Exoskeleton robot rehabilitation training control method and system based on motion end point distribution

Abstract

The application discloses a kind of exoskeleton robot rehabilitation training control method and system based on motion end point distribution, the method collects the limb movement data of patient same task multiple attempts, extracts motion end point;Grouped according to patient, affected side, task type, filters time sequence front effective end to establish history set;Adaptive end point tolerance domain is constructed by tolerance center determined by median of historical end point, quantile deviation combined with minimum half-width and safety margin;Calculate each axis out-of-bound amount and define out-of-bound cost, construct continuous tolerance potential based on out-of-bound amount, generate recovery action pointing to tolerance domain;Recovery action is mapped as joint torque, correction amount or virtual guide force output control instruction;Update history end point set after training iteration tolerance domain parameter.The system includes exoskeleton and controller, and multiple modules are internally arranged to realize the above process.The application has the characteristics that it can maintain zero additional correction within the tolerance range, and only generates constrained recovery action for out-of-bound part.

Description

Technical Field

[0001] This invention relates to the field of intelligent rehabilitation robots, and specifically to a rehabilitation training control method and system for exoskeleton robots based on the distribution of motion endpoints. Background Technology

[0002] With the deepening of population aging and the increase in the number of people with upper limb injuries and neurological dysfunction, the demand for upper limb rehabilitation training, exoskeleton assistance, and human-machine collaborative training is growing. Whether the hand, joint, or device end effector can reach a reasonable target area is an important basis for evaluating the quality of movement completion and determining the degree of assistance. Due to differences in the physiological structure, motor ability, and rehabilitation stage of different users, it is difficult to use completely consistent target positions or training parameters for the same task among different individuals.

[0003] In existing technologies, preset target points, standard motion trajectories, fixed distance thresholds, or fixed stiffness potential fields are typically used to determine motion deviations and generate auxiliary effects. However, even when the same user repeatedly performs the same task, the endpoint of their motion will fluctuate normally; the endpoint distribution also differs depending on the side of motion, target direction, task type, and training stage. Treating any deviation as an error requiring correction can easily lead to continuous fine-tuning and unnecessary intervention; while uniformly using a wide fixed threshold may reduce the ability to identify actual boundary violations.

[0004] Some rehabilitation or assistive control systems already collect motion data using joint angle sensors, inertial measurement units, electromyography sensors, and visual posture estimation modules, and evaluate training effectiveness based on movement amplitude, duration, or overall error. These methods often focus on a single movement process or a pre-set reference trajectory, failing to fully utilize the individualized distribution characteristics of historical effective movement endpoints under the same task conditions. They also struggle to continuously update the reasonable landing point range as historical samples accumulate and training stages change. Furthermore, the measurement results of the movement endpoint may be affected by sensor noise, spatial registration errors, and target estimation uncertainties. Directly mapping all positional deviations to control inputs can easily lead to frequent starts and stops, abrupt output changes, or jitter near the boundaries. Summary of the Invention

[0005] To address the aforementioned shortcomings, this invention proposes a rehabilitation training control method and system for exoskeleton robots based on motion endpoint distribution, which can maintain zero additional correction within a tolerable range and continuously generate constrained recovery effects only for out-of-bounds portions.

[0006] The present invention provides a rehabilitation training control method for exoskeleton robots based on motion endpoint distribution, comprising the following steps:

[0007] Step S1, Data Acquisition: The motion detection device of the exoskeleton robot acquires limb motion data of the same patient under the same rehabilitation training task for multiple completed training trials, preprocesses each limb motion data, and extracts the limb motion endpoint from the end of the active motion segment of each training trial.

[0008] Step S2, Historical Set Establishment: The limb motor endpoints are grouped according to grouping conditions including patient, affected side, training target location or training task type. For the current training trial, a historical valid endpoint set corresponding to the training task group is established using limb motor endpoints that are earlier in time than the current training trial and have passed quality screening.

[0009] Step S3, Endpoint Tolerance Domain Construction: Determine the tolerance center based on the median of the historical valid endpoint set in each coordinate dimension, and determine the tolerance half-width based on the empirical quantile of the absolute deviation of the historical valid endpoints relative to the tolerance center in each coordinate dimension, and construct the endpoint tolerance domain with the tolerance center and the tolerance half-width.

