Massage robot massage path generation method

Abstract

This invention discloses a method for generating massage paths for a massage robot, belonging to the field of massage robot technology. The method includes: S1 multimodal data acquisition and fusion, S2 human body modeling and acupoint localization, S3 adaptive path planning, S4 force-position hybrid control and safety mechanism, and S5 dynamic interaction. This invention, in addition to generating massage paths for the robot, can also achieve multimodal perception fusion and dynamically adapt to the environment.

Description

Technical Field

[0001] This invention relates to the field of massage robot technology, and more specifically, to a method for generating massage paths for a massage robot. Background Technology

[0002] Massage, also known as Tuina, refers to the act of pressing or stretching the muscles under the skin through physical contact to unblock meridians, lubricate joints, promote blood circulation, regulate organ function, and enhance the body's disease resistance, thereby achieving the purpose of curing ailments. With the improvement of living standards, people's demand for massage is increasing, leading to the widespread use of massage robots.

[0003] Traditional solutions rely on a single sensor and lack cross-modal alignment and fusion mechanisms. Visual positioning is greatly affected by occlusion. Biosignals are not associated with spatial coordinates, resulting in relatively large acupoint positioning errors. Point cloud denoising relies on simple filtering. Force control signals are affected by mechanical vibration. Environmental perception is static, leading to positioning failure in low-light environments. Although traditional round-trip algorithms guarantee full coverage, the path jitter rate is high. Fixed-weight evaluation functions cannot adapt to dense obstacles and have low throughput in narrow passages. Summary of the Invention

[0004] 1. Technical problems to be solved In view of the problems existing in the prior art, the purpose of this invention is to provide a method for generating massage paths for a massage robot. On the basis of generating massage paths for the robot, this invention can also realize multimodal perception fusion and dynamically adapt to the environment.

[0005] 2. Technical Solution

[0006] To solve the above problems, the present invention adopts the following technical solution: A method for generating massage paths for a massage robot, comprising: S1. Multimodal data acquisition and fusion; S2. Human body modeling and acupoint location; S3, Adaptive Path Planning; S4, Force-position hybrid control and safety mechanism; S5, dynamic interaction.

[0007] As a preferred embodiment of the present invention, step S1, multimodal data acquisition and fusion, includes: Multi-source sensor collaborative acquisition: Visual data and biosignal acquisition, as well as dynamic environmental perception, are performed, and the acquisition and perception timestamps are synchronized to ensure spatiotemporal alignment of multi-sensor data. Visual data is acquired through RGB-D cameras to collect 3D point cloud data of the human back and limbs, obtaining surface contour, spinal curvature, and body length spatial information. The body length spatial information is also the three-dimensional spatial information of the body. 33 skeletal key points are identified through the MediaPipe Pose model, muscle tension changes are monitored in real time through electromyography sensors, inflammation areas are located through infrared thermal imaging, contact force distribution is fed back through a six-dimensional force sensor, micro-movements of the human body are detected through an inertial measurement unit, and visual positioning capability is maintained in low-light environments through near-infrared supplementary lighting in conjunction with a high-sensitivity IR camera. Data cleaning and standardization: Visual data is denoised by point cloud based on statistical outlier filtering, and normal vector smoothing is used. Force control data is removed by Kalman filtering to remove mechanical vibration noise. Electromyography signals are extracted by Fourier transform to extract frequency domain features. Thermal imaging data is normalized to the [0,1] interval. ICP algorithm is used to register point cloud pose. Homogeneous transformation matrix T is used to compensate for human movement deviation to ensure spatiotemporal alignment. Cross-modal feature extraction: A lightweight YOLOv4 network is used to extract acupoint coordinates (u,v), which are mapped to 3D space. The amplitude of electromyography signal ΔA is used to identify muscle tension, and the thermal imaging temperature gradient ∇T is used to mark the inflammatory area. B-spline surface is used to fit the back contour, and the normal velocity of the massage point is calculated to ensure vertical force application. Feature-level fusion: Visual features H v With biological characteristics H p Acupoint-muscle associations are modeled using a weighted fusion approach based on attention mechanisms and cross-modal graphical neural networks. Decision-level fusion: When some modes fail, the independent model output is weighted and fused, and w1 and w2 are adjusted in real time according to the signal quality to avoid interference from single-mode noise; Real-time force control fusion: Six-dimensional force sensor data is input into a PID controller to dynamically adjust the trajectory normal velocity; Biofeedback optimization: Sudden changes in muscle signal amplitude trigger local path replanning, extending the massage duration of stiff areas; Edge computing acceleration: Feature extraction and preliminary fusion are completed at the acquisition end, and only compressed features are uploaded to the main control system.

[0008] As a preferred embodiment of the present invention, the visual coordinate normalization formula is as follows: ,in The maximum and minimum values ​​of the point cloud on the X-axis. These represent the maximum and minimum values ​​of the point cloud along the y-axis. These represent the maximum and minimum values ​​of the point cloud along the z-axis. The biosignal normalization formula is as follows: ,in The original amplitude of the muscle signal. The average amplitude within the signal window. The standard deviation of the amplitude; The contact force normalization formula is as follows: Where F is the contact force measured in real time by the six-dimensional force sensor and the unit is Newtons / N. Maximum contact force threshold, in Newtons per N. This represents the minimum contact force threshold, and its unit is Newtons per N. Attention-weighted fusion formula is , ,in The frequency domain characteristics are those of the muscle signal after Fourier transform. MLP(·) is a multilayer perceptron that maps frequency domain features to... In the same dimensional space, W and b are learnable weight matrices and biases, is the Sigmoid activation function, and it outputs weights in the range [0, 1]; || represents the feature vector concatenation operation; When thermal imaging is obstructed, the weighted fusion independent model output formula is: , , ,in For visual location reliability, Inferring probabilities from biological signals The signal-to-noise ratio of each modal signal; PID control to compensate for contact force error The formula for converting to normal velocity compensation is as follows: , This is the normalized force error. Gain and dimensionless It is an integral gain and is dimensionless. The maximum normal velocity is expressed in m / s.

