evaluation model of the transfer nursing robot. S21. Randomly initialize the initial values ​​of the network parameters of the task learning network, configure an independent learning rate for each network parameter, and form a vector with the same dimension and corresponding element positions as the initial values ​​of the network parameters, i.e., the task learning rate. S22. Take a batch of training tasks, and for each training task, perform the following steps S23-S26; S23. Train the task learning network using the support set of the current training task, and calculate the task adaptation gradient; S24. Iteratively update the network parameters based on the task adaptation gradient and the task learning rate; S25. Pass the updated network parameters to the task learning network to obtain the updated task learning network; S26. Train the updated task learning network using the query set of the current training task to obtain the loss function corresponding to the query set; S27. Based on the loss function corresponding to the query set of each training task in this batch of tasks, calculate the average gradient of the network parameters and the average gradient of the task learning rate in this batch of tasks. S28. Based on the average gradient of the network parameters and the average gradient of the task learning rate on this batch of training tasks, update the initial values ​​of the network parameters and the task learning rate in the parameter optimizer. The formula for updating the initial values ​​of network parameters and the task learning rate in step S28 is as follows: in, These are the updated initial values ​​of the network parameters from step S28. These are the initial values ​​for the network parameters. The maximum learning rate of the network parameters on the parameter optimizer is the initial value of the network parameters. The minimum learning rate of the network parameters on the parameter optimizer is the initial value of the network parameters. This represents the total number of training iterations in the current training cycle on the parameter optimizer. For the current training cycle on the parameter optimizer, the first... This training session For symbolic functions, For the decimal function, The updated task learning rate in step S28. The learning rate of the task before the update. The maximum learning rate of the task on the parameter optimizer. The minimum learning rate of the task on the parameter optimizer. The average gradient of the network parameters, For task learning rate The average gradient; S29. Pass the updated initial values ​​of network parameters and task learning rate to the task learning network, and jump to step S22 until the set number of training times is reached to obtain the training completed comfort evaluation model of the transfer nursing robot. S3. The trained comfort assessment model of the transfer nursing robot is used to evaluate the task to be evaluated, and the comfort of the transfer nursing robot is obtained.

2. The method for evaluating the comfort of a transfer nursing robot according to claim 1, characterized in that, Step S1 includes the following sub-steps: S11. Acquire pressure characteristic data and position and attitude data through sensors; S12. Normalize the pressure characteristic data and position and attitude data to obtain normalized data; S13. Label the normalized data for comfort levels to obtain the sample dataset; S14. Based on human characteristics, the sample dataset is divided into different task groups to obtain multiple task groups; S15. Divide each task group into multiple training tasks, and divide each training task into two disjoint sets, which are called the support set and the query set.

3. The method for evaluating the comfort of a transfer nursing robot according to claim 1, characterized in that, The comfort evaluation model for the transfer nursing robot in step S2 includes: a task learning network and a parameter optimizer; The task learning network comprises an input layer, a hidden layer, and an output layer connected in sequence, wherein the number of hidden layers is greater than or equal to two.

4. The method for evaluating the comfort of a transfer nursing robot according to claim 3, characterized in that, The loss function of the task learning network is: in, For loss function, For comfort level, As input to the task learning network, The output of the task learning network, This is the error threshold parameter.

5. The method for evaluating the comfort of a transfer nursing robot according to claim 1, characterized in that, The formula for calculating the task adaptation gradient in step S23 is as follows: in, To adapt gradients for the task, To support the loss function corresponding to the set, For the first The iteration of the ... The network parameters corresponding to each training task For network parameters Calculate the partial derivatives for each element.

6. The method for evaluating the comfort of a transfer nursing robot according to claim 5, characterized in that, The formula for iteratively updating the network parameters in step S24 is as follows: in, For the first The iteration of the ... The network parameters corresponding to each training task For the 0th iteration The network parameters corresponding to each training task These are the initial values ​​for the network parameters. For the first The iteration of the ... The network parameters corresponding to each training task For task learning rate, For Hadamard product operations, Adapt gradients for the task.

7. The method for evaluating the comfort of a transfer nursing robot according to claim 6, characterized in that, The formulas for calculating the average gradient of the network parameters and the average gradient of the task learning rate in step S27 are as follows: in, The average gradient of the network parameters, This refers to the number of training tasks in the batch of training tasks selected in step S22. This is the training task number. The loss function corresponding to the query set. For the first The iteration of the ... The network parameters corresponding to each training task For task learning rate, For task learning rate The average gradient, To the task learning rate Calculate the partial derivatives by element. For network parameters Calculate the partial derivatives for each element.

Images