Gait model adjustment method and system for exoskeleton robot

Abstract

This application belongs to the field of intelligent control technology, specifically relating to a method and system for adjusting the gait model of an exoskeleton robot. The method for adjusting the gait model of an exoskeleton robot involves: acquiring relevant data from the exoskeleton robot; inputting the relevant data into the gait model to obtain relevant results; generating a final environmental complexity evaluation result in response to fluctuations in the frequency of changes in the action data set; acquiring a gait subset and a historical subset in response to safety hazards in the environment in front of the exoskeleton robot; determining the user's walking intention in response to a mismatch between the gait subset and the historical subset, and determining a temporary gait adjustment strategy based on the final environmental complexity evaluation result; correcting the temporary gait adjustment strategy based on re-acquired user action data to obtain a final gait adjustment strategy, which matches the final environmental complexity evaluation result; adjusting some coefficient values ​​in the gait model to obtain the adjusted gait model.

Description

Technical Field

[0001] This application belongs to the field of intelligent control technology, specifically relating to a method and system for adjusting the gait model of an exoskeleton robot. Background Technology

[0002] Exoskeleton robots, as intelligent devices that assist human movement, have shown great potential in medical rehabilitation, industrial operations, and daily life. Their core value lies in helping users perform complex actions such as walking, climbing stairs, and overcoming obstacles through precise gait control, thereby improving quality of life and work efficiency. However, in practical applications, accurately sensing user intentions and adapting to complex and ever-changing dynamic environments remains a key challenge limiting the performance of exoskeleton robots.

[0003] Currently, most existing gait detection methods for exoskeleton robots rely on a single type of sensing device, such as inertial measurement units (IMUs) or other body-attached sensors, to capture user motion data. This single-sensor approach struggles to comprehensively acquire and understand changes in the surrounding environment when faced with complex terrain or sudden environmental changes. This leads to the system's inability to adjust gait strategies in a timely manner, resulting in misjudgments or slow responses. This not only affects the device's practicality and comfort but may also pose a potential threat to the user's safety.

[0004] Therefore, how to effectively combine the terrain features ahead with the user's own action information in a dynamic environment to achieve accurate identification and advance adjustment of complex gait patterns is a technical problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] To address the above issues, this application provides a gait model adjustment method and system for exoskeleton robots, which solves the problems of incomplete environmental information acquisition, lack of integration between environmental information and user action information, slow response, and gait misjudgment caused by the reliance on a single sensor in existing exoskeleton robot gait detection methods.

[0006] To achieve the above objectives, this application provides a method for adjusting the gait model of an exoskeleton robot, comprising:

[0007] Acquire a set of motion data for the exoskeleton robot, wherein the set of motion data includes at least user motion data and terrain feature data;

[0008] The user action data and the terrain feature data are input into the gait model to obtain the slope data and obstacle density data of the area in front of the exoskeleton robot, and an initial environmental complexity evaluation result is generated based on the slope data and obstacle density data of the area in front of the exoskeleton robot.

[0009] The frequency of change of the action data set is detected, and in response to fluctuations in the frequency of change of the action data set, a final evaluation result of the environmental complexity is generated by combining the initial evaluation result of the environmental complexity.

[0010] In response to a safety hazard in the environment in front of the exoskeleton robot, a gait subset and a corresponding historical subset are obtained from the user's motion data.

[0011] In response to a mismatch between the gait subset and the historical subset, the user's walking intention is determined based on the user action data, and a temporary gait adjustment strategy is determined in conjunction with the final environmental complexity evaluation result.

[0012] The user action data is acquired again, and the gait adjustment temporary strategy is corrected based on the acquired user action data to obtain the gait adjustment final strategy;

[0013] In response to the matching between the ultimate gait adjustment strategy and the ultimate environmental complexity evaluation result, the coefficient values ​​of at least some of the components in the gait model used to control the gait generation logic of the exoskeleton robot are adjusted according to the ultimate gait adjustment strategy to obtain the adjusted gait model.

[0014] On the one hand, this application provides a gait model adjustment system for an exoskeleton robot, including:

[0015] The data acquisition module is used to acquire a set of motion data of the exoskeleton robot, which includes at least user motion data and terrain feature data;

[0016] The initial evaluation module is used to input the user action data and the terrain feature data into the gait model, obtain the slope data and obstacle density data of the area in front of the exoskeleton robot, and generate an initial evaluation result of environmental complexity based on the slope data and obstacle density data.

[0017] The final evaluation module is used to detect the frequency of change of the action data set, and in response to fluctuations in the frequency of change of the action data set, it generates a final evaluation result of the environmental complexity based on the initial evaluation result of the environmental complexity.

[0018] The subset acquisition module is used to acquire a gait subset and a historical subset corresponding to the gait subset in response to a safety hazard in the environment in front of the exoskeleton robot.

[0019] A temporary strategy generation module is used to determine the user's walking intention based on the user action data and to determine a temporary gait adjustment strategy in response to a mismatch between the gait subset and the historical subset, and in conjunction with the final environmental complexity evaluation result.

[0020] The strategy correction module is used to acquire user action data again and correct the gait adjustment temporary strategy based on the acquired user action data to obtain the gait adjustment final strategy.

[0021] The model adjustment module is used to adjust the coefficient values ​​of at least some of the gait models used to control the gait generation logic of the exoskeleton robot according to the gait adjustment final strategy and the environmental complexity final evaluation result in response to the gait adjustment final strategy, so as to obtain the adjusted gait model.

[0022] This application provides a method and system for adjusting the gait model of an exoskeleton robot. Through a deep integration of multi-sensor fusion and dynamic gait adjustment, combined with visual intelligence technology, it not only improves the accuracy of environmental perception but also enhances the dynamic adaptability and safety reliability of the exoskeleton robot. Specifically, by fusing user action data and terrain feature data to form a gait model, and using visual intelligence to perform structured perception of the terrain ahead, it comprehensively acquires information on ground slope changes and obstacle distribution, solving the problem of incomplete information from a single sensor. Simultaneously, this application embodiment continuously detects the frequency of changes in the action data set. Once a fluctuation exceeds a preset threshold, a dynamic interference assessment is immediately triggered, incorporating environmental change factors into the evaluation result to generate a more accurate final evaluation result of environmental complexity. This mechanism enables the exoskeleton robot to perceive changes during sudden environmental shifts, rather than passively reacting after gait errors occur, thus effectively overcoming the slow response issues in related technologies.

[0023] Furthermore, the embodiments of this application integrate the final evaluation result of environmental complexity, gait subset, historical subset, and user intent information for judgment. By utilizing the visual intelligence to accurately identify the terrain features ahead, the exoskeleton robot can make an overall judgment by simultaneously combining the terrain ahead and the user's current posture. This avoids gait misjudgment caused by insufficient information in a single dimension, and improves the gait detection accuracy, dynamic response speed, and walking safety of the exoskeleton robot in complex terrain.

[0024] Finally, in the embodiments of this application, user action data is acquired again as feedback, and the temporary strategy is corrected through an adaptive parameter update mechanism to generate a more accurate final strategy. Furthermore, after confirming that the final strategy matches the environmental complexity evaluation results, the method adjusts the coefficient values ​​in the gait model (such as stride adjustment coefficient, gait frequency adjustment coefficient, joint torque distribution coefficient, center of gravity offset compensation coefficient, etc.) so that the gait model can directly call the optimized parameters in subsequent similar terrain conditions. This closed-loop feedback and continuous model optimization mechanism enables the exoskeleton robot to continuously and intelligently improve with changes in the environment and user state, rather than recalculating each time. This allows the exoskeleton robot to have better visual perception and gait adaptation capabilities in complex environments, thereby achieving continuous adaptability in complex environments. Attached Figure Description

[0025] Figure 1 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the first embodiment of this application;

[0026] Figure 2 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the second embodiment of this application;

[0027] Figure 3 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the third embodiment of this application;

[0028] Figure 4 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the fourth embodiment of this application;

[0029] Figure 5 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the fifth embodiment of this application;

[0030] Figure 6 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the sixth embodiment of this application;

[0031] Figure 7 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the seventh embodiment of this application;

[0032] Figure 8 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the eighth embodiment of this application;

[0033] Figure 9 This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the ninth embodiment of this application;

[0034] Figure 10This is a flowchart illustrating the gait model adjustment method for the exoskeleton robot provided in the tenth embodiment of this application;

[0035] Figure 11 This application also provides a schematic diagram of a gait model adjustment system for an exoskeleton robot. Detailed Implementation

[0036] To enable those skilled in the art to better understand the technical solution, the present application will be described in detail below with reference to the embodiments. The description in this section is only exemplary and explanatory, and should not be used to limit the scope of protection of the present application in any way.

[0037] Exoskeleton robots, as intelligent devices that assist human movement, have shown great potential in medical rehabilitation, industrial operations, and daily life. Their core value lies in helping users complete complex actions such as walking, climbing stairs, and overcoming obstacles through precise gait control. However, accurately sensing user intentions and adapting to complex and ever-changing dynamic environments remains a key challenge that limits their performance.

[0038] Most existing gait detection methods for exoskeleton robots rely on a single type of sensing device, such as capturing user motion data solely through an inertial measurement unit. When faced with complex terrain or sudden environmental changes, this approach struggles to comprehensively acquire information about the surrounding environment, causing the exoskeleton robot to be unable to adjust its gait strategy in a timely manner. This can lead to misjudgments or slow reactions, which not only affects the practicality and comfort of the exoskeleton robot but may also endanger the user's personal safety.

[0039] The inventors discovered that the core challenges of gait detection technology lie in two aspects: First, how to accurately acquire feature information about the terrain ahead in a dynamic environment, such as the elevation changes of the ground or the position and shape of obstacles. Second, the lack of effective integration between environmental information and the user's own action information makes it difficult for the system to form a holistic judgment in complex scenarios. For example, when faced with a sudden step, the robot cannot simultaneously combine the terrain ahead and the user's current posture, missing the best adjustment opportunity and causing gait errors or unnatural movement transitions.