[0009] As a preferred embodiment of the present invention, step S2, human body modeling and acupoint location, includes: Skeleton topology generation: Construct skeleton connections based on key points and calculate the connection probability between bones; Dynamic skeleton optimization: It uses IMU sensors to capture micro-movements caused by breathing and updates joint angles in real time; Skin surface reconstruction: Point cloud data is fitted using B-spline surfaces to generate a smooth skin model; Internal structure annotation: Segment CT / MRI data, reconstruct bones, muscles and blood vessels, and establish a layered anatomical model; Bone measurement and positioning; based on the traditional Chinese medicine bone measurement scale, the distance on the body surface is proportionally calculated; Dynamic adaptation of body measurement: The distance between the inner creases when the middle phalanx of the user's middle finger is flexed is 1 inch to locate acupoints on the limbs. The transverse finger body measurement method is used for positioning on the abdomen and back. Geometric relationship deduction: Based on the known key points of the shoulder, elbow, and wrist, calculate the midpoint of the forearm, and locate the acupoints on the outer side by taking the knee joint as the origin and combining the angle between the thigh and the calf. AI-assisted localization optimization: Utilizing an improved mmPose network, it takes a 256×256 RGB image as input and outputs an acupoint heatmap and confidence score P. v Overlapping prediction boxes are filtered out by nonmaximum suppression.

[0010] As a preferred embodiment of the present invention, a skeleton connection is constructed based on skeletal key points, and the skeleton topology generation formula for calculating the connection probability between bones is as follows: ,in This represents the Euclidean distance between key points i and j, which intuitively reflects the actual physical distance between the two points in space. E represents the maximum Euclidean distance from keypoint i to the set of all keypoints v. When E > 0.05, it is considered a valid skeletal connection. The skin surface reconstruction formula is: ,in For the coordinates of the control points, For p-th degree B-spline basis functions; The bone measurement method for acupoint location is: acupoint coordinates = reference point + k × bone measurement unit length, where k is the converted cun (inches) from the acupoint to the reference point. ; The middle finger's body measurement method is 1 cun = the distance between the two transverse creases on the inner side of the middle joint of the middle finger when it is bent; The four-finger measurement method is 3 inches, which is the total width of the second joint of the four fingers held together.

[0011] As a preferred embodiment of the present invention, step S3 adaptive path planning includes: Digital coding of the environment: The physical environment is coded as a 0-1 matrix, with obstacle areas marked as 1 and feasible areas as 0; Establish a situation information model: Collect temperature and humidity, signal strength, and obstacle location parameters through sensors, and normalize them into situation characteristic values ​​with unified dimensions; Algorithm parameter initialization: Set the particle position, velocity, individual optimality, and global optimality for the particle swarm optimization algorithm, or set the mutation probability P using a genetic algorithm. m ; Feasible path space construction: Calculate the velocity constraint intersection Vr=V using the dynamic window method. d ∩V a ∩V s Candidate velocity trajectories are generated, and an initial path population is randomly generated using a swarm intelligence algorithm. Path evaluation function design: Design a comprehensive index for path length, obstacle avoidance safety, and smoothness; Environmental characteristics drive parameter adjustments; reduce heading angle weight for dense obstacles to avoid over-turning, and reduce distance weight for narrow gap obstacles to avoid jitter. Dynamic obstacle prediction and obstacle avoidance: Based on the least squares method, the trajectory of moving obstacles is fitted, and the predicted position is added to the map as a virtual static obstacle; Local path smoothing: B-spline curve interpolation is used to optimize path turning points, improve motion continuity, and provide emergency stop when the force applied to the robot arm's Z-axis exceeds a preset safety threshold.

[0012] As a preferred embodiment of the present invention, the comprehensive evaluation function formula for the candidate path is as follows: ,in The angle between the path and the target point, in rad. This is the distance from the path to the nearest obstacle, expressed in meters (m). Path speed, in m / s. , These are the weighting coefficients, and + , For normalization function, ; The global path economy and safety tradeoff is evaluated using the AAFMS evaluation function, the formula of which is: ,in Normalized value of path length , Normalized value of safe distance , Normalized value of the rate of change of steering angle , , and satisfy ; The formula for adjusting the weight of dense obstacles is as follows: ,in The distance between the current position and the nearest obstacle is expressed in meters (m). The distance threshold is 0.5m. and For attenuation coefficient, distance term Convert physical distance into a proportionality coefficient to avoid dimensional interference in attenuation calculations; specifically, These are the initial weighting coefficients, dimensionless, and range from 0.8 to 1.0. This is the distance sensitivity coefficient, dimensionless, with a value ranging from 2.0 to 5.0. and Used to control the rate at which weights decay with distance; The formula for adjusting the weight in a confined space is: ,in This represents the remaining distance from the current location to the target, in meters (m). The channel width characteristic value is expressed in meters (m). This converts the distance ratio into a dimensionless quantity, controlling the input range of the exponential function. Preferably, This refers to the minimum width of the feasible path in front of the robot, which is perceived and calculated in real time using LiDAR or a depth camera.

[0013] As a preferred embodiment of the present invention, the force-position hybrid control and safety mechanism in step S4 includes: Task space decomposition and mapping: The space is decomposed into a position control subspace and a force control subspace, and dynamically allocated; State prediction model construction: Based on the robot dynamics model, the pose and force coupling relationship is established by combining the Jacobian matrix to predict the motion and force state under different control strategies. At the same time, user vital sign data is integrated to dynamically update the model parameters and adapt to individual differences. Control Increment Calculation: Calculate the initial joint torque based on the desired contact force, and simultaneously compensate for force or position tracking errors in real time through a PI controller; Torque synthesis: The feedforward and feedback torques are superimposed to drive the joint motor, achieving coordinated control of position and force; Dynamic force threshold setting: Based on the tissue stiffness measured by ultrasound elastography, the maximum safe pressure is set individually, and the pressure is set to a three-level response.

[0014] Laser ranging warning: Real-time monitoring of the distance between the robotic arm and the human body; triggers path replanning when the distance is less than 50mm. Joint airbag protection: In the event of a sudden collision, the airbag at the joint expands instantly to absorb the impact force and reduce impact damage; Temperature monitoring: When the joint temperature is >65°C, the power is automatically reduced and air cooling is activated; Abnormal posture detection: The system uses an RGB-D camera to detect sudden user movement and suspends service until the posture stabilizes.

[0015] As a preferred embodiment of the present invention, the position control subspace is used to control the movement of the end effector of the robotic arm along a predetermined massage trajectory, and to correct the path deviation in real time through joint angle feedback; the force control subspace is used to adjust the applied force in the normal degree of freedom of contacting the human body, and to dynamically adjust the output torque in combination with feedback from a six-dimensional force sensor; the dynamic allocation is used to switch the control weight in real time according to task requirements. The three-level response includes mild overpressure, moderate overpressure, and sustained overpressure, when the pressure is >70%F. max The pressure was slightly too high, and a voice prompt indicated that the pressure was too high and to reduce it to 0.5F.max When the pressure is greater than 85%F max If the pressure is moderately high, immediately stop and retract 5mm.