[0010] Step S4, Calculate the cross-axis out-of-bounds amount: Obtain the current limb movement endpoint of the current training trial, calculate the cross-axis out-of-bounds amount of the current limb movement endpoint relative to the endpoint tolerance domain, make the out-of-bounds amount corresponding to the coordinate component located within the endpoint tolerance domain zero, and retain only the actual out-of-bounds part of the coordinate component that crosses the boundary of the endpoint tolerance domain, and further define the endpoint out-of-bounds cost to quantify the degree to which the current landing point exceeds the historical reasonable distribution.

[0011] Step S5, Tolerance potential energy generation: Based on the axial crossover amount, a tolerance potential energy is constructed, and a recovery effect pointing to the endpoint tolerance domain is generated according to the tolerance potential energy, so that the tolerance potential energy and the recovery effect are zero in the endpoint tolerance domain and continuously increase with the axial crossover amount after the crossover.

[0012] Step S6, Control Output: Based on the controller of the exoskeleton robot, the recovery action is mapped to the exoskeleton robot joint reference torque, joint control input correction amount, assistance level or virtual guiding force acting on the human limb end, and exoskeleton robot control commands are generated.

[0013] Step S7, History Update: After the current training trial is completed, the limb movement data is quality filtered, and only the current limb movement endpoints that pass the quality filter are written into the historical valid endpoint set of the corresponding training task group. The tolerance center and the tolerance half-width are updated based on the updated historical valid endpoint set.

[0014] The above-mentioned exoskeleton robot rehabilitation training control method based on the distribution of motion endpoints, wherein, in step S1, the acquisition of the limb motion endpoint includes: collecting limb motion data from the joint position sensors, encoders, inertial measurement units, depth cameras, optical motion capture systems, or visual pose estimation modules of the exoskeleton robot, and representing it as the limb task spatial position, limb joint spatial state, or exoskeleton robot end effector state that changes over time; the preprocessing includes missing value processing, filtering, active motion segment identification, or quality screening; the limb motion endpoint is the mean of N consecutive valid sampling points selected from the end of the active motion segment, where N is a positive integer.

[0015] In the above-mentioned exoskeleton robot rehabilitation training control method based on the distribution of movement endpoints, in step S2, the historical effective endpoint set is grouped according to the patient, affected side, training target position or training task type, and the grouping conditions also include at least one of the exoskeleton robot state or patient rehabilitation training stage; the historical effective endpoint set does not include the limb movement endpoint of the current training trial and the limb movement endpoint that is later than the current training trial.

[0016] The above-mentioned exoskeleton robot rehabilitation training control method based on motion endpoint distribution, wherein, in step S3, the historical valid endpoint set is set as follows: , The tolerance center and the tolerance half-width are determined according to the following formulas:

[0017]

[0018]

[0019] Where, p i Let R represent the i-th valid historical endpoint, n represent the number of valid historical endpoints, and Ri represent the number of valid historical endpoints. d Let p represent a d-dimensional real space, where d represents the coordinate dimension of the endpoint of the motion. i，j c represents the j-th dimension coordinate of the i-th valid historical endpoint. j Let j represent the j-th dimension coordinate of the tolerance center, median i h represents taking the median of the data corresponding to index i. j Let Qq represent the tolerance half-width in the j-th dimension, Qq represent the empirical q-quantile, and q represent the quantile level. Minimum tolerance half-width, For safety, the endpoint tolerance domain satisfies ch≤p≤c+h, where c represents the tolerance center, h represents the vector composed of the tolerance half-widths of each dimension, and p represents the limb movement endpoint.

[0020] In the above-mentioned exoskeleton robot rehabilitation training control method based on motion endpoint distribution, in step S4, the axis-by-axis out-of-bounds amount and endpoint out-of-bounds cost are determined according to the following formulas:

[0021]

[0022]

[0023] in, The term "axis-by-axis out-of-bounds" indicates the value of the cross-axis, "sign" indicates the sign function, "⊙" indicates element-by-element multiplication, and "|·|" indicates element-by-element absolute value. Let represent the cost of the endpoint going out of bounds, and ||·|2 represent the L2 norm.

[0024] In the above-described exoskeleton robot rehabilitation training control method based on motion endpoint distribution, in step S5, the tolerance potential energy and the recovery effect are determined according to the following formulas:

[0025]

[0026]

[0027] in, This represents the positive definite diagonal gain matrix, where each diagonal gain... U(p) represents the tolerable potential energy at the current limb movement endpoint p. express The transpose of F(p) represents the recovery effect. This represents the gradient operator with respect to the current limb movement endpoint p.