[0040] Therefore, how to effectively combine the terrain features ahead with the user's own motion information to achieve accurate identification and advance adjustment of complex gait patterns has become a key problem that urgently needs to be solved in the field of gait detection for exoskeleton robots.

[0041] To address the aforementioned technical problems, one or more embodiments and combinations thereof in this specification provide a gait model adjustment method for exoskeleton robots, thereby solving the problem of the lack of effective fusion between terrain features and user action information in gait detection under complex environments.

[0042] In one embodiment, see Figure 1The method for adjusting the gait model of an exoskeleton robot includes the following steps:

[0043] Step S100: Obtain the motion data set of the exoskeleton robot. The motion data set includes at least user motion data and terrain feature data.

[0044] Step S200: Input user action data and terrain feature data into the gait model, obtain slope data and obstacle density data of the area in front of the exoskeleton robot, and generate an initial evaluation result of environmental complexity based on the slope data and obstacle density data of the area in front of the exoskeleton robot.

[0045] Step S300: Detect the frequency of change in the action data set, and in response to fluctuations in the frequency of change in the action data set, combine the initial evaluation results of environmental complexity to generate the final evaluation results of environmental complexity.

[0046] Step S400: In response to a safety hazard in the environment in front of the exoskeleton robot, acquire a gait subset from the user's motion data and a corresponding historical subset.

[0047] Step S500: In response to the mismatch between the gait subset and the historical subset, determine the user's walking intention based on the user's action data, and determine a temporary gait adjustment strategy based on the final evaluation result of environmental complexity.

[0048] Step S600: Acquire user action data again, and revise the temporary gait adjustment strategy based on the newly acquired user action data to obtain the final gait adjustment strategy.

[0049] Step S700: In response to the matching of the gait adjustment ultimate strategy and the environmental complexity ultimate evaluation result, adjust at least some of the coefficient values ​​in the gait model used to control the gait generation logic of the exoskeleton robot according to the gait adjustment ultimate strategy to obtain the adjusted gait model.

[0050] In step S100, a motion data set of the exoskeleton robot is acquired, which includes at least user motion data and terrain feature data.

[0051] Specifically, motion data sets can refer to the combined data collected by various sensors mounted on an exoskeleton robot. As an example, during walking, the exoskeleton robot collects the angular velocity and acceleration data of the user's lower limbs at a frequency of 100Hz using an inertial measurement unit (IMU) as user motion data.

[0052] Terrain feature data can refer to spatial geometric information describing the road conditions and environment in front of the exoskeleton robot, including but not limited to ground slope, obstacle position and shape, road surface smoothness, and step height. For example, depth images within a 50-meter range in front can be collected by a depth camera, and point cloud data of the surrounding environment can be generated by LiDAR at a frequency of 10Hz as terrain feature data. These data together constitute the motion data set.

[0053] In step S200, user action data and terrain feature data are input into the gait model to obtain slope data and obstacle density data of the area in front of the exoskeleton robot, and an initial evaluation result of environmental complexity is generated based on the slope data and obstacle density data of the area in front of the exoskeleton robot.

[0054] A gait model can refer to a mathematical model used to describe and predict the gait behavior of an exoskeleton robot. It contains multiple adjustable coefficient parameters, such as stride adjustment coefficient and joint torque distribution coefficient.

[0055] The slope data in front of the exoskeleton robot can refer to the spatial distribution information describing the degree of inclination of the ground in front of the exoskeleton robot. Specifically, it can be represented by the slope angle or slope percentage of each grid cell after discretizing the terrain in front of the robot into a grid. Obstacle density data can refer to the statistical value of the number of obstacles per unit area, specifically a quantitative indicator of the density of obstacle distribution in the area in front of the robot.

[0056] Specifically, through visual intelligence technology, environmental perception distribution information can be generated from spatial distribution data formed by quantifying and describing the environment surrounding the exoskeleton robot, thus achieving a visual understanding of the terrain ahead. This environmental perception distribution information can be a heatmap or grid map generated based on the slope data and obstacle density data. That is, the slope data and obstacle density data are the foundational data constituting the environmental perception distribution information, which is a visual spatial distribution representation formed by fusing the two.

[0057] Subsequently, an initial environmental complexity evaluation result can be generated based on the environmental perception distribution information. As an example, the lidar point cloud data is converted into a 3D grid map, and the slope value of each grid is calculated. For example, a slope exceeding 15 degrees is marked as a high-risk area. Simultaneously, a deep learning model in visual intelligence, such as YOLOv5, is used to perform target detection on the depth image, identifying obstacle locations and generating a heatmap that divides the area in front into low, medium, and high complexity levels as environmental perception distribution information. Then, the initial environmental complexity evaluation result is calculated using a preset scoring formula. For example, if the slope risk coefficient of a certain area is 0.8 and the obstacle density is 0.5 per square meter, the comprehensive score is 0.65. The above data is only for illustrative purposes, and this embodiment does not impose specific limitations on the initial environmental complexity evaluation result.

[0058] In step S300, the frequency of change of the action data set is detected, and in response to fluctuations in the frequency of change of the action data set, the final evaluation result of the environmental complexity is generated by combining the initial evaluation result of the environmental complexity.

[0059] As an example, when the acceleration fluctuation of the inertial measurement unit exceeds 2 m / s², or the change rate of the lidar point cloud density exceeds 30%, the system determines that the data change frequency exceeds a preset threshold, triggering a dynamic interference assessment. Combined with potential risk prediction, a more accurate final evaluation result of environmental complexity is generated. The above data is only for illustrative purposes, and this embodiment does not impose specific limitations on the change frequency of the action data set.

[0060] In step S400, in response to a safety hazard in the environment in front of the exoskeleton robot, a gait subset and a corresponding historical subset are obtained from the user's motion data.

[0061] A gait subset can refer to a set of key feature parameters related to gait extracted from user motion data, such as stride length sequence, gait frequency sequence, or joint angle change sequence within a specific time window. As an example, when the system detects a steep slope exceeding 15 degrees or an obstacle with a diameter greater than 0.5 meters ahead, it determines that there is a safety hazard and extracts the stride length sequence and gait frequency sequence within the current time window as a gait subset.

[0062] Simultaneously, historical gait data under the same terrain conditions can be retrieved from the historical database as a historical subset. A historical subset can refer to a collection of historical gait data in the exoskeleton robot's historical operation records that is similar to the current terrain conditions and gait pattern, including but not limited to historical stride sequences, historical step frequency sequences, historical joint angle change sequences, and their corresponding control effect evaluation data.

[0063] It is understandable that the final environmental complexity evaluation result generated in step S300 is used to determine whether the environment ahead poses a safety hazard. This step will only be triggered when the final evaluation result exceeds the preset safety threshold.

[0064] In step S500, in response to the mismatch between the gait subset and the historical subset, the user's walking intention is determined based on the user's action data, and a temporary gait adjustment strategy is determined in combination with the final evaluation result of environmental complexity.

[0065] User walking intent can refer to the user's movement tendency inferred by analyzing user action data, such as the user's intention to accelerate, decelerate, turn, go uphill, or go downhill. As an example, if the matching degree between the current gait subset and the historical subset is lower than a preset deviation threshold (e.g., 85%), then the user's intent is inferred to be cautiously going uphill by analyzing the user action data.

[0066] Then, by combining the slope information from the final environmental complexity evaluation results, a temporary gait adjustment strategy can be generated. This temporary gait adjustment strategy can refer to a transitional gait control scheme generated based on the current environmental complexity and the user's walking intention, including but not limited to control parameters such as stride adjustment, cadence adjustment, joint angle correction values, and joint torque distribution ratios. For example, the cadence can be adjusted from 80 steps / minute to 75 steps / minute, and the stride length can be reduced from 0.7 meters to 0.65 meters. The above data is for illustrative purposes only, and this embodiment does not impose specific limitations on the matching degree threshold.

[0067] In step S600, user action data is acquired again, and the temporary gait adjustment strategy is corrected based on the acquired user action data to obtain the final gait adjustment strategy.

[0068] The ultimate gait adjustment strategy can refer to the final gait control scheme obtained by modifying and optimizing the temporary gait adjustment strategy based on real-time feedback data. Compared with the temporary strategy, this strategy has higher environmental adaptability and movement matching degree, including but not limited to control parameters such as the modified stride adjustment amount, the modified cadence adjustment amount, the modified joint angle correction value, and the modified joint torque distribution ratio.

[0069] As an example, during the execution of the temporary strategy, feedback data on the user's joint angles are collected in real time. If a large deviation is detected between the predicted gait sequence and the actual action data (e.g., the predicted angle of the left knee joint is 35 degrees, but the actual measured angle is 30 degrees), the control command is corrected through an adaptive parameter update mechanism to generate the adjusted final strategy.

[0070] In step S700, in response to the matching of the gait adjustment ultimate strategy and the environmental complexity ultimate evaluation result, at least some of the coefficient values ​​in the gait model used to control the gait generation logic of the exoskeleton robot are adjusted according to the gait adjustment ultimate strategy to obtain the adjusted gait model.

[0071] The coefficients in the gait model refer to adjustable parameters used within the gait model to control the gait generation logic. Adjusting these coefficients directly affects the final gait output of the exoskeleton robot. For example, adjustable coefficients include, but are not limited to, stride adjustment coefficient, joint torque distribution coefficient, center of gravity offset compensation coefficient, gait phase switching threshold coefficient, terrain adaptation sensitivity coefficient, and energy consumption optimization coefficient.

[0072] As an example, it was verified that the ultimate gait adjustment strategy effectively mitigated potential risks (such as reducing the probability of falling from 30% to 18%), confirming that it matched the ultimate environmental complexity evaluation results. Based on this, some coefficient values ​​in the gait model were updated, enabling the model to directly call the optimized parameters in subsequent similar terrain conditions, thus achieving continuous improvement of the gait model.