[0016] As a preferred embodiment of the present invention, the dynamic interaction step S5 includes: Dynamic scanning of vital signs: The human body contour is reconstructed in real time using an RGB-D camera, combined with infrared thermal imaging to identify inflammatory areas, an electromyography sensor to detect muscle tension, and a six-dimensional force sensor to record the distribution of contact pressure, forming a real-time portrait of the user's vital signs. Integrated environmental parameters: Temperature and humidity sensors monitor environmental comfort and automatically adjust the massage chamber temperature; Voice command parsing: Supports natural language commands, parses intent, generates task tags, asks follow-up questions about massage details, and automatically skips confirmed parameter setting steps by combining the user's historical massage records to avoid repeated confirmation; Non-verbal interaction response: The massage pauses when the user turns over, the RGB-D camera automatically resumes service after repositioning, and the camera captures frowning or muscle contraction signals to reduce the intensity by 10% in real time; Sensitive point self-learning: Records user pain feedback on neighboring points and iteratively updates the optimal acupoint coordinates; Cloud-based knowledge base collaboration: Anonymous data is uploaded in encrypted form for federated learning to optimize the global model; Transparent progress updates: Voice prompts indicate the remaining time and the next action; Instant health report generation: A report is pushed to the app after the massage, which includes a muscle tone change graph, a comparison of inflamed areas, and home exercise suggestions.

[0017] 3. Beneficial effects

[0018] Compared with the prior art, the advantages of this invention are: (1) This invention reduces acupoint positioning error by using multimodal sensor fusion and attention mechanism weighting, dynamically sets user force threshold based on ultrasonic elastography, and reduces soft tissue damage rate by combining three-level response mechanism. At the same time, the adaptive path planning algorithm dynamically adjusts weights according to obstacle density to improve the passability in narrow spaces and improve path smoothness. In addition, edge computing compresses feature transmission to reduce system latency, and the emergency stop response is fast, which significantly ensures human-computer interaction safety.

[0019] (2) This invention models the acupoint muscle association through multimodal graph neural network, combines federated learning to encrypt and upload anonymous data to optimize the global model, improves the acupoint positioning coverage, and uses natural language parsing to link user commands with electromyographic signals to adjust the intensity in real time, solving the pain point of traditional solutions where pain cannot be responded to in time. After massage, a muscle tension change map and a home exercise suggestion report are generated to improve user rehabilitation compliance. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the steps of a massage robot massage path generation method according to the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0022] Example:

[0023] Please see Figure 1 A method for generating massage paths for a massage robot includes: S1, multimodal data acquisition and fusion; S2. Human body modeling and acupoint location; S3, Adaptive Path Planning; S4, Force-position hybrid control and safety mechanism; S5, dynamic interaction.

[0024] In specific embodiments of the present invention, multimodal sensor fusion and attention mechanism weighting reduce acupoint positioning errors. Based on ultrasonic elastography, the user force threshold is dynamically set, and combined with a three-level response mechanism, the soft tissue injury rate is reduced. At the same time, the adaptive path planning algorithm dynamically adjusts the weights according to the obstacle density, improving the passability in narrow spaces and the smoothness of the path. In addition, edge computing compresses feature transmission to reduce system latency, and the emergency stop response is fast, significantly ensuring the safety of human-computer interaction. Multimodal graph neural network is used to model the acupoint-muscle relationship, and federated learning is used to optimize the global model by encrypting and uploading anonymous data, significantly improving the acupoint positioning coverage. Natural language parsing of user commands and electromyographic signals are linked to adjust the force in real time, solving the pain point of the inability to respond to pain in traditional solutions. After massage, a muscle tension change map and a home exercise suggestion report are generated to improve user rehabilitation compliance.

[0025] Specifically, step S1, multimodal data acquisition and fusion, includes: Multi-source sensor collaborative acquisition: Visual data and biosignal acquisition, as well as dynamic environmental perception, are performed, and the acquisition and perception timestamps are synchronized to ensure spatiotemporal alignment of multi-sensor data. Visual data is acquired through RGB-D cameras to collect 3D point cloud data of the human back and limbs, obtaining body surface contour, spinal curvature, and body length spatial information. The body length spatial information is also the three-dimensional spatial information of the body. 33 skeletal key points are identified through the MediaPipe Pose model, muscle tension changes are monitored in real time through electromyography sensors, inflammation areas are located through infrared thermal imaging, contact force distribution is fed back through a six-dimensional force sensor, micro-movements of the human body are detected through an inertial measurement unit, and visual positioning ability is maintained in low-light environments through near-infrared supplementary lighting with an 850nm wavelength and a high-sensitivity IR camera. Preferably, in low-light environments, near-infrared supplementary lighting is used with a wavelength of 850nm, which, together with a high-sensitivity IR camera, maintains visual positioning capability, with an ISO sensitivity of >12800; Preferably, locating the inflammatory area through infrared thermal imaging means using infrared thermal imaging to identify areas with abnormal temperatures as an auxiliary basis for judging inflammation. Specifically, the MediaPipe Pose model identifies 33 skeletal key points by processing images captured by an RGB-D camera, identifying the two-dimensional pixel coordinates of 33 skeletal key points on the human back and limbs, and then mapping them to three-dimensional space by combining depth information.

[0026] Data cleaning and standardization: Visual data is denoised using statistical outlier filtering and normal vector smoothing. Force control data is denoised using Kalman filtering to remove mechanical vibration noise. Electromyography signals are extracted for frequency domain features by Fourier transform. Thermal imaging data is normalized to the [0,1] interval. Point cloud pose is registered using the ICP algorithm. Homogeneous transformation matrix T is used to compensate for human movement deviations to ensure spatiotemporal alignment. The homogeneous transformation matrix is ​​used to uniformly represent linear transformations such as rotation and translation as matrix multiplication. Preferably, the point clouds of two frames are registered using the ICP algorithm, the rotation matrix R and translation vector t reflecting human movement are solved, a homogeneous transformation matrix T is constructed, and the data of the next frame is transformed to the coordinate system of the previous frame through T, thereby achieving spatiotemporal alignment.