[0028] The above-described exoskeleton robot rehabilitation training control method based on motion endpoint distribution, wherein, in step S6, the mapping result of the recovery effect includes the exoskeleton robot joint reference torque, the exoskeleton robot joint control input correction amount, the assistance level, or the virtual guiding force applied to the patient's limb; when the joint reference torque mapping is performed using a Jacobian matrix, the reference torque is... ,in, The reference torque is represented by J, the Jacobian matrix is ​​represented by F, and the recovery effect is represented by F. The output constraint processing includes amplitude limiting, low-pass smoothing, rate of change limiting, boundary hysteresis, dead zone expansion, or minimum trigger time.

[0029] The above-mentioned exoskeleton robot rehabilitation training control method based on the distribution of motion endpoints includes, in step S7, the quality screening includes integrity, missing value ratio, motion amplitude, motion segment length, and sensor effectiveness checks; when the number of historical valid endpoints in the corresponding task group is insufficient, the group parameters or preset cold start parameters are used as the tolerance center and the tolerance half-width, and the tolerance center and the tolerance half-width are updated as the number of motion endpoints that pass the quality screening increases, so as to transition from the group parameters or preset cold start parameters to individualized parameters.

[0030] A rehabilitation training control system for an exoskeleton robot based on motion endpoint distribution, used to execute the control method, is characterized by comprising an exoskeleton robot and a controller communicatively connected to the exoskeleton robot; the exoskeleton robot includes a motion detection device and a joint drive component; the controller includes a historical endpoint processing module, a tolerance domain construction module, an out-of-bounds calculation module, and an auxiliary control module; the historical endpoint processing module is used to establish a historical valid endpoint set based on limb motion data collected by the motion detection device; the tolerance domain construction module is used to construct an endpoint tolerance domain based on the median and absolute deviation empirical quantiles of the historical valid endpoint set in each coordinate dimension; the out-of-bounds calculation module is used to calculate the axial out-of-bounds amount of the current limb motion endpoint relative to the endpoint tolerance domain; the auxiliary control module is used to maintain zero additional assistance when the current limb motion endpoint is within the endpoint tolerance domain, and generate continuous recovery action based on the actual out-of-bounds portion when the current limb motion endpoint crosses the boundary of the endpoint tolerance domain, mapping the continuous recovery action to a joint reference torque, joint control input correction, or assistance level output to the joint drive component.

[0031] Compared with the prior art, the present invention has significant advantages. As can be seen from the above solutions, the present invention has the following advantages:

[0032] (1) By using the median and empirical quantile to estimate the center and tolerance half-width, and combining the minimum tolerance half-width, safety margin and cold start parameter, the impact of a small number of abnormal trials, insufficient samples, sensor noise and spatial registration error on boundary estimation can be reduced, and the robustness and interpretability of tolerance range calculation can be improved.

[0033] (2) When the current motion endpoint is within the tolerance domain, the outbound vector, tolerance potential energy and additional correction are all kept to zero; only when the endpoint exceeds the boundary is the recovery action generated according to the actual outbound part, which helps to preserve the user's autonomy and reduce unnecessary continuous intervention.

[0034] (3) The flexible correction output can be mapped to virtual guiding force, reference torque, control input correction amount, etc. according to the device interface, which is convenient to connect to exoskeletons, rehabilitation or collaborative robots and training prompt terminals, and has good hardware compatibility.

[0035] (4) By updating the history of newly completed and qualified exercise trials, the tolerance center and tolerance half-width can be gradually transitioned from group or preset cold start parameters to individual parameters, thereby adapting to the accumulation of training data and changes in user status; the endpoint out-of-bounds vector and out-of-bounds cost provide calculable and traceable quantitative basis for training evaluation.

[0036] The following specific embodiments further illustrate the beneficial effects of the present invention. Attached Figure Description

[0037] Figure 1 This is a schematic diagram of the method flow of the present invention;

[0038] Figure 2 This is a schematic diagram of the control system structure of the present invention. Detailed Implementation

[0039] The following, in conjunction with the accompanying drawings and preferred embodiments, details the specific implementation, features, and effects of an exoskeleton robot rehabilitation training control method and system based on motion endpoint distribution proposed in accordance with the present invention.

[0040] like Figure 1-2 As shown, the present invention provides an exoskeleton robot rehabilitation training control method and system based on motion endpoint distribution, comprising:

[0041] Step S1, Data Acquisition: The motion detection device of the exoskeleton robot acquires limb motion data of the same patient under the same rehabilitation training task for multiple completed training trials, preprocesses each limb motion data, and extracts the limb motion endpoint from the end of the active motion segment of each training trial.