[0073] The stride adjustment factor controls the stride scaling of the exoskeleton robot under different terrain conditions. For example, when an uphill slope of 15 degrees is detected ahead, adjusting the stride adjustment factor from the default value of 1.0 to 0.85 will shorten the robot's stride from 0.8 meters to 0.68 meters. On flat terrain, the stride adjustment factor can be adjusted to 1.0 or 1.05 to restore or slightly increase the normal stride.

[0074] The cadence adjustment factor is used to control the gait cycle length of the exoskeleton robot. For example, when a slippery surface is detected ahead, adjusting the cadence adjustment factor from the default value of 1.0 to 1.15 increases the cadence from 80 steps / minute to 92 steps / minute, reducing the risk of slipping by increasing the cadence and decreasing the single-foot contact time. On a stable surface, the cadence adjustment factor can be adjusted to 0.95, reducing the cadence to 76 steps / minute to save energy.

[0075] Joint torque distribution coefficients are used to control the torque output ratio of each joint (such as the hip, knee, and ankle joints) of an exoskeleton robot in different gait phases. For example, in a stair-climbing scenario, adjusting the knee joint torque distribution coefficient from 0.3 to 0.5 and the hip joint torque distribution coefficient from 0.4 to 0.3 allows the knee joint to receive more driving force to complete the leg-lifting and step-over movement. In a downhill scenario, adjusting the knee joint torque distribution coefficient to 0.2 and the ankle joint torque distribution coefficient from 0.2 to 0.4 enhances the cushioning and support capabilities during downhill descents.

[0076] The center of gravity offset compensation coefficient is used to control the magnitude of center of gravity adjustment during the walking process of the exoskeleton robot. For example, when it detects that the user is carrying heavy objects (such as a backpack) or has weak strength in one side of the limb, the center of gravity offset compensation coefficient is adjusted from the default value of 1.0 to 1.2 to increase the lateral offset compensation of the center of gravity and maintain walking balance. In flat and open terrain, this coefficient can be adjusted to 0.9 to reduce unnecessary center of gravity adjustments and improve walking smoothness.

[0077] The gait phase switching threshold coefficient controls the timing of the gait transition between the support and swing phases. For example, when an increase in user walking speed is detected, the gait phase switching threshold coefficient is adjusted from the default value of 1.0 to 0.9, causing the transition from the support to the swing phase to occur 0.05 seconds earlier to accommodate the user's acceleration intention. When there is a clear intention to decelerate or stop, the coefficient is adjusted to 1.1 to delay the phase switching, providing a longer support time to enhance stability.

[0078] The terrain adaptation sensitivity coefficient is used to control the response speed of the gait model to environmental changes. For example, in complex and variable environments such as gravel roads or stairs, adjusting the terrain adaptation sensitivity coefficient from the default value of 1.0 to 1.3 makes the robot respond to environmental changes faster and adjust its gait more promptly. In flat and stable road environments, this coefficient is adjusted to 0.8 to reduce response sensitivity and avoid frequent misadjustments caused by sensor noise.

[0079] This is used to balance energy consumption and stability in gait control. For example, when the battery is low, the energy consumption optimization coefficient is adjusted from the default value of 1.0 to 1.2, causing the gait model to prioritize combinations of gait parameters with lower energy consumption, such as appropriately reducing joint torque output and minimizing unnecessary gait correction movements, thereby extending the exoskeleton robot's endurance. In scenarios with sufficient battery power or higher stability requirements, this coefficient is adjusted to 0.9, prioritizing gait stability over energy consumption.

[0080] It is understood that the adjustment of the above coefficient values ​​can be an independent adjustment of a single coefficient or a joint adjustment of multiple coefficients. For example, in an uphill scenario, the stride adjustment coefficient (lowered to 0.85), cadence adjustment coefficient (lowered to 0.94), knee joint torque distribution coefficient (increased to 0.5), and center of gravity shift compensation coefficient (slightly forward) can be adjusted simultaneously to achieve comprehensive optimization of uphill gait. The above data and adjustment methods are only illustrative examples. This embodiment does not impose any limitations on the specific coefficient values ​​and their adjustment ranges. In practical applications, they can be flexibly configured according to specific terrain conditions, user vital signs, and real-time feedback data.

[0081] In this embodiment, by fusing user motion data and terrain feature data to form a gait model, and leveraging visual intelligence to perceive the terrain ahead, a comprehensive understanding of ground slope changes and obstacle distribution information is obtained, solving the problem of incomplete information from a single sensor. Simultaneously, this embodiment continuously monitors the frequency of changes in the motion data set. Once a fluctuation exceeds a preset threshold, a dynamic interference assessment is immediately triggered, incorporating environmental change factors into the evaluation result to generate a more accurate final evaluation of environmental complexity. This mechanism enables the exoskeleton robot to perceive changes during sudden environmental shifts, rather than reacting passively after gait errors occur, effectively overcoming the slow response issues inherent in related technologies.

[0082] Furthermore, in this embodiment, the final evaluation result of environmental complexity, gait subset, historical subset, and user intent information are fused together for judgment. The visual intelligence is used to accurately identify the terrain features ahead, so that the exoskeleton robot can make an overall judgment by combining the terrain ahead and the user's current posture at the same time. This avoids gait misjudgment caused by insufficient information in a single dimension, and improves the gait detection accuracy, dynamic response speed, and walking safety of the exoskeleton robot in complex terrain.

[0083] Finally, in this embodiment, user action data is acquired again as feedback, and the temporary strategy is corrected through an adaptive parameter update mechanism to generate a more accurate final strategy. Furthermore, after confirming that the final strategy matches the environmental complexity evaluation results, the method adjusts the coefficient values ​​in the gait model (such as stride adjustment coefficient, gait frequency adjustment coefficient, joint torque distribution coefficient, center of gravity shift compensation coefficient, etc.) so that the gait model can directly call the optimized parameters in subsequent similar terrain conditions. This closed-loop feedback and continuous model optimization mechanism enables the exoskeleton robot to continuously and intelligently improve with changes in the environment and user state, rather than recalculating each time. This allows the exoskeleton robot to have better visual perception and gait adaptation capabilities in complex environments, thereby achieving continuous adaptability in complex environments.

[0084] In one embodiment, see Figure 2 Step S700 includes:

[0085] Step S710: Based on the gait adjustment final strategy and the final evaluation results of environmental complexity, confirm whether the safety hazards have been eliminated.

[0086] Step S720: In response to the elimination of safety hazards, determine whether the ultimate gait adjustment strategy matches the ultimate environmental complexity evaluation result, and adjust at least some of the coefficient values ​​in the gait model according to the ultimate gait adjustment strategy to obtain the adjusted gait model.

[0087] In step S710, as an example, the criteria for determining whether a safety hazard has been eliminated may include, but are not limited to: whether the user's fall probability has decreased below a safe threshold after gait adjustment, whether the distance between the exoskeleton robot and obstacles remains within a safe range, whether the user's joint angle deviation has converged to a normal range, and whether the ground reaction force has returned to a normal level. This embodiment does not impose specific limitations on these aspects.

[0088] Furthermore, after generating the final gait adjustment strategy in the aforementioned steps, environmental data and user action data can be re-collected to calculate the adjusted safety indicators. For example, suppose the system predicted a fall probability of 30% before adjustment, which is higher than the safety threshold of 15%, indicating a potential safety hazard. After implementing the final gait adjustment strategy, the system uses an inertial measurement unit and depth camera to monitor user posture and terrain changes in real time, reassessing the fall probability to 12%, which is lower than the preset safety threshold of 15%, thus determining that the safety hazard has been eliminated.

[0089] For example, the system detected that the distance to the obstacle in front increased from 0.8 meters before the adjustment to 1.5 meters after the adjustment, which is higher than the safe distance threshold of 1.0 meter. At the same time, the user's center of gravity shift decreased from 0.12 meters before the adjustment to 0.05 meters, which is lower than the shift threshold of 0.08 meters. Based on the above indicators, it can be determined that the safety hazard has been eliminated.

[0090] If the safety hazard is not eliminated, it is necessary to return to step S500 or S600 to regenerate and correct the gait adjustment strategy until the safety hazard is eliminated or a safety protection mechanism (such as emergency braking) is triggered.

[0091] In step S720, in response to the elimination of safety hazards, it is determined that the gait adjustment ultimate strategy matches the environmental complexity ultimate evaluation result, and at least some coefficient values ​​in the gait model are adjusted according to the gait adjustment ultimate strategy to obtain the adjusted gait model.

[0092] It is understandable that the elimination of safety hazards means that the generated gait adjustment ultimate strategy can effectively cope with the current environmental complexity. Therefore, there is a matching relationship between this strategy and the current environmental complexity evaluation results. At this time, the gait adjustment logic has reference value and reuse significance, and it is worth recording and associating.

[0093] As an example, once the safety hazard is confirmed to be eliminated, the system extracts key parameters from the final gait adjustment strategy (such as adjusting stride length from 0.7 meters to 0.65 meters, cadence from 80 steps / minute to 75 steps / minute, and knee joint torque distribution coefficient from 0.3 to 0.5), and updates the corresponding coefficient values ​​in the gait model accordingly. The coefficient values ​​in the gait model can refer to adjustable parameters used within the gait model to control the gait generation logic. Adjusting these coefficient values ​​directly affects the final gait output effect of the exoskeleton robot. For example, adjustable coefficient values ​​include, but are not limited to, stride length adjustment coefficient, cadence adjustment coefficient, joint torque distribution coefficient, center of gravity shift compensation coefficient, gait phase switching threshold coefficient, terrain adaptation sensitivity coefficient, and energy consumption optimization coefficient. Specific definitions and adjustment examples for various coefficients have been described in detail in previous steps and will not be repeated here.