[0027] Cross-modal feature extraction: A lightweight YOLOv4 network is used to extract acupoint coordinates (u,v), which are mapped to 3D space. The amplitude of electromyography (EMG) signal ΔA is used to identify muscle tension, and the thermal imaging temperature gradient ∇T is used to mark the inflammatory area. B-spline surface is used to fit the back contour, and the normal velocity of the massage point is calculated to ensure vertical force application. Preferably, the root mean square value (RMS) of the EMG signal or its normalized amplitude is used as a quantitative indicator of muscle tension, which is more suitable for reflecting the degree of muscle activation. Feature-level fusion: Visual features H v With biological characteristics H p Acupoint-muscle association is modeled across modal graph neural networks using weighted fusion based on an attention mechanism; preferably, acupoints are treated as nodes in the graph, and node features include visual coordinates H. v and muscle tension H p Based on the theory of meridians in traditional Chinese medicine and the anatomical structure of muscles, the connection edges between nodes are preset to construct a graph structure; through the message passing mechanism of graph neural network (GNN), the deep relationship between acupoints and muscle state is learned, and finally more accurate acupoint locations and massage priorities are output. In one embodiment, the method for constructing the cross-modal graph neural network is as follows: each acupoint located in step S2 is taken as a node in the graph, and the initial feature vector of the node is determined by the visual coordinate feature H. v and the corresponding muscle tension characteristic H p It is composed of spliced ​​parts; based on the theory of meridians in traditional Chinese medicine and the human anatomical structure, the connection relationship between acupoints is preset as the edge of the graph; message passing is carried out through a 2-layer graph convolutional network (GCN) to update node features; finally, the node features output by the GNN are used to optimize the accurate coordinates of acupoints and calculate massage priority. For example, higher path planning weights are assigned to the area around acupoints with high muscle tension.

[0028] Decision-level fusion: When some modes fail, the independent model output is weighted and fused, and w1 and w2 are adjusted in real time according to the signal quality to avoid interference from single-mode noise; Real-time force control fusion: Six-dimensional force sensor data is input into a PID controller to dynamically adjust the trajectory normal velocity; Biofeedback optimization: Sudden changes in muscle signal amplitude trigger local path replanning, extending the massage duration of stiff areas; Edge computing acceleration: Feature extraction and preliminary fusion are completed at the acquisition end, and only compressed features are uploaded to the main control system.

[0029] In specific embodiments of this invention, visual, biological signal, and environmental data are acquired using multiple sensors, including an RGB-D camera, an electromyography (EMG) sensor, an infrared thermal imaging system, a six-dimensional force sensor, and an inertial measurement unit, all with synchronized timestamps. Statistical outlier filtering is used to denoise the point cloud, Kalman filtering removes mechanical vibration noise from the force control data, Fourier transform extracts the frequency domain features of the EMG signals, and the ICP algorithm is used to register the point cloud pose, ensuring spatiotemporal alignment of the multimodal data. Feature-level fusion uses an attention mechanism to weight visual and biological features, while decision-level fusion adjusts weights based on signal quality when some modalities fail, avoiding interference from single noise sources. Edge computing performs feature extraction and compression at the acquisition end, reducing latency in the main control system.

[0030] Specifically, the formula for visual coordinate normalization is: ,in The maximum and minimum values ​​of the point cloud on the X-axis. These represent the maximum and minimum values ​​of the point cloud along the y-axis. These represent the maximum and minimum values ​​of the point cloud along the z-axis.

[0031] The formula for normalizing biological signals is: ,in The original amplitude of the muscle signal. The average amplitude within the signal window. This represents the standard deviation of the amplitude.

[0032] The contact force normalization formula is: Where F is the contact force measured in real time by the six-dimensional force sensor and the unit is Newtons / N. Maximum contact force threshold, in Newtons per N. This is the minimum contact force threshold, and its unit is Newtons / N.

[0033] Attention-weighted fusion formula is , ,in The frequency domain characteristics are those of the muscle signal after Fourier transform. MLP(·) is a multilayer perceptron that maps frequency domain features to... In the same dimensional space, W and b are learnable weight matrices and biases, is the Sigmoid activation function, and it outputs weights in the range [0, 1]; || represents the feature vector concatenation operation.

[0034] When thermal imaging is obstructed, the weighted fusion independent model output formula is: , , ,in For visual location reliability, Inferring probabilities from biological signals The signal-to-noise ratio of each modal signal; Preferred, ∈[0,1] represents the maximum value of the acupoint heatmap output by the mmPose network, indicating the confidence level of visual localization; The output of a logistic regression model, determined by the amplitude of muscle signals and the temperature gradient, represents the probability that the area is an acupoint or discomfort point.

[0035] Force control PID compensation for contact force error The formula for converting to normal velocity compensation is as follows: , This is the normalized force error. Gain and dimensionless It is an integral gain and is dimensionless. The maximum normal velocity is expressed in m / s.

[0036] Preferred, , Determined by the Ziegler-Nichols tuning method or by empirical trial and error. The speed is typically set to 0.01 m / s to 0.05 m / s depending on the massage task.

[0037] In a specific embodiment of the present invention, visual coordinate normalization scales point cloud data to the [0,1] interval to eliminate scale differences; biosignal standardization eliminates individual differences based on the mean and standard deviation within the window; contact force normalization is personalized based on the maximum tolerance threshold of soft tissue and the Young's modulus of tissue measured by ultrasound elastography; attention-weighted fusion outputs [0,1] weights through the Sigmoid function to optimize feature contribution; when thermal imaging is occluded, multimodal output is weighted based on signal-to-noise ratio; force control PID converts contact force error into normal velocity compensation to achieve compliant control, unifies data dimensions, improves algorithm stability and generalization ability, personalized force control protects user safety, attention mechanism and weighted fusion enhance system fault tolerance, and PID compensation ensures force control accuracy and response speed.

[0038] Specifically, step S2, human body modeling and acupoint location, includes: Skeleton topology generation: Construct skeleton connections based on key points and calculate the connection probability between bones; Dynamic skeleton optimization: It uses IMU sensors to capture micro-movements caused by breathing and updates joint angles in real time; Skin surface reconstruction: Point cloud data is fitted using B-spline surfaces to generate a smooth skin model; Internal structure annotation: CT / MRI data is segmented, bones, muscles, and blood vessels are reconstructed, and a layered anatomical model is established. Preferably, the steps for establishing the layered anatomical model are as follows: standard human CT / MRI data is segmented beforehand, bones, muscles, and blood vessels are reconstructed, and a universal layered anatomical model database is established. During actual massage, the system acquires the user's body surface contour through non-contact sensing such as an RGB-D camera, registers it with the standard model in the database, and performs non-rigid deformation, thereby indirectly inferring the approximate location of the user's subcutaneous bones and muscles, providing a reference for force control and safety mechanisms. Bone measurement and positioning; based on the traditional Chinese medicine bone measurement scale, the distance on the body surface is proportionally calculated; Dynamic adaptation of body measurement: The distance between the inner creases when the middle phalanx of the user's middle finger is flexed is 1 inch to locate acupoints on the limbs. The transverse finger body measurement method is used for positioning on the abdomen and back. Geometric relationship deduction: Based on the known key points of the shoulder, elbow, and wrist, calculate the midpoint of the forearm, and locate the acupoints on the outer side by taking the knee joint as the origin and combining the angle between the thigh and the calf. AI-assisted localization optimization: Utilizing an improved mmPose network, it takes a 256×256 RGB image as input and outputs an acupoint heatmap and confidence score P. v Overlapping prediction boxes are filtered out by nonmaximum suppression.