[0042] Step S2, Historical Set Establishment: The limb motor endpoints are grouped according to grouping conditions including patient, affected side, motor side, training target location or training task type. For the current training trial, a historical valid endpoint set corresponding to the training task group is established using limb motor endpoints that are earlier in time than the current training trial and have passed quality screening.

[0043] Step S3, Endpoint Tolerance Domain Construction: Determine the tolerance center based on the median of the historical valid endpoint set in each coordinate dimension, and determine the tolerance half-width based on the empirical quantile of the absolute deviation of the historical valid endpoints relative to the tolerance center in each coordinate dimension, and construct the endpoint tolerance domain with the tolerance center and the tolerance half-width.

[0044] Step S4, Calculate the cross-axis out-of-bounds amount: Obtain the current limb movement endpoint of the current training trial, calculate the cross-axis out-of-bounds amount of the current limb movement endpoint relative to the endpoint tolerance domain, make the out-of-bounds amount corresponding to the coordinate component located within the endpoint tolerance domain zero, and retain only the actual out-of-bounds part of the coordinate component that crosses the boundary of the endpoint tolerance domain, and further define the endpoint out-of-bounds cost to quantify the degree to which the current landing point exceeds the historical reasonable distribution.

[0045] Step S5, Tolerance potential energy generation: Based on the axial crossover amount, a tolerance potential energy is constructed, and a recovery effect pointing to the endpoint tolerance domain is generated according to the tolerance potential energy, so that the tolerance potential energy and the recovery effect are zero in the endpoint tolerance domain and continuously increase with the axial crossover amount after the crossover.

[0046] Step S6, Control Output: Based on the controller of the exoskeleton robot, the recovery action is mapped to the exoskeleton robot joint reference torque, joint control input correction amount, assistance level or virtual guiding force acting on the human limb end, and exoskeleton robot control commands are generated.

[0047] Step S7, History Update: After the current training trial is completed, the limb movement data is quality filtered, and only the current limb movement endpoints that pass the quality filter are written into the historical valid endpoint set of the corresponding training task group. The tolerance center and the tolerance half-width are updated based on the updated historical valid endpoint set.

[0048] The method can map the restorative effect as virtual guiding force, reference torque, joint control input correction amount, control weight, feedback intensity or assistance level, etc., and can be connected to exoskeleton rehabilitation equipment.

[0049] This invention is applied to the rehabilitation training control method and system of exoskeleton robots. At the end point of hand or joint movement, a task-related tolerance domain can be constructed. By adapting to the accumulation of training data and changes in the user's state, the flexible correction output can be guided to gradually transition from group or preset cold start parameters to individualized parameters. The endpoint out-of-bounds vector and out-of-bounds cost can provide calculable and traceable quantitative basis for training evaluation.

[0050] Effect Analysis:

[0051] To validate this method, the three types of data used are all publicly available datasets, named as follows: Hand MotionCatalog (formal name: Hand Motion Catalog of Human Center-Out TransportTrajectories Measured Redundantly in 3D Task-Space), OpenPack (formal name: OpenPack: A Large-Scale Dataset for Recognizing Packaging Works in IoT-Enabled Logistic Environments), and Upper Limb Torque Prediction from EMGSignals (formal name: A Dataset for the Investigation of Upper Limb Torque Prediction from EMG Signals).

[0052] In the Hand Motion Catalog, filtered and aligned 3D out-of-center transport data of the hand were used to filter low-quality trials and trials with an anomalous number of groups. Trials were grouped by participant, movement side, and target category, and the motion endpoint was constructed using the mean of the 10 valid samples at the end of each trial. This dataset was used for primary validation of endpoint tolerance centers, quantile boundaries, out-of-bounds vectors, and endpoint out-of-bounds costs.

[0053] In OpenPack, repetitive logistics packaging operation segments are segmented based on publicly available operation annotations. Robust motion endpoints of the left or right wrist at the end of each segment are extracted from 3D skeleton data and grouped according to subject, session, operation category, and wrist side. This dataset is used to demonstrate that the same endpoint layer construction process can be transferred to real repetitive operation data. Since this dataset does not provide anomaly labels directly corresponding to the endpoint out-of-bounds definition of this invention, out-of-bounds results are not expressed as fault detection rate or diagnostic accuracy.