[0094] Furthermore, a correlation can be established between the ultimate gait adjustment strategy and environmental perception distribution information. This correlation involves binding and storing the control parameters of the ultimate gait adjustment strategy with the environmental perception distribution information that triggers the strategy (such as a specific slope range, obstacle density level, road surface smoothness index, etc.), forming a mapping pair between environmental features and gait strategies. For example, the system records the current environmental perception distribution information: an upcoming slope of 15 degrees, an obstacle density of 0.3 per square meter, and a road surface smoothness index of 0.7. Simultaneously, it records the ultimate gait adjustment strategy used to successfully eliminate the safety hazard: stride adjustment coefficient 0.85, cadence adjustment coefficient 0.94, and knee joint torque distribution coefficient 0.5. The system stores these two pieces of information in a knowledge base, forming an empirical rule: when encountering similar environmental features again (slope 12 to 18 degrees, obstacle density 0.2 to 0.5 per square meter, smoothness index 0.6 to 0.8), this gait strategy can be directly invoked as the initial solution. It is understandable that exoskeleton robots can quickly invoke proven and effective strategies when facing similar terrain, significantly shortening response time and improving the efficiency and accuracy of gait adjustment.

[0095] In this embodiment, the effectiveness of the gait adjustment strategy is verified by confirming whether safety hazards have been eliminated based on the final gait adjustment strategy and the final environmental complexity evaluation results. After performing gait adjustment, the method re-collects environmental data and user action data, calculates the adjusted safety indicators, and comprehensively judges whether safety hazards have been eliminated. If safety hazards have not been eliminated, the process returns to the previous steps to regenerate and correct the strategy. This verification mechanism ensures the effectiveness of each gait adjustment and user safety, avoiding safety incidents caused by ineffective adjustments.

[0096] In this embodiment, dynamic optimization of the gait model coefficients is implemented. The coefficient values ​​in the gait model are only updated after the ultimate gait adjustment strategy has passed security verification and is confirmed to effectively cope with the current environmental complexity. This method ensures that only effective strategies, proven in practice, can influence the behavior of the gait model, avoiding contamination of the model by invalid or erroneous strategies, and allowing the gait model to continuously optimize with the accumulation of successful experiences.

[0097] Furthermore, in this embodiment, the control parameters in the ultimate gait adjustment strategy that successfully eliminates safety hazards are bound and stored with the environmental perception distribution information that triggers the strategy. Based on visual intelligence environmental feature extraction and matching technology, a mapping pair between environmental features and gait strategies is formed. When similar environmental features are encountered again, the gait strategy can be directly invoked as the initial solution. This method effectively solves the problems of existing technologies that require recalculating the gait strategy every time similar terrain is encountered, resulting in large computational load and slow response speed. It significantly shortens the response time of gait adjustment, reduces the real-time computational burden, and improves the overall efficiency of the system.

[0098] As can be understood, in this embodiment, security verification ensures the effectiveness of the strategy, coefficient updates solidify the effective strategy into the model, and association construction enables the environmental storage and rapid retrieval of experience. These three elements work together to allow the exoskeleton robot to continuously accumulate experience and optimize its behavior during use, significantly improving its long-term adaptability and intelligence level in complex environments.

[0099] In one embodiment, see Figure 3 Step S710 includes:

[0100] Step S711: Calculate the gait stability improvement rate based on the gait stability data before and after implementing the final gait adjustment strategy.

[0101] Step S712: Calculate the risk mitigation rate based on the risk index before and after adjusting the final strategy according to the execution gait.

[0102] Step S713: In response to the fact that both the gait stability improvement rate and the risk mitigation rate exceed the first threshold, it is confirmed that the safety hazard has been eliminated.

[0103] In step S711, gait stability data can refer to index data used to quantitatively evaluate the walking stability of an exoskeleton robot, including but not limited to center of gravity shift amplitude, joint angle variance, plantar pressure distribution uniformity, and stride consistency. Gait stability improvement rate can refer to the degree of improvement in gait stability compared to before the final gait adjustment strategy. As an example, it can be calculated as: (stability index value before adjustment - stability index value after adjustment) / stability index value before adjustment × 100%.

[0104] For example, before executing the final gait adjustment strategy, the system collects the user's gait stability data through an inertial measurement unit and plantar pressure sensors. It can be understood that the improvement rate of stability data from different dimensions can be calculated individually, or a weighted fusion method can be used to calculate the overall stability improvement rate.

[0105] In step S712, the risk index can refer to a comprehensive indicator used to quantitatively assess the likelihood of a safety accident occurring in the exoskeleton robot under current conditions, and may include, but is not limited to, fall risk index, collision risk index, and slip risk index. The risk mitigation rate can refer to the degree to which various risks are reduced compared to before the gait adjustment final strategy is implemented. For example, it can be calculated as: (Risk index before adjustment - Risk index after adjustment) / Risk index before adjustment × 100%.

[0106] As an example, before executing the final gait adjustment strategy, the system calculates the current risk index by fusing environmental awareness data and user action data. It can be understood that the mitigation rate of different dimensions of the risk index can be calculated individually, or a weighted fusion method can be used to calculate the overall risk mitigation rate. The specific weights can be configured according to the security requirements of the application scenario.

[0107] In step S713, the first threshold can be a pre-set critical value for determining whether a safety hazard has been eliminated. This threshold can be dynamically configured based on factors such as the safety requirements of the application scenario, user characteristics, and environmental complexity. When both the gait stability improvement rate and the risk mitigation rate exceed this threshold, it indicates that the ultimate gait adjustment strategy has effectively improved walking stability and significantly reduced various safety risks, thus confirming that the safety hazard has been eliminated. As an example, the first threshold can be 30%, 50%, 60%, etc.

[0108] It is understandable that the setting of the first threshold directly affects the rigor of safety hazard assessment. In scenarios with high safety requirements (such as rehabilitation training and elderly assistance), a lower threshold can be used to ensure sufficient safety redundancy, while in scenarios where efficiency is prioritized, a relatively higher threshold can be used to balance safety and performance. In addition, the system can also dynamically adjust the first threshold according to the complexity of the real-time environment. For example, the threshold can be appropriately lowered in complex terrain to increase safety redundancy, while the threshold can be appropriately raised in simple terrain to avoid over-conservatism.

[0109] In this embodiment, multi-dimensional gait stability data, including center of gravity shift amplitude, joint angle variance, plantar pressure distribution uniformity, and stride consistency, are collected before and after implementing the ultimate gait adjustment strategy. The stability improvement rate for each dimension or the overall data is then calculated. Compared to existing technologies that rely solely on subjective observation or a single indicator to judge the degree of gait improvement, this embodiment provides an objective, multi-dimensional quantitative evaluation method, making the assessment of gait adjustment effects more scientific and accurate.

[0110] Furthermore, in this embodiment, multiple risk indicators such as fall risk index, collision risk index, and slip risk index are evaluated before and after implementing the final gait adjustment strategy, and the risk mitigation rate of each dimension or the whole is calculated. Compared with the shortcomings of existing technologies that lack risk quantification assessment methods and cannot objectively judge the degree of safety improvement, this embodiment provides reliable data support for the determination of safety hazards.

[0111] Meanwhile, this embodiment requires that both the gait stability improvement rate and risk mitigation rate exceed a preset first threshold before confirming that the safety hazard has been eliminated. Compared to the shortcomings of existing technologies that are prone to misjudgment due to subjective judgment or single-indicator judgment, this step adopts a dual-dimensional quantitative indicator joint judgment method, which significantly improves the accuracy and reliability of safety hazard judgment. The first threshold can be dynamically configured according to factors such as the security requirements of the application scenario, user characteristics, and environmental complexity, making the judgment criteria flexible and adaptable.

[0112] In one embodiment, user motion data is acquired via an inertial measurement unit. Terrain feature data is acquired via LiDAR and a depth camera.

[0113] Please see Figure 4 Step S200 includes:

[0114] Step S210: Use the Kalman filter algorithm to perform time synchronization and spatial registration of user action data and terrain feature data.

[0115] Step S220: Use a deep learning model to identify the location of obstacles in the terrain feature data.

[0116] Step S230: Convert the point cloud data in the terrain feature data into a 3D grid map, calculate the slope value of each grid, and obtain the slope data.

[0117] Step S240: Use a clustering algorithm to identify obstacle boundaries in the terrain feature data, and count the number of obstacles per unit area based on the obstacle location and obstacle boundaries to obtain obstacle density data.

[0118] In step S210, time synchronization can refer to aligning data collected by different sensors on the time axis so that user action data and terrain feature data collected at the same time can correspond. Spatial registration can refer to transforming data from different sensor coordinate systems to the same reference coordinate system so that the data can correspond in spatial location.

[0119] Because the inertial measurement unit (IMU), LiDAR, and depth camera have different sampling frequencies (e.g., the IMU acquires data at 100Hz, the LiDAR generates point cloud data at 10Hz, and the depth camera acquires depth images at 30 frames per second), and the installation positions and coordinate systems of each sensor differ, time synchronization and spatial registration of the multi-source data are necessary. As an example, a Kalman filter algorithm is used to timestamp-align the angular velocity and acceleration data acquired by the IMU, the point cloud data generated by the LiDAR, and the depth images acquired by the depth camera, controlling the error within 0.05 meters. All data is then converted to a world coordinate system with the exoskeleton robot's center of gravity as the origin to ensure the accuracy of subsequent fusion processing.

[0120] Time synchronization of multi-source sensor data can be achieved as follows: using the sampling frame period of the lidar as the reference time axis, the user action data collected by the inertial measurement unit and the terrain feature data collected by the depth camera are estimated at the lidar sampling time through the state prediction mechanism of the Kalman filter algorithm, thereby achieving the alignment of sensor data of different frequencies on the time axis.