[0039] The improved mmPose network adds an attention module after the backbone network of the standard mmPose network to enhance the feature extraction capability for small targets such as acupoints.

[0040] In a specific embodiment of the present invention, a skeleton topology is constructed based on skeletal key points, a skin model is fitted by B-spline surface, and acupoint locations are dynamically adapted by combining the traditional Chinese medicine bone measurement method. An improved mmPose network is used to generate acupoint heatmaps, and overlapping prediction boxes are filtered by non-maximum suppression. The bone measurement method makes the acupoint positioning error <3mm, which is better than traditional mechanical positioning. AI-assisted positioning adapts to individual body shape differences and improves coverage. The layered anatomical model supports precise force application to deep muscles.

[0041] Specifically, the skeleton topology generation formula for constructing skeleton connections based on skeletal keypoints and calculating the connection probability between bones is as follows: ,in This represents the Euclidean distance between key points i and j, which intuitively reflects the actual physical distance between the two points in space. E represents the maximum Euclidean distance from keypoint i to the set of all keypoints v. When E > 0.05, it is considered a valid skeletal connection. This threshold has been experimentally calibrated and is used to exclude abnormal connections caused by keypoint detection errors. The formula for skin surface reconstruction is: ,in For the coordinates of the control points, For p-th degree B-spline basis functions; The bone measurement method for acupoint location is: acupoint coordinates = reference point + k × bone measurement unit length, where k is the converted cun (inches) from the acupoint to the reference point. ; The middle finger's body measurement method is 1 cun = the distance between the two transverse creases on the inner side of the middle joint of the middle finger when it is bent; The four-finger measurement method is 3 inches, which is the total width of the second joint of the four fingers held together.

[0042] In a specific embodiment of the present invention, the skeleton topology generation formula calculates the connection probability based on the coordinates of adjacent key points and the local connected region map; the skin surface reconstruction formula generates a smooth surface based on B-spline basis functions and control point coordinates; the bone measurement positioning method calculates acupoint coordinates through reference points and conversion distances; the middle finger and four-finger body measurements provide dynamic scales; mathematical expression improves the computability and consistency of the model; B-spline surface ensures the smoothness of the skin model; and dynamic body measurement adaptation improves the flexibility of acupoint positioning.

[0043] Specifically, step S3, adaptive path planning, includes: Digital coding of the environment: The physical environment is coded as a 0-1 matrix, with obstacle areas marked as 1 and feasible areas as 0; Establish a situation information model: Collect temperature and humidity, signal strength, and obstacle location parameters through sensors, and normalize them into situation characteristic values ​​with unified dimensions; Algorithm parameter initialization: Set the particle position, velocity, individual optimality, and global optimality for the particle swarm optimization algorithm, or set the mutation probability P using a genetic algorithm. m ; Feasible path space construction: Calculate the velocity constraint intersection Vr=V using the dynamic window method. d ∩V a ∩V s Candidate velocity trajectories are generated, and an initial path population is randomly generated using a swarm intelligence algorithm. Path evaluation function design: Design a comprehensive index for path length, obstacle avoidance safety, and smoothness; Environmental characteristics drive parameter adjustments; reduce heading angle weight for dense obstacles to avoid over-turning, and reduce distance weight for narrow gap obstacles to avoid jitter. Dynamic obstacle prediction and obstacle avoidance: Based on the least squares method, the trajectory of moving obstacles is fitted, and the predicted position is added to the map as a virtual static obstacle; Local path smoothing: B-spline curve interpolation is used to optimize path inflection points, improving motion continuity, and an abrupt stop is implemented when the force applied to the robot arm's Z-axis exceeds a preset safety threshold, such as when the applied force F > Fz. max Emergency stop, of which F max The maximum safe pressure threshold is dynamically set based on the user's organizational stiffness.

[0044] In a specific embodiment of the present invention, the physical environment is encoded as a 0-1 matrix, marking obstacles and feasible areas. Parameters such as temperature, humidity, and signal strength are collected by sensors and normalized into situational feature values. Particle swarm optimization (PSO) parameters are initialized. The intersection of velocity constraints is calculated using a dynamic window method to generate candidate velocity trajectories. A comprehensive evaluation function is designed to balance path length, obstacle avoidance safety, and smoothness. The heading angle and distance weights are dynamically adjusted according to environmental characteristics. Dynamic obstacle trajectories are predicted based on the least squares method and treated as virtual static obstacles. B-spline curve interpolation is used to optimize path turning points. An emergency stop is implemented when the robotic arm applies excessive force, improving the real-time performance and adaptability of path planning. Dynamic weight adjustment optimizes path generation in complex environments. Obstacle prediction and B-spline smoothing enhance safety, and the emergency stop mechanism ensures user safety.

[0045] Specifically, the comprehensive evaluation function formula for candidate paths is as follows: ,in The angle between the path and the target point, in rad. This is the distance from the path to the nearest obstacle, expressed in meters (m). Path speed, in m / s. , These are the weighting coefficients, and + , For normalization function, ; The global path economy and safety tradeoff is evaluated using the AAFMS evaluation function, the formula of which is: ,in Normalized value of path length , Normalized value of safe distance , Normalized value of the rate of change of steering angle , , and satisfy ; The formula for adjusting the weight of dense obstacles is as follows: ,in The distance between the current position and the nearest obstacle is expressed in meters (m). The distance threshold is 0.5m. and For attenuation coefficient, distance term Convert physical distance into a proportionality coefficient to avoid dimensional interference in attenuation calculations; specifically, These are the initial weighting coefficients, dimensionless, and range from 0.8 to 1.0. This is the distance sensitivity coefficient, dimensionless, with a value ranging from 2.0 to 5.0. and Used to control the rate at which weights decay with distance; The formula for adjusting the weight in a confined space is: ,in This represents the remaining distance from the current location to the target, in meters (m). The channel width characteristic value is expressed in meters (m). This converts the distance ratio into a dimensionless quantity, controlling the input range of the exponential function. Preferably, This refers to the minimum width of the feasible path in front of the robot, which is perceived and calculated in real time using LiDAR or a depth camera.