[0054] Upper Limb Torque Prediction from EMG Signals does not participate in endpoint tolerance boundary training. Shoulder and elbow joint torque data from their publicly available multi-joint tasks are read to construct an external measured torque magnitude envelope, which is used to assist in calibrating the gain matrix K, reference torque mapping coefficients, and output upper limit, ensuring that the torque proxy generated by endpoint out-of-bounds events remains within a conservative range.

[0055] The above three types of publicly available data form a verification sequence of "endpoint distribution modeling - real operation endpoint migration - output magnitude verification". The verification objects all revolve around the tolerance potential field of the endpoint layer and do not involve the interpretation of the boundary of the complete motion process or the state of different historical memory lengths.

[0056] Table 1 Comparison of Mean Squared Correction Costs for Different Endpoint Constraint Strategies

[0057]

[0058] In the Hand Motion Catalog endpoint experiment, the mean of the last 10 sampling points of each motion trial was first taken as the 3D motion endpoint. Then, the endpoints were grouped by participant, motion side, and target category. Within each group, the top 70% of valid endpoints were used to construct the center and 95% of the quantile boundaries, with the remaining 30% as a hold-out test. A total of 1610 evaluable endpoints were obtained. The mean squared correction costs for the precise target point, the fixed tolerance domain, and the adaptive endpoint tolerance domain of this invention were 43.61, 17.01, and 3.88, respectively. Compared to the fixed tolerance domain, the mean squared correction cost of this invention was reduced by 77.21%. This result demonstrates that the adaptive quantile boundary after task grouping can reduce the out-of-bounds burden caused by a single fixed boundary.

[0059] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A method for controlling rehabilitation training of an exoskeleton robot based on the distribution of motion endpoints, characterized in that, Includes the following steps: Step S1, Data Acquisition: The motion detection device of the exoskeleton robot acquires limb motion data of the same patient under the same rehabilitation training task for multiple completed training trials, preprocesses each limb motion data, and extracts the limb motion endpoint from the end of the active motion segment of each training trial. Step S2, Historical Set Establishment: The limb motor endpoints are grouped according to grouping conditions including patient, affected side, training target location or training task type. For the current training trial, a historical valid endpoint set corresponding to the training task group is established using limb motor endpoints that are earlier in time than the current training trial and have passed quality screening. Step S3, Endpoint Tolerance Domain Construction: Determine the tolerance center based on the median of the historical valid endpoint set in each coordinate dimension, and determine the tolerance half-width based on the empirical quantile of the absolute deviation of the historical valid endpoints relative to the tolerance center in each coordinate dimension, and construct the endpoint tolerance domain with the tolerance center and the tolerance half-width. Step S4, Calculate the cross-axis out-of-bounds amount: Obtain the current limb movement endpoint of the current training trial, calculate the cross-axis out-of-bounds amount of the current limb movement endpoint relative to the endpoint tolerance domain, make the out-of-bounds amount corresponding to the coordinate component located within the endpoint tolerance domain zero, and retain only the actual out-of-bounds part of the coordinate component that crosses the boundary of the endpoint tolerance domain, and further define the endpoint out-of-bounds cost to quantify the degree to which the current landing point exceeds the historical reasonable distribution. Step S5, Tolerance potential energy generation: Based on the axial crossover amount, a tolerance potential energy is constructed, and a recovery effect pointing to the endpoint tolerance domain is generated according to the tolerance potential energy, so that the tolerance potential energy and the recovery effect are zero in the endpoint tolerance domain and continuously increase with the axial crossover amount after the crossover. Step S6, Control Output: Based on the controller of the exoskeleton robot, the recovery action is mapped to the exoskeleton robot joint reference torque, joint control input correction amount, assistance level or virtual guiding force acting on the human limb end, and exoskeleton robot control commands are generated. Step S7, History Update: After the current training trial is completed, the limb movement data is quality filtered, and only the current limb movement endpoints that pass the quality filter are written into the historical valid endpoint set of the corresponding training task group. The tolerance center and the tolerance half-width are updated based on the updated historical valid endpoint set.

2. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 1, characterized in that, In step S1, the acquisition of the limb movement endpoint includes: collecting limb movement data from the joint position sensors, encoders, inertial measurement units, depth cameras, optical motion capture systems, or visual pose estimation modules of the exoskeleton robot, and representing it as the limb task spatial position, limb joint spatial state, or exoskeleton robot end effector state that changes over time; the preprocessing includes missing value processing, filtering, active motion segment identification, or quality screening; the limb movement endpoint is the mean of N consecutive valid sampling points selected from the end of the active motion segment, where N is a positive integer.

3. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 1, characterized in that, In step S2, the historical valid endpoint set is grouped according to patient, affected side, training target location or training task type, and the grouping conditions also include at least one of exoskeleton robot status or patient rehabilitation training stage; the historical valid endpoint set does not include the limb movement endpoint of the current training trial and the limb movement endpoint that is later than the current training trial.

4. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 1, characterized in that, In step S3, let the set of historical valid endpoints be... , The tolerance center and the tolerance half-width are determined according to the following formulas: Where, p i Let R represent the i-th valid historical endpoint, n represent the number of valid historical endpoints, and Ri represent the number of valid historical endpoints. d Let p represent a d-dimensional real space, where d represents the coordinate dimension of the endpoint of the motion. i，j c represents the j-th dimension coordinate of the i-th valid historical endpoint. j Let j represent the j-th dimension coordinate of the tolerance center, median i h represents taking the median of the data corresponding to index i. j Let Qq represent the tolerance half-width in the j-th dimension, Qq represent the empirical q-quantile, and q represent the quantile level. Minimum tolerance half-width, For safety, the endpoint tolerance domain satisfies ch≤p≤c+h, where c represents the tolerance center, h represents the vector composed of the tolerance half-widths of each dimension, and p represents the limb movement endpoint.

5. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 4, characterized in that, In step S4, the axis-by-axis out-of-bounds amount and the endpoint out-of-bounds cost are determined according to the following formulas: in, The term "axis-by-axis out-of-bounds" indicates the value of the cross-axis, "sign" indicates the sign function, "⊙" indicates element-by-element multiplication, and "|·|" indicates element-by-element absolute value. Let represent the cost of the endpoint going out of bounds, and ||·|2 represent the L2 norm.

6. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 5, characterized in that, In step S5, the tolerance potential energy and the recovery effect are determined according to the following formulas: in, This represents the positive definite diagonal gain matrix, where each diagonal gain... U(p) represents the tolerable potential energy at the current limb movement endpoint p. express The transpose of F(p) represents the recovery effect. This represents the gradient operator with respect to the current limb movement endpoint p.

7. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 6, characterized in that, In step S6, the mapping result of the recovery effect includes the exoskeleton robot joint reference torque, the exoskeleton robot joint control input correction amount, the assistance level, or the virtual guiding force applied to the patient's limb. When the Jacobian matrix is ​​used for the joint reference torque mapping, the reference torque is: ,in, The reference torque is represented by J, the Jacobian matrix is ​​represented by F, and the recovery effect is represented by F. The output constraint processing includes amplitude limiting, low-pass smoothing, rate of change limiting, boundary hysteresis, dead zone expansion, or minimum trigger time.

8. The exoskeleton robot rehabilitation training control method based on motion endpoint distribution according to claim 1, characterized in that, In step S7, the quality screening includes integrity, missing value ratio, motion amplitude, motion segment length, and sensor validity check. When the number of historical valid endpoints in the corresponding task group is insufficient, the group parameter or preset cold start parameter is used as the tolerance center and the tolerance half-width. The tolerance center and the tolerance half-width are updated as the number of motion endpoints that pass the quality screening increases, so as to transition from the group parameter or preset cold start parameter to individualized parameter.

9. A rehabilitation training control system for an exoskeleton robot based on motion endpoint distribution, used to execute the control method according to any one of claims 1 to 8, characterized in that, The system includes an exoskeleton robot and a controller communicatively connected to the exoskeleton robot. The exoskeleton robot includes a motion detection device and a joint drive assembly. The controller includes a historical endpoint processing module, a tolerance domain construction module, an out-of-bounds calculation module, and an auxiliary control module. The historical endpoint processing module is used to establish a set of historical valid endpoints based on limb motion data collected by the motion detection device. The tolerance domain construction module is used to construct an endpoint tolerance domain based on the median and absolute deviation empirical quantiles of the historical valid endpoint set in each coordinate dimension. The out-of-bounds calculation module is used to calculate the out-of-bounds amount of the current limb motion endpoint relative to the endpoint tolerance domain. The auxiliary control module is used to maintain zero additional assistance when the current limb motion endpoint is within the endpoint tolerance domain, and generate a continuous recovery action based on the actual out-of-bounds portion when the current limb motion endpoint crosses the boundary of the endpoint tolerance domain. The continuous recovery action is mapped to a joint reference torque, joint control input correction, or assistance level output to the joint drive assembly.

Images