[0121] As an optional implementation, a data buffer is established based on the time delay characteristics of different sensors. Data collected by each sensor is stored in its respective buffer according to timestamps, and time synchronization is performed based on the time of information acquisition by the inertial measurement unit. The state equation and observation equation of the Kalman filter algorithm are pre-calibrated according to the clock drift statistics and transmission delay statistics of each sensor, and the synchronization error is adjusted in real time through covariance matching of the innovation sequence. Specifically, for each sensor data stream, the system time at the time of acquisition and the system time at the time of arrival at the processing unit are recorded, and the transmission delay is calculated. The Kalman filter is used to estimate the delay online, and the delay estimate is introduced into the state equation to compensate for the delay of the sensor data, eliminating the data timing misalignment caused by the transmission delay. The time synchronization method described above is suitable for multi-source sensor scenarios where the start-up times of each sensor are asynchronous, the sampling frequencies are inconsistent, and the transmission delay fluctuates. It can unify multi-source sensor data with different frequencies and arrival times under the same time reference.

[0122] In step S220, obstacles can refer to objects in the environment in front of the exoskeleton robot that may affect its walking safety, including but not limited to steps, rocks, potholes, and the edges of slopes. As an example, the YOLOv5 deep learning model in visual intelligence can be used to detect objects in depth images captured by a depth camera. This model can identify obstacle categories in the image in real time and output their bounding box coordinates, leveraging the advantages of visual intelligence in object detection. By mapping the obstacle positions in the image coordinate system to the spatially registered world coordinate system, the precise position of the obstacle in the world coordinate system is obtained. During the model training phase, the YOLOv5 model is trained using a labeled image dataset containing various terrain and obstacle types, enabling it to achieve an obstacle recognition accuracy of over 90% in practical applications.

[0123] A deep learning model (such as YOLOv5) is used to perform object detection on depth images captured by a depth camera. Understandably, to accommodate the computational resource constraints of the embedded platform of the exoskeleton robot, the deep learning model can employ a lightweight network structure. For example, the YOLO network can be improved by introducing a Fire module and a 1×1 bottleneck layer, or a lightweight backbone network can replace the original detection network's backbone network, significantly reducing the amount of forward computation and parameter data. As another optional implementation, the parameters of the deep learning model can be quantized using INT8 and deployed on the embedded platform to reduce model memory usage and improve inference speed.

[0124] In step S230, the three-dimensional grid map can refer to discretizing the terrain space in front into several grid cells, each grid cell containing terrain height information for that area. The slope value can refer to the degree of ground inclination corresponding to each grid cell, usually expressed as an angle or percentage. The slope data can be a collection of multiple slope values.

[0125] As an example, point cloud data generated by LiDAR is downsampled using voxel filtering and projected onto a horizontal grid. Each grid cell is set to 0.1m x 0.1m in size, and the average height of all point clouds within the grid is taken as the ground height value for that grid. For each grid cell, the slope value is calculated based on the height difference between it and its adjacent cells. For example, when the height difference between a grid cell and the grid in front of it is 0.3m and the horizontal distance is 1.0m, the slope value is approximately 16.7 degrees. When the slope value exceeds 15 degrees, the grid cell is marked as a high-risk area for subsequent environmental complexity assessment. The above data is for illustrative purposes only.

[0126] In step S240, the obstacle boundary can refer to the spatial outline of the obstacle, used to determine the size and shape of the obstacle so that the exoskeleton robot can plan an obstacle avoidance path. As an example, the DBSCAN clustering algorithm is used to cluster the point cloud data generated by the LiDAR, setting the minimum clustering distance to 0.3 meters, and grouping spatially adjacent point clouds with a distance of less than 0.3 meters into the same obstacle cluster.

[0127] Obstacle density data refers to the statistical value of the number of obstacles per unit area, used to quantitatively describe the density of obstacle distribution in the area in front of an exoskeleton robot. The obstacle density can be calculated as follows: within a specified spatial area, count the number of identified obstacles, divide by the total area of ​​that area, and obtain the number of obstacles per unit area.

[0128] For each obstacle cluster obtained from clustering, its minimum bounding rectangle or convex hull boundary is calculated as the obstacle's boundary profile. Obstacles with a diameter greater than 0.5 meters are identified, and their center coordinates and boundary range are labeled. For smaller obstacles (such as pebbles with a diameter less than 0.1 meters), they can be treated as road noise rather than as independent obstacles to avoid overreaction. The above data is for illustrative purposes only.

[0129] In addition, after step S240, the process may further include dividing the environment in front of the exoskeleton robot into regions of different complexity levels based on the slope value, obstacle position, and obstacle boundary, thereby generating environmental perception distribution information.

[0130] Environmental perception distribution information refers to spatial distribution data formed after quantifying the environment surrounding the exoskeleton robot, typically represented in the form of a heatmap or grid map. As an example, the slope value of each grid calculated in step S230 is fused with the obstacle positions and boundary information identified in steps S220 and S240, and the area in front is divided into three complexity levels: low, medium, and high, according to a preset complexity classification rule. For example, the scoring formula can be set as: Complexity Score = 0.4 × Slope Risk Coefficient + 0.3 × Obstacle Density + 0.3 × Data Volatility, with a scoring range of 0 to 10. The slope risk coefficient is mapped based on the grid slope value (e.g., 0.2 for slopes less than 5 degrees, 0.5 for slopes between 5 and 15 degrees, and 0.8 for slopes greater than 15 degrees). The obstacle density is calculated based on the number of obstacles per unit area, and the data volatility is determined based on the real-time changes in sensor data. Based on the above scoring results, an environmental perception distribution map is generated. Areas with a score less than 3 are marked as low complexity (green), areas with a score between 3 and 6 are marked as medium complexity (yellow), and areas with a score greater than 6 are marked as high complexity (red). This data is stored in the system database in the form of a heatmap to provide a basis for the formulation of subsequent gait strategies. The above data is for illustrative purposes only.

[0131] Furthermore, in steps S220 to S250, through visual intelligence technology, this embodiment preprocesses obstacle target detection and point cloud data based on deep learning models, calculates slope values ​​based on three-dimensional grid maps and extracts obstacle boundaries based on spatial clustering algorithms, combines multi-source fusion of slope data and obstacle density data, quantitative classification of environmental complexity levels, and visualization output in the form of heat maps or raster maps, to provide a decision basis for subsequent gait model adjustments.

[0132] In this embodiment, a Kalman filter algorithm is used to synchronize and spatially register user action data and terrain feature data. This aligns multi-source sensor data from different sampling frequencies and coordinate systems to a unified spatiotemporal reference, providing an accurate data alignment foundation for subsequent multi-source data fusion processing. Based on this, visual intelligence technology is used to fuse and perceive the multi-source data, significantly improving the accuracy of environmental understanding. Furthermore, this embodiment employs a deep learning model to identify obstacle locations in the terrain feature data. It can detect various obstacles such as steps, rocks, and potholes from depth images in real time and output their precise coordinates, leveraging the advantages of visual intelligence in target detection and achieving intelligent identification and localization of obstacles in the foreground environment.

[0133] In this embodiment, by converting point cloud data from terrain feature data into a 3D grid map and calculating the slope value of each grid, discrete point cloud data is transformed into a structured gridded terrain representation, achieving a refined quantitative assessment of ground slope changes. Furthermore, this embodiment utilizes the spatial modeling capabilities of visual intelligence. By employing a clustering algorithm to identify obstacle boundaries in the terrain feature data, it can extract the outline range and size information of obstacles from the point cloud data, providing geometric feature data of obstacles for subsequent obstacle avoidance path planning and gait adjustment.

[0134] Furthermore, in this embodiment, by dividing the environment in front of the exoskeleton robot into regions of different complexity levels based on slope value, obstacle position, and obstacle boundary, and generating environmental perception distribution information, multi-dimensional environmental features are integrated into an intuitive spatial distribution representation. This enables the exoskeleton robot to comprehensively and quantitatively perceive the spatial heterogeneity and complexity distribution of the environment in front, providing a scientific and accurate environmental data foundation for subsequent environmental complexity evaluation and gait strategy adjustment.

[0135] In one embodiment, see Figure 5 Step S300 includes:

[0136] Step S310: Use a time series anomaly detection method based on a sliding window to calculate the change frequency of the action data set.

[0137] Step S320: In response to the change frequency exceeding the second threshold, determine that the change frequency of the action data set has fluctuated, and use a Bayesian probability model, combined with the change frequency, to calculate the probability of interference, and use a random forest algorithm to calculate the risk index in the future time period.

[0138] Step S330: Generate the final environmental complexity evaluation result based on the probability of interference impact and the risk index.

[0139] In step S310, the sliding window can refer to a method of segmenting continuously acquired time-series data according to a fixed time length. The frequency of change can refer to the number of times the data values ​​in the action data set fluctuate significantly within a unit of time. As an example, the sliding window size is set to 5 minutes, and the step size is 1 minute, that is, the action data set within the most recent 5 minutes is analyzed every 1 minute.

[0140] Within a sliding window, the rate of change of various dimensions of data, such as the acceleration data of the inertial measurement unit, the point cloud data of the lidar, and the depth image data of the depth camera, is calculated. For example, the rate of change of acceleration data is defined as the percentage of times the absolute value of the difference between adjacent sampling points exceeds a preset fluctuation threshold, and the rate of change of point cloud density is defined as the relative change in the number of point clouds between the current frame and the previous frame. By combining the rates of change of each dimension of data, the sum of the number of times each threshold is exceeded per unit time is calculated, and then divided by the window duration to obtain the overall change frequency. For example, when 18 abrupt changes are detected within a 5-minute window, the calculated change frequency is 3.6 times / minute. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0141] In step S320, the second threshold can refer to a critical value for determining the frequency of changes in the environment that indicate a sudden change. The probability of interference can refer to the likelihood that current environmental changes will interfere with the gait stability of the exoskeleton robot. The risk index can refer to the overall risk level of a predicted safety accident occurring within a future time period.