[0046] In a specific embodiment of the present invention, the comprehensive evaluation function combines the angle between the path and the target point, the distance to the obstacle, and the speed, and calculates them by normalization and weighting. The global path economy and safety trade-off adopts the AAFMS function, which combines the path length, safety distance, and turning angle change rate. The weight adjustment of dense obstacles is based on the Euclidean distance decay distance term weight, and the weight adjustment of narrow space is based on the remaining distance and the channel width characteristic exponential decay distance term weight.

[0047] Specifically, step S4, the force-position hybrid control and safety mechanism, includes: Task space decomposition and mapping: The space is decomposed into a position control subspace and a force control subspace, and dynamically allocated; State prediction model construction: Based on the robot dynamics model, the pose and force coupling relationship is established by combining the Jacobian matrix to predict the motion and force state under different control strategies. At the same time, user vital sign data is integrated to dynamically update the model parameters and adapt to individual differences. Control Increment Calculation: Calculate the initial joint torque based on the desired contact force, and simultaneously compensate for force or position tracking errors in real time through a PI controller; Torque synthesis: The feedforward and feedback torques are superimposed to drive the joint motor, achieving coordinated control of position and force; Dynamic force threshold setting: Based on the tissue stiffness measured by ultrasound elastography, the maximum safe pressure is set individually, and the pressure is set to a three-level response.

[0048] In one specific embodiment, based on the tissue stiffness measured by ultrasonic elastography, a handheld ultrasonic elastography device can be used to perform a rapid pre-scan of the user's main massage areas before the massage begins. The measured tissue stiffness information is then entered into the system for personalized dynamic force threshold setting. In another specific embodiment, the system retrieves an approximate tissue stiffness model from a database as an initial setting based on the user's age, BMI, and other basic information. During the massage, data from a six-dimensional force sensor and displacement sensor are used to estimate the dynamic stiffness of local tissues online, serving as a supplementary basis for adjusting the dynamic force threshold setting.

[0049] Laser ranging warning: Real-time monitoring of the distance between the robotic arm and the human body; triggers path replanning when the distance is less than 50mm. Joint airbag protection: In the event of a sudden collision, the airbag at the joint expands instantly to absorb the impact force and reduce impact damage; Temperature monitoring: When the joint temperature is >65°C, the power is automatically reduced and air cooling is activated; Abnormal posture detection: The system uses an RGB-D camera to detect sudden user movement and suspends service until the posture stabilizes.

[0050] In a specific embodiment of the present invention, the task space is decomposed into a position control subspace and a force control subspace, control weights are dynamically allocated, a state prediction model is constructed based on the robot dynamics model and the Jacobian matrix, parameters are updated by integrating user vital sign data, initial joint torque is calculated based on the expected contact force, tracking errors are compensated by a PI controller, feedforward and feedback torques are superimposed to drive the joint motor, personalized safety pressure is set based on ultrasonic elastography, and a three-level response is implemented: laser ranging warning triggers replanning; joint airbags absorb collision impacts; temperature monitoring automatically reduces power; RGB-D camera detects abnormal postures and suspends service, achieving precise force-position coordinated control and multiple safety protections. The prediction model enhances system adaptability, the three-level response mechanism handles overpressure in stages, and multiple safety mechanisms ensure the safety of users and equipment.

[0051] Specifically, the position control subspace is used to control the movement of the robotic arm end effector along a predetermined massage trajectory and to correct path deviations in real time through joint angle feedback. The force control subspace is used to adjust the applied force in the normal degree of freedom of contacting the human body, and to dynamically adjust the output torque in combination with feedback from a six-dimensional force sensor. Dynamic allocation is used to switch control weights in real time according to task requirements. Level 3 response includes mild overpressure, moderate overpressure, and sustained overpressure, when the pressure is >70%F. max The pressure was slightly too high, and a voice prompt indicated that the pressure was too high and to reduce it to 0.5F. max When the pressure is greater than 85%F max If the pressure is moderately high, immediately stop and retract 5mm.

[0052] In a specific embodiment of the present invention, the position control subspace corrects path deviation through joint angle feedback, and the force control subspace dynamically adjusts the output torque in combination with feedback from a six-dimensional force sensor, dynamically allocating control weights that are switched in real time according to task requirements.

[0053] Specifically, step S5 dynamic interaction includes: Dynamic scanning of vital signs: The human body contour is reconstructed in real time using an RGB-D camera, combined with infrared thermal imaging to identify inflammatory areas, an electromyography sensor to detect muscle tension, and a six-dimensional force sensor to record the distribution of contact pressure, forming a real-time portrait of the user's vital signs. Integrated environmental parameters: Temperature and humidity sensors monitor environmental comfort and automatically adjust the massage chamber temperature; Voice command parsing: Supports natural language commands, parses intent, generates task tags, asks follow-up questions about massage details, and automatically skips confirmed parameter setting steps by combining the user's historical massage records to avoid repeated confirmation; Non-verbal interaction response: The massage pauses when the user turns over, the RGB-D camera automatically resumes service after repositioning, and the camera captures frowning or muscle contraction signals to reduce the intensity by 10% in real time; Sensitive point self-learning: Records user pain feedback on neighboring points and iteratively updates the optimal acupoint coordinates; Cloud-based knowledge base collaboration: Anonymous data is uploaded with encryption for federated learning to optimize the global model. Specifically, each massage robot terminal trains its local model using user massage data, such as acupoint coordinates, user feedback, and massage effects. Only the gradient update parameters of the model are encrypted and uploaded to the cloud server. The cloud server aggregates the gradients from multiple terminals, updates the global model, and then distributes the updated global model to each terminal. Throughout the entire process, the original user data remains locally, ensuring user privacy. Transparent progress updates: Voice prompts indicate the remaining time and the next action; Instant health report generation: A report is pushed to the app after the massage, which includes a muscle tone change graph, a comparison of inflamed areas, and home exercise suggestions.

[0054] In a specific embodiment of the present invention, the user's vital signs are scanned in real time using an RGB-D camera, infrared thermal imaging, electromyography (EMG) sensor, etc., to form a vital sign profile. Temperature and humidity sensors are integrated to adjust environmental comfort. Natural language command parsing is supported, task tags are generated and details are requested. Non-verbal interaction responses include turning over, frowning, etc., adjusting the intensity, recording user pain feedback, iteratively updating optimal acupoint coordinates, encrypting and uploading anonymous data, optimizing the global model through federated learning, providing voice feedback on progress and actions, and generating a health report after the massage, including changes in muscle tone, inflammation comparison, and exercise suggestions. This achieves natural and intelligent human-computer interaction and continuous learning. Multimodal perception comprehensively captures the user's state, voice and non-verbal interaction enhance convenience, a self-learning mechanism optimizes personalized services, and the health report provides added value.

[0055] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concept, should be covered within the scope of protection of the present invention.

Claims

1. A massage robot massage path generation method characterized by, include: S1. Multimodal data acquisition and fusion; S2. Human body modeling and acupoint location; S3, Adaptive Path Planning; S4, Force-position hybrid control and safety mechanism; S5, dynamic interaction.

2. The massage robot massage path generation method according to claim 1, characterized by, The step S1, multimodal data acquisition and fusion, includes: Multi-source sensor collaborative acquisition: Visual data and biosignal acquisition, as well as dynamic environmental perception, are performed, and the acquisition and perception timestamps are synchronized to ensure spatiotemporal alignment of multi-sensor data. Visual data is acquired through RGB-D cameras to collect 3D point cloud data of the human back and limbs, obtaining surface contour, spinal curvature, and body length spatial information. The body length spatial information is also the three-dimensional spatial information of the body. 33 skeletal key points are identified through the MediaPipe Pose model, muscle tension changes are monitored in real time through electromyography sensors, inflammation areas are located through infrared thermal imaging, contact force distribution is fed back through a six-dimensional force sensor, micro-movements of the human body are detected through an inertial measurement unit, and visual positioning capability is maintained in low-light environments through near-infrared supplementary lighting in conjunction with a high-sensitivity IR camera. Data cleaning and standardization: Visual data is denoised by point cloud based on statistical outlier filtering, and normal vector smoothing is used. Force control data is removed by Kalman filtering to remove mechanical vibration noise. Electromyography signals are extracted by Fourier transform to extract frequency domain features. Thermal imaging data is normalized to the [0,1] interval. ICP algorithm is used to register point cloud pose. Homogeneous transformation matrix T is used to compensate for human movement deviation to ensure spatiotemporal alignment. Cross-modal feature extraction: A lightweight YOLOv4 network is used to extract acupoint coordinates (u,v), which are mapped to 3D space. The amplitude of electromyography signal ΔA is used to identify muscle tension, and the thermal imaging temperature gradient ∇T is used to mark the inflammatory area. B-spline surface is used to fit the back contour, and the normal velocity of the massage point is calculated to ensure vertical force application. Feature-level fusion: visual feature H v With biological characteristics H p By attention mechanism weighted fusion, cross-modal graph neural network models acupoint muscle correlation; Decision-level fusion: When some modes fail, the independent model output is weighted and fused, and w1 and w2 are adjusted in real time according to the signal quality to avoid interference from single-mode noise; Real-time force control fusion: Six-dimensional force sensor data is input into a PID controller to dynamically adjust the trajectory normal velocity; Biofeedback optimization: Sudden changes in muscle signal amplitude trigger local path replanning, extending the massage duration of stiff areas; Edge computing acceleration: Feature extraction and preliminary fusion are completed at the acquisition end, and only compressed features are uploaded to the main control system.

3. The method for generating a massage path for a massage robot according to claim 2, characterized in that, The visual coordinate normalization formula is wherein is the maximum and minimum value of the point cloud in the X axis, is the maximum and minimum value of the point cloud in the y axis, is the maximum and minimum value of the point cloud in the z axis; The biological signal standardization formula is wherein is the original amplitude of the muscle signal, is the amplitude mean value within the signal window, is the amplitude standard deviation; The contact force normalization formula is where F is the contact force measured in real time by the six-dimensional force sensor and has units of Newtons / N, is a maximum contact force threshold value and has units of Newtons / N, is a minimum contact force threshold value and has units of Newtons / N; Attention-weighted fusion formula is , ,in The frequency domain characteristics are those of the muscle signal after Fourier transform. MLP(·) is a multilayer perceptron that maps frequency domain features to... In the same dimensional space, W and b are learnable weight matrices and biases, is the Sigmoid activation function, and it outputs weights in the range [0, 1]; || represents the feature vector concatenation operation; When thermal imaging is obstructed, the weighted fusion independent model output formula is: , , ,in For visual location reliability, Inferring probabilities from biological signals The signal-to-noise ratio of each modal signal; Force control PID compensation for contact force error The formula for converting to normal velocity compensation is as follows: , This is the normalized force error. It is a proportional gain and dimensionless. It is an integral gain and is dimensionless. The maximum normal velocity is expressed in m / s.

4. The method for generating a massage path for a massage robot according to claim 3, characterized in that, Step S2, human body modeling and acupoint location, includes: Skeleton topology generation: Construct skeleton connections based on key points and calculate the connection probability between bones; Dynamic skeleton optimization: It uses IMU sensors to capture micro-movements caused by breathing and updates joint angles in real time; Skin surface reconstruction: Point cloud data is fitted using B-spline surfaces to generate a smooth skin model; Internal structure annotation: Segment CT / MRI data, reconstruct bones, muscles and blood vessels, and establish a layered anatomical model; Bone measurement and positioning; based on the traditional Chinese medicine bone measurement scale, the distance on the body surface is proportionally calculated; Dynamic adaptation of body measurement: The distance between the inner creases when the middle phalanx of the user's middle finger is flexed is 1 inch to locate acupoints on the limbs. The transverse finger body measurement method is used for positioning on the abdomen and back. Geometric relationship deduction: Based on the known key points of the shoulder, elbow, and wrist, calculate the midpoint of the forearm, and locate the acupoints on the outer side by taking the knee joint as the origin and combining the angle between the thigh and the calf. AI-assisted positioning optimization: use improved mmPose network, input 256x256 RGB image, output acupoint heat map and confidence P v , filter overlapping prediction boxes through non-maximum suppression.