[0142] As an example, a second threshold is set to 3 times / minute. When the frequency of change calculated in step S310 is 3.6 times / minute, it is determined that it exceeds the second threshold, indicating a fluctuation in the frequency of change in the action data set, i.e., a sudden change in the environment. Subsequently, a Bayesian probability model is used, with the frequency of change as the observed variable, combined with the probability of gait interference corresponding to different frequencies of change in historical data, to calculate the probability of current interference impact. Based on the calculation results, it is determined whether there is a significant risk of interference. Simultaneously, a random forest algorithm is used, with current environmental parameters (including slope value, obstacle density, frequency of change, etc.) as input features, to predict the risk index for the next hour. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0143] In step S330, the final evaluation result of environmental complexity can refer to the comprehensive environmental complexity quantification value after integrating static environmental information and dynamic change information. Compared with the initial evaluation result of environmental complexity generated in step S200, the final evaluation result more comprehensively reflects the real-time dynamic characteristics of the environment.

[0144] As an example, a weighted fusion method can be used to calculate the final evaluation result of environmental complexity. The formula for calculating the final evaluation result is: Final evaluation result = Initial evaluation result × (1 + Interference impact probability × Risk index), i.e., 0.65 × (1 + 0.75 × 0.82) = 0.65 × 1.615 ≈ 1.05. If the evaluation result is set to a range of 0 to 10, 1.05 can be mapped to the corresponding interval through linear mapping. When the final evaluation result exceeds a preset attention threshold (e.g., 6.0), the system determines that the current environmental complexity is high and requires close attention and timely adjustment of gait strategies. When the danger threshold (e.g., 8.0) is exceeded, the system triggers an emergency protection mechanism. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment. It can be understood that the final evaluation result of environmental complexity will serve as an important basis for judging safety hazards and generating gait adjustment strategies in subsequent steps.

[0145] In this embodiment, by employing a time series anomaly detection method based on a sliding window to calculate the change frequency of the action data set, the dynamic change characteristics of the action data set on the time axis can be captured in real time, accurately quantifying the severity of environmental changes. This provides an objective data basis for judging whether a sudden change has occurred in the environment. Combined with the ability of visual intelligence to model time series, the accuracy of change detection is further improved.

[0146] Moreover, in this embodiment, by using a Bayesian probability model combined with the frequency of change to calculate the probability of interference, and using probabilistic reasoning methods to quantify the possibility of current environmental changes interfering with gait stability, and at the same time using a random forest algorithm to calculate the risk index in the future time period, joint prediction of short-term and medium-term risks is achieved on the basis of visual intelligence.

[0147] Meanwhile, in this embodiment, by generating the final environmental complexity evaluation result based on the interference impact probability and risk index, the static initial environmental complexity evaluation is integrated with the dynamic environmental change characteristics, so that the final environmental complexity evaluation result can simultaneously reflect the static structure and dynamic change trend of the environment, providing a more comprehensive and accurate environmental basis for subsequent safety hazard judgment and gait strategy adjustment.

[0148] In one embodiment, see Figure 6 Step S400 includes:

[0149] Step S410: Combine principal component analysis algorithm with user action data to extract key gait features and obtain gait subsets.

[0150] Step S420: Obtain historical action data corresponding to the current time interval from the historical database, and use principal component analysis algorithm to extract key features of historical gait from the historical action data to obtain a historical subset.

[0151] In step S410, the principal component analysis algorithm can refer to a data processing method that reduces high-dimensional data to low-dimensional data through linear transformation while retaining the main information of the data. Gait key features can refer to core parameters that can characterize the user's walking pattern, including but not limited to stride length, cadence, joint angle change curves, and center of gravity shift trajectory.

[0152] A gait subset can refer to a set of key feature parameters related to gait extracted from user motion data. Since the user motion data collected by the inertial measurement unit has high dimensionality (e.g., containing multiple dimensions such as three-axis acceleration, three-axis angular velocity, and attitude angles), directly using the raw data for subsequent processing would result in excessive computation and redundant information. Therefore, principal component analysis (PCA) is required for dimensionality reduction.

[0153] As an example, a sampling frequency of 50Hz can be set to collect user action data within a 1-second time window, resulting in multidimensional data samples at 50 time points. The original features such as acceleration, angular velocity, and attitude angle at each time point are used as input vectors to construct a 50×N matrix (N being the original feature dimension). Principal component analysis (PCA) is used to calculate the covariance matrix, eigenvalues, and eigenvectors of this matrix. The top K principal components with a cumulative variance contribution rate of over 85% are selected as the dimensionality-reduced feature vectors. The extracted key gait features include stride length, stride frequency, and standard deviation of center of gravity shift, which together constitute a subset of the gait. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment. It can be understood that dimensionality reduction using PCA can significantly reduce the feature dimension while preserving the main information of the original data, thereby reducing the computational load of subsequent matching calculations and strategy generation.

[0154] In step S420, the historical database can refer to a database that stores user action data, terrain feature data, gait strategies, and their execution effects collected during the past operation of the exoskeleton robot. The historical subset can refer to a set of historical gait data in the exoskeleton robot's historical operation records that are similar to the current terrain conditions and gait patterns.

[0155] As an example, the system first retrieves historical records with similar matching conditions from the historical database based on the current time interval (e.g., the current time is 10:00 AM, and the current walking segment is a 15-degree uphill section). This includes historical motion data from the same time period (e.g., historical morning periods) and the same terrain conditions (e.g., slope range of 12 to 18 degrees). After obtaining the historical motion data, the system uses the same principal component analysis algorithm parameters as in step S410 (e.g., cumulative variance contribution rate threshold of 85%) to perform dimensionality reduction on the historical motion data, extracting key features of historical gait, including historical stride sequences (mean of 1.25 meters), historical step frequency sequences (mean of 2.1 steps / second), historical joint angle change sequences, and their corresponding control effect evaluation data (e.g., gait stability score, energy consumption score, etc.). These features together constitute a historical subset. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0156] It is understandable that steps S420 and S410 use the same principal component analysis algorithm parameters, ensuring the comparability of the current gait subset and the historical subset in the feature space, thus laying a unified data foundation for the matching degree calculation in subsequent steps. By comparing the current gait subset with the historical subset, the exoskeleton robot can determine whether the current gait is adapted to the environment and make gait adjustments based on past successful experiences.

[0157] In this embodiment, by combining principal component analysis algorithm with user action data to extract key gait features to obtain a gait subset, the high-dimensional user action data can be reduced to a low-dimensional set of key features. While retaining the main information of the original data, the data dimensionality is significantly reduced, the computational load of subsequent matching calculations and strategy generation is reduced, and the real-time response capability of the system is improved.

[0158] Moreover, in this embodiment, by obtaining historical action data corresponding to the current time interval from the historical database and extracting key features of the historical gait using the same principal component analysis algorithm, a historical subset is obtained. This ensures the uniformity and comparability of the current gait subset and the historical subset in the feature space, and establishes the comparison between the current gait and the historical gait under the same feature measurement standard, providing a reliable data foundation for subsequent matching degree calculation.

[0159] Furthermore, in this embodiment, by comparing the current gait subset with the historical subset, the exoskeleton robot can draw on past successful experiences to determine whether the current gait is adapted to the environment and provide a reference for gait adjustment. With the help of visual intelligence, the effective reuse of experience is realized.

[0160] In one embodiment, see Figure 7 Step S500 includes:

[0161] Step S510: Calculate the feature distance between the gait subset and the history subset using the Euclidean distance algorithm.

[0162] Step S520: Determine the matching degree between the gait subset and the history subset based on the feature distance.

[0163] Step S530: In response to the matching degree being lower than the third threshold, determine that the gait subset does not match the historical subset.

[0164] In step S510, the Euclidean distance algorithm can be a mathematical method for calculating the straight-line distance between two vectors, used to measure the similarity between two points in a multidimensional space. Feature distance can refer to the geometric distance between the gait subset and the history subset in the feature space; the smaller the distance, the more similar they are in terms of features. Since both the gait subset and the history subset are K-dimensional feature vectors after dimensionality reduction using principal component analysis, the Euclidean distance algorithm can be used to calculate their similarity. It can be understood that the smaller the feature distance, the closer the current gait is to the historical gait; the larger the feature distance, the greater the difference between the current gait and the historical gait.

[0165] In step S520, the matching degree can refer to an indicator used to quantify the similarity between a gait subset and a historical subset. It is typically negatively correlated with the feature distance; that is, the smaller the feature distance, the higher the matching degree, and the larger the feature distance, the lower the matching degree. In other words, a matching degree closer to 100% indicates a greater consistency between the current gait and the historical gait, while a lower matching degree indicates a greater difference between the current gait and the historical gait. This embodiment does not impose any restrictions on the specific calculation method of the matching degree, as long as it can reflect the similarity between the gait subset and the historical subset.

[0166] In step S530, the third threshold can refer to a matching degree threshold used to determine whether the current gait deviates from the historical normal gait. When the matching degree is lower than this threshold, it indicates that there is a significant difference between the current gait and the historical gait, and the gait calibration process needs to be triggered. When the matching degree is higher than or equal to this threshold, it indicates that the current gait is within the normal range and no adjustment is required.

[0167] As an example, the third threshold is set to 85%. Following the matching degree calculated in step S520, which is 89.9%, since 89.9% is higher than 85%, the gait subset is determined to match the historical subset, and the current gait is within the normal range; therefore, no gait calibration process needs to be triggered. If, in another scenario, the calculated matching degree is 80%, which is lower than 85%, then the gait subset is determined to not match the historical subset, and the system will trigger a gait sequence calibration process to extract outliers from the significantly different gait parameters and determine the key areas for calibration. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0168] It is understood that the specific value of the third threshold can be personalized based on factors such as user physical characteristics, walking habits, and environmental requirements. For example, for elderly users or users in the rehabilitation period, the threshold can be appropriately increased (e.g., 90%) to provide greater safety redundancy, while for skilled users, the threshold can be appropriately decreased (e.g., 80%) to reduce unnecessary gait adjustments. This embodiment does not impose any restrictions on the specific value of the third threshold.

[0169] In this embodiment, by employing the Euclidean distance algorithm to calculate the feature distance between the gait subset and the historical subset, the geometric difference between the current gait and the historical gait in the feature space can be quantified, providing an objective mathematical metric for gait similarity judgment. Furthermore, the matching degree between the gait subset and the historical subset is determined based on the feature distance, transforming the abstract feature distance into an intuitive percentage-based matching degree index. This facilitates threshold comparison and decision-making within the system, improving the interpretability and operability of gait difference assessment.

[0170] Moreover, in this embodiment, by determining that the gait subset does not match the historical subset in response to the matching degree being lower than the third threshold, an automatic detection and triggering mechanism for gait abnormality is realized. The third threshold can be personalized according to factors such as user physical characteristics, walking habits, and environmental requirements, making gait abnormality judgment flexible and adaptable.

[0171] In one embodiment, see Figure 8 Step S500 includes:

[0172] Step S540: Collect muscle signals from the user's lower limbs.

[0173] Step S550: Use the Kalman filter algorithm to perform noise reduction processing on the user's lower limb muscle signals.

[0174] Step S560: Analyze the noise-reduced lower limb muscle signals and user action data using a deep learning model to predict the user's walking intention.

[0175] In step S540, the lower limb muscle signal can refer to the electrophysiological signal generated by the user's lower limb muscle activity collected by the electromyography sensor, which can reflect the user's muscle contraction state and movement intention. Common lower limb muscle collection sites include the quadriceps femoris, hamstrings, tibialis anterior, and gastrocnemius muscles. The force exertion pattern and temporal changes of these muscles during walking are closely related to the user's walking intention.

[0176] As an example, surface electromyography (EMG) electrodes are attached to the quadriceps and gastrocnemius muscles of the user's lower limbs, and the sampling frequency is set to 1000 Hz to collect EMG signals in real time during the user's walking process. The amplitude of the EMG signal is typically in the microvolt range, and the raw signal contains various noises such as power frequency interference and motion artifacts. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0177] It is understandable that electromyography (EMG) signals have an earlier predictive advantage compared to motion data collected by inertial measurement units (IMUs), because muscle activation typically occurs about 30 to 100 milliseconds before actual joint movement. Therefore, EMG signals can reflect the user's movement intentions in advance, providing a time window for early adjustment of gait strategies.

[0178] In step S550, the Kalman filter algorithm can refer to a recursive filtering algorithm that uses the linear system state equation and observation data to optimally estimate the system state, effectively filtering out noise and extracting the true signal from noisy observation signals. Since electromyography (EMG) signals are easily affected by environmental electromagnetic interference, electrode motion artifacts, power frequency interference, and other factors during acquisition, resulting in a low signal-to-noise ratio, directly using the raw signal for analysis will lead to a decrease in the accuracy of intent recognition. Therefore, noise reduction processing is necessary.

[0179] As an example, we define the state equation of the Kalman filter algorithm as the time evolution model of muscle activation state, and the observation equation as the noisy signal collected by the electromyography (EMG) sensor. The Kalman filter algorithm recursively estimates the EMG signal through two steps: prediction and updating. First, it predicts the muscle activation state at the current moment based on the state at the previous moment. Then, it corrects the prediction result using the observation value at the current moment, obtaining the filtered muscle activation signal. After Kalman filtering, high-frequency noise and motion artifacts in the original EMG signal are effectively suppressed, and the signal-to-noise ratio is significantly improved. It can be understood that using the Kalman filter algorithm for noise reduction can remove noise while retaining the feature components related to movement intention in the EMG signal, providing a high-quality input signal for subsequent deep learning model analysis.

[0180] In step S560, the user's walking intention can refer to the user's movement tendency inferred by parsing the user's lower limb muscle signals and user action data, including but not limited to intention categories such as acceleration, deceleration, starting, stopping, turning, going uphill, going downhill, and crossing obstacles.

[0181] Deep learning models can refer to pre-trained neural network models, such as convolutional neural networks, long short-term memory networks, or a combination of both, which can automatically extract features from time-series signals and perform classification or regression predictions. When using long short-term memory networks to analyze lower limb muscle signals to predict a user's walking intention, a lightweight long short-term memory network model can be used, and inference can be accelerated through dedicated hardware accelerators to reduce processing latency and improve computational parallelism.

[0182] As an example, a long short-term memory (LSTM) network-based intention recognition model can be constructed by combining temporal modeling techniques from visual intelligence. The model's input includes denoised electromyography (EMG) signals (such as temporal and frequency domain features of the quadriceps and gastrocnemius muscles) and user motion data (such as acceleration and angular velocity data collected by an inertial measurement unit). The output is a probability distribution of the user's walking intention. During model training, a large number of user EMG signals, motion data, and their corresponding real intention labels are collected in different walking scenarios. The LSM network is then subjected to supervised training to learn the mapping relationship between EMG signals, motion data, and walking intention. In practical use, the denoised EMG signals from step S520 and the user motion data collected by the inertial measurement unit are synchronously input into the trained deep learning model. The model outputs probability values ​​for various intentions and selects the intention with the highest probability as the current user's walking intention. For example, when the model outputs a probability of 0.85 for an uphill intention, 0.10 for a flat ground intention, and 0.05 for a downhill intention, the predicted user walking intention is uphill. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0183] It is understandable that by fusing information from both electromyography (EMG) signals and motion data, deep learning models can capture users' movement intentions more comprehensively and accurately. Compared to intention recognition methods that rely solely on motion data, fusing EMG signals can predict intentions in advance before the user's movements occur, providing a lead time for gait strategy adjustments.

[0184] In this embodiment, by collecting lower limb muscle signals from the user, electrophysiological information reflecting the user's muscle contraction state and movement intention can be obtained. Electromyography (EMG) signals have an earlier predictive advantage compared to motion data, providing a time window for advance adjustment of gait strategies. Simultaneously, by employing a Kalman filter algorithm to denoise the user's lower limb muscle signals, power frequency interference and motion artifacts in the original EMG signals can be effectively filtered out, significantly improving the signal-to-noise ratio while retaining feature components related to movement intention, providing high-quality input signals for deep learning model analysis.

[0185] In this embodiment, the user's walking intention is predicted by analyzing the denoised lower limb muscle signals and user motion data based on the deep learning model. This integrates information from two modalities: electromyography signals and motion data. This allows the intention recognition model to comprehensively utilize the user's neurophysiological signals and physical motion data, achieving accurate prediction of various walking intentions such as acceleration, deceleration, uphill, and downhill. Compared with intention recognition methods that rely solely on motion data, this method has higher prediction accuracy and earlier prediction timing.

[0186] In one embodiment, see Figure 9 Step S600 includes:

[0187] Step S610: Correct the temporary gait adjustment strategy based on the reacquired user action data to obtain the final gait adjustment strategy: calculate the deviation between the predicted gait sequence in the temporary gait adjustment strategy and the reacquired user action data.

[0188] Step S620: In response to the deviation value exceeding the fourth threshold, the Kalman filter algorithm is used to smooth the user action data acquired again.

[0189] Step S630: Use the gradient descent algorithm to adjust the control parameters in the gait adjustment temporary strategy.

[0190] Step S640: Based on the inverse kinematics algorithm, the smoothed user action data, and the adjusted gait adjustment temporary strategy, recalculate the target angles of each joint of the exoskeleton robot to obtain the final gait adjustment strategy.

[0191] In step S610, the predicted gait sequence can refer to the time series data that is pre-generated in the gait adjustment temporary strategy and describes the expected motion trajectory of the exoskeleton robot, including information such as joint angles, joint angular velocities, and foot positions at each time point.

[0192] The user action data acquired again can refer to the actual user action data collected in real time during the execution of the temporary gait adjustment strategy. The deviation value can refer to the degree of difference between the predicted gait sequence and the actual action data, and is used to measure the effectiveness of the temporary strategy.

[0193] As an example, during the execution of a temporary gait adjustment strategy, the system collects the angle data of the user's left knee joint in real time at a frequency of 100 times per second as the user's action data for subsequent acquisitions. Assuming that at the current time point, the predicted left knee joint angle in the temporary gait adjustment strategy is 35 degrees, while the actual collected left knee joint angle is 30 degrees, the deviation value is 5 degrees. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0194] Understandably, a smaller deviation value indicates a better execution effect of the temporary strategy, and a closer match between the actual and expected actions. A larger deviation value indicates a significant difference between the temporary strategy and the actual execution, requiring modification of the temporary strategy.

[0195] In step S620, the fourth threshold can refer to a critical value used to determine whether the deviation value is large enough to trigger a correction mechanism. The Kalman filter algorithm can be a recursive filtering algorithm that uses the state equations of a linear system to optimally estimate the observed data, capable of filtering out random noise and extracting the true signal from noisy observed data.

[0196] Since the user action data collected by the sensor inevitably contains measurement noise, directly using the raw data for correction may lead to jitter and instability in the control commands. Therefore, it is necessary to smooth the data before correction. As an example, the fourth threshold is set to 3 degrees. Continuing from the example in step S610, the calculated deviation value is 5 degrees, which is greater than the fourth threshold of 3 degrees, thus triggering the correction mechanism. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0197] A Kalman filter algorithm is used to smooth the real-time acquired joint angle sequence. The state equation is set as the motion model of the joint angles, and the observation equation is set as the measurement model of the sensors. The true joint angle values ​​are recursively estimated through two steps: prediction and update. After Kalman filtering, high-frequency noise in the original measurement data is effectively suppressed, resulting in a smoothed joint angle variation curve.

[0198] In step S630, the gradient descent algorithm can refer to a first-order optimization algorithm that finds optimal parameters by iteratively optimizing the objective function. The control parameters can refer to the key parameters used to generate gait control commands in the temporary gait adjustment strategy, and may include, but are not limited to, stride adjustment, cadence adjustment, joint angle correction, and joint torque distribution ratio. The gradient descent algorithm can be an algorithm that progressively updates parameters along the negative direction of the objective function's gradient, continuously decreasing the objective function value until it converges to the optimal solution.