5. The method for generating a massage path for a massage robot according to claim 4, characterized in that, The skeleton topology generation formula, which constructs skeleton connections based on skeletal keypoints and calculates the connection probability between bones, is as follows: ,in , This represents the Euclidean distance between key points i and j, which intuitively reflects the actual physical distance between the two points in space. E represents the maximum Euclidean distance from keypoint i to the set of all keypoints v. When E > 0.05, it is considered a valid skeletal connection. The skin surface reconstruction formula is: ,in For the coordinates of the control points, For p-th degree B-spline basis functions; The bone measurement method for acupoint location is: acupoint coordinates = reference point + k × bone measurement unit length, where k is the converted cun (inches) from the acupoint to the reference point. ; The middle finger's body measurement method is 1 cun = the distance between the two transverse creases on the inner side of the middle joint of the middle finger when it is bent; The four-finger measurement method is 3 inches, which is the total width of the second joint of the four fingers held together.

6. The method for generating a massage path for a massage robot according to claim 5, characterized in that, The adaptive path planning step S3 includes: Digital coding of the environment: The physical environment is coded as a 0-1 matrix, with obstacle areas marked as 1 and feasible areas as 0; Establish a situation information model: Collect temperature and humidity, signal strength, and obstacle location parameters through sensors, and normalize them into situation characteristic values ​​with unified dimensions; Algorithm parameter initialization: set the particle swarm algorithm particle position, velocity, individual optimal and global optimal, or use genetic algorithm to set mutation probability P m ; Feasible path space construction: use dynamic window method to calculate speed constraint intersection Vr=V d ∩V a ∩V s , generate candidate speed trajectory, use swarm intelligence algorithm to randomly generate initial path population; Path evaluation function design: Design a comprehensive index for path length, obstacle avoidance safety, and smoothness; Environmental characteristics drive parameter adjustments; reduce heading angle weight for dense obstacles to avoid over-turning, and reduce distance weight for narrow gap obstacles to avoid jitter. Dynamic obstacle prediction and obstacle avoidance: Based on the least squares method, the trajectory of moving obstacles is fitted, and the predicted position is added to the map as a virtual static obstacle; Local path smoothing: B-spline curve interpolation is used to optimize path turning points, improve motion continuity, and provide emergency stop when the force applied to the robot arm's Z-axis exceeds a preset safety threshold.

7. The method for generating a massage path for a massage robot according to claim 6, characterized in that, The comprehensive evaluation function formula for the candidate path is as follows: ,in The angle between the path and the target point, expressed in rad. This is the distance from the path to the nearest obstacle, expressed in meters (m). Path speed, in m / s. , These are the weighting coefficients, and + , For normalization function, ; The global path economy and safety tradeoff is evaluated using the AAFMS evaluation function, the formula of which is: ,in Normalized value of path length , Normalized value of safe distance , Normalized value of the rate of change of steering angle , , and satisfy ; The formula for adjusting the weight of dense obstacles is as follows: ,in The distance between the current position and the nearest obstacle is expressed in meters (m). With a distance threshold of 0.5m, the distance term Convert the physical distance into a proportionality coefficient to avoid dimensional interference in attenuation calculations; These are the initial weighting coefficients, dimensionless. The distance sensitivity coefficient is dimensionless. and Used to control the rate at which weights decay with distance; The formula for adjusting the weight in a confined space is: ,in This represents the remaining distance from the current location to the target, in meters (m). The channel width characteristic value is expressed in meters (m). This converts the distance ratio into a dimensionless quantity, controlling the input range of the exponential function.

8. The method for generating a massage path for a massage robot according to claim 7, characterized in that, The force-position hybrid control and safety mechanism in step S4 includes: Task space decomposition and mapping: The space is decomposed into a position control subspace and a force control subspace, and dynamically allocated; State prediction model construction: Based on the robot dynamics model, the pose and force coupling relationship is established by combining the Jacobian matrix to predict the motion and force state under different control strategies. At the same time, user vital sign data is integrated to dynamically update the model parameters and adapt to individual differences. Control Increment Calculation: Calculate the initial joint torque based on the desired contact force, and simultaneously compensate for force or position tracking errors in real time through a PI controller; Torque synthesis: The feedforward and feedback torques are superimposed to drive the joint motor, achieving coordinated control of position and force; Dynamic force threshold setting: Based on the tissue stiffness measured by ultrasound elastography, the maximum safe pressure is set individually, and the pressure is set to a three-level response. Laser ranging warning: Real-time monitoring of the distance between the robotic arm and the human body; triggers path replanning when the distance is less than 50mm. Joint airbag protection: In the event of a sudden collision, the airbag at the joint expands instantly to absorb the impact force and reduce impact damage; Temperature monitoring: When the joint temperature is >65°C, the power is automatically reduced and air cooling is activated; Abnormal posture detection: The system uses an RGB-D camera to detect sudden user movement and suspends service until the posture stabilizes.

9. A method for generating a massage path for a massage robot according to claim 8, characterized in that, The position control subspace is used to control the movement of the robotic arm end effector along a predetermined massage trajectory and to correct path deviations in real time through joint angle feedback. The force control subspace is used to adjust the applied force in the normal degree of freedom of contacting the human body and to dynamically adjust the output torque in combination with feedback from a six-dimensional force sensor. The dynamic allocation is used to switch control weights in real time according to task requirements. The three level response includes light overpressure, moderate overpressure and sustained overpressure, when pressure > 70% F max Light overpressure, voice prompt pressure is too large, reduce to 0.5 F max Moderate overpressure, immediately emergency stop and retract 5mm max when pressure > 85% F 10. A method for generating a massage path for a massage robot according to claim 9, characterized in that, The dynamic interaction in step S5 includes: Dynamic scanning of vital signs: The human body contour is reconstructed in real time using an RGB-D camera, combined with infrared thermal imaging to identify inflammatory areas, an electromyography sensor to detect muscle tension, and a six-dimensional force sensor to record the distribution of contact pressure, forming a real-time portrait of the user's vital signs. Integrated environmental parameters: Temperature and humidity sensors monitor environmental comfort and automatically adjust the massage chamber temperature; Voice command parsing: Supports natural language commands, parses intent, generates task tags, asks follow-up questions about massage details, and automatically skips confirmed parameter setting steps by combining the user's historical massage records to avoid repeated confirmation; Non-verbal interaction response: The massage pauses when the user turns over, the RGB-D camera automatically resumes service after repositioning, and the camera captures frowning or muscle contraction signals to reduce the intensity by 10% in real time; Sensitive point self-learning: Records user pain feedback on neighboring points and iteratively updates the optimal acupoint coordinates; Cloud-based knowledge base collaboration: Anonymous data is uploaded in encrypted form for federated learning to optimize the global model; Transparent progress updates: Voice prompts indicate the remaining time and the next action; Instant health report generation: A report is pushed to the app after the massage, which includes a muscle tone change graph, a comparison of inflamed areas, and home exercise suggestions.

Images