[0199] As an example, the objective function is defined as the mean squared error between the predicted gait sequence and the actual motion data. The optimization objective is to minimize this mean squared error by adjusting the control parameters. The initial values ​​for the stride length adjustment coefficient and the cadence adjustment coefficient are set to 1.0 and 1.0 respectively. The partial derivative (gradient) of the objective function with respect to the stride length adjustment coefficient is calculated, yielding a current gradient value of 0.3. With a learning rate of 0.1, the update amount for the stride length adjustment coefficient is -0.1 × 0.3 = -0.03, resulting in an updated stride length adjustment coefficient of 0.97. The update of the cadence adjustment coefficient is calculated similarly. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0200] After multiple iterations, the objective function value gradually converged, and the control parameters were optimized to the optimal values ​​that make the predicted gait sequence closer to the actual motion data.

[0201] In step S640, the inverse kinematics algorithm can refer to a calculation method that solves for the angles of each joint based on the desired pose of the robot's end effector (such as the foot). The ultimate gait adjustment strategy can refer to the final gait control scheme obtained by modifying and optimizing the temporary gait adjustment strategy based on real-time feedback data, which has higher environmental adaptability and motion matching degree compared to the temporary strategy.

[0202] As an example, based on the adjusted gait adjustment temporary strategy in step S630, the desired trajectory of the foot (including the position and posture sequence of the foot in three-dimensional space) is determined. Using smoothed user motion data as constraints (such as maintaining the trend characteristics of the user's actual joint angles), and combined with the kinematic model of the exoskeleton robot (including parameters such as link lengths and joint degrees of freedom), an inverse kinematics algorithm is used to solve for the target angle sequences of the hip, knee, and ankle joints required to achieve the desired foot trajectory. It is assumed that the calculated target angle of the left knee joint has a peak value of 65 degrees in the swing phase and 5 degrees in the stance phase, with the target angle of the right knee joint being symmetrically matched. The calculated target angle sequences of each joint are integrated to form a complete gait control command, which is the ultimate gait adjustment strategy. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0203] In this embodiment, by calculating the deviation between the predicted gait sequence in the temporary gait adjustment strategy and the reacquired user action data, a real-time quantitative evaluation of the temporary strategy's execution effect is achieved, providing a clear triggering basis and correction direction for subsequent strategy correction. Simultaneously, by employing a Kalman filter algorithm to smooth the reacquired user action data in response to the deviation exceeding a fourth threshold, high-frequency noise and random interference in the sensor measurement data can be filtered out. This provides smooth and reliable input data for subsequent parameter optimization and joint angle calculation, avoiding control command jitter and instability caused by noise.

[0204] In this embodiment, by employing the gradient descent algorithm to adjust the control parameters in the temporary gait adjustment strategy, the optimal control parameters that best match the predicted gait sequence with the actual motion data can be automatically found through iterative optimization, achieving automatic optimization and adaptive adjustment of the control parameters. Furthermore, by recalculating the target angles of each joint of the exoskeleton robot based on the inverse kinematics algorithm, smoothed user motion data, and the adjusted temporary gait adjustment strategy, the final gait adjustment strategy is obtained. The optimized control parameters are then converted into an executable sequence of target angles for each joint, forming a complete closed loop from strategy correction to joint command generation. This ensures that the final gait adjustment strategy simultaneously meets the requirements of environmental adaptability and user motion matching.

[0205] In one embodiment, see Figure 10 Step S700 includes:

[0206] Step S710: Extract the optimized control parameters from the gait adjustment final strategy.

[0207] Step S720: Using the optimized control parameters as feedback signals, adjust the feature extraction weights of the corresponding user action data and the fusion coefficients of the corresponding terrain feature data in the gait model to obtain the adjusted gait model.

[0208] In step S710, the optimized control parameters can refer to the set of optimal gait control parameters obtained after correction and optimization in step S600. These parameters have been verified by actual execution results and can effectively adapt to the current environmental complexity and user walking status.

[0209] The types of control parameters include, but are not limited to, stride adjustment coefficient, cadence adjustment coefficient, joint torque distribution coefficient, center of gravity offset compensation coefficient, gait phase switching threshold coefficient, terrain adaptation sensitivity coefficient, and energy consumption optimization coefficient.

[0210] As an example, following the optimization results of step S600, the following optimized control parameters are extracted from the gait adjustment final strategy: stride adjustment coefficient is optimized from the default value of 1.0 to 0.85, cadence adjustment coefficient is optimized from the default value of 1.0 to 0.94, knee joint torque distribution coefficient is optimized from the default value of 0.3 to 0.5, center of gravity offset compensation coefficient is optimized from the default value of 1.0 to 1.1, and terrain adaptation sensitivity coefficient is optimized from the default value of 1.0 to 1.2. The above data are for illustrative purposes only and do not limit the scope of protection of this embodiment.

[0211] It is understandable that these optimized control parameters reflect the optimal gait configuration under the current terrain conditions (such as a 15-degree uphill slope) and user status, and have important reference value and reuse significance.

[0212] In step S720, the gait model can refer to a mathematical model used to describe and predict the gait behavior of an exoskeleton robot, which contains multiple adjustable parameters. Feature extraction weights can refer to the importance coefficients assigned by the gait model to different feature dimensions when extracting gait features from user action data. Fusion coefficients can refer to the weight coefficients assigned by the gait model to different data sources when fusing user action data and terrain feature data. By adjusting these weights and coefficients, the gait model can focus more on features and data sources that contribute significantly to the current scene during subsequent operation, thereby improving the model's adaptability and accuracy.

[0213] As an example, when the gait model extracts gait features from user motion data, the original feature extraction weights are set as follows: stride length feature weight is 0.3, cadence feature weight is 0.3, joint angle feature weight is 0.2, and center of gravity shift feature weight is 0.2. Based on the optimized control parameters extracted in step S710, it was found that the stride length adjustment coefficient changed significantly from 1.0 to 0.85, the cadence adjustment coefficient changed slightly from 1.0 to 0.94, while the joint torque distribution coefficient changed significantly. Based on this analysis, in the current uphill scenario, stride length control and joint torque distribution contribute significantly to gait stability. Therefore, the feature extraction weights corresponding to user motion data in the gait model are adjusted as follows: stride length feature weight is increased to 0.4, joint angle feature weight is increased to 0.3, cadence feature weight is decreased to 0.2, and center of gravity shift feature weight is decreased to 0.1. Simultaneously, when the gait model fuses user motion data and terrain feature data, the original fusion coefficients are set as follows: user motion data weight is 0.5, and terrain feature data weight is 0.5. The optimized control parameters show that the terrain adaptation sensitivity coefficient has increased from 1.0 to 1.2, indicating that terrain features have a greater impact on gait adjustment in the current environment. Therefore, the fusion coefficient of terrain feature data is increased to 0.6, and the fusion coefficient of user action data is decreased to 0.4. After the above adjustments, an optimized gait model is obtained. The above data is for illustrative purposes only and does not limit the scope of protection of this embodiment.

[0214] It is understandable that the adjustment of feature extraction weights and fusion coefficients can be achieved in various ways, such as rule-based methods, gradient descent-based optimization methods, or reinforcement learning-based methods. The adjustment range of weights and coefficients can be determined based on the degree of change of the optimized control parameters relative to their default values; the greater the change in parameters, the greater the increase in the weight of the corresponding feature or data source. By using the validated optimized control parameters as feedback signals to feed back into the gait model, the gait model can learn from successful experiences and directly generate a better initial gait strategy when encountering similar scenarios in the future, reducing the number of iterations and improving the system's response speed.

[0215] In this embodiment, by extracting optimized control parameters from the ultimate gait adjustment strategy, the effective strategy parameters, verified through actual execution, are separated from the temporary control scheme, providing high-quality learning samples for the continuous optimization of the gait model. Simultaneously, by using the optimized control parameters as feedback signals to adjust the feature extraction weights of the corresponding user action data in the gait model, the gait model can dynamically adjust its focus on different gait features according to different scenarios. In scenarios where the model has been verified to be effective, the weights of key features are increased, enabling the model to extract features that contribute significantly to gait stability more quickly in subsequent similar scenarios.

[0216] In this embodiment, by adjusting the feature extraction weights of the corresponding user action data and the fusion coefficients of the corresponding terrain feature data in the gait model, the adaptive optimization of the multi-source data fusion strategy is achieved. This enables the gait model to dynamically adjust the relative importance of user action data and terrain feature data in decision-making according to different scenarios. The perception results of visual intelligence are used as the optimization basis to optimize the fusion weight configuration in scenarios where the verification is effective.

[0217] Furthermore, by using the optimized control parameters as feedback signals to feed back into the gait model, a complete closed-loop learning mechanism from policy execution to model optimization is formed. By integrating the continuous learning capabilities of visual intelligence, the gait model can continuously learn and evolve from successful experiences. When encountering similar scenarios in the future, it can directly call the optimized weight configuration to generate a better initial gait strategy, reducing the number of iterations and enabling the relevant algorithms of the exoskeleton robot to have the learning ability from visual perception to gait adaptation, thereby improving the system's response speed and intelligence level.

[0218] In one embodiment, see Figure 11 This application also provides a gait model adjustment system for an exoskeleton robot. The gait model adjustment system for the exoskeleton robot includes:

[0219] The data acquisition module is used to acquire the motion data set of the exoskeleton robot. The motion data set includes at least user motion data and terrain feature data.

[0220] The initial evaluation module is used to input user action data and terrain feature data into the gait model, obtain slope data and obstacle density data of the area in front of the exoskeleton robot, and generate an initial evaluation result of environmental complexity based on the slope data and obstacle density data.

[0221] The final evaluation module is used to detect the frequency of change of the action data set, and in response to fluctuations in the frequency of change of the action data set, it generates the final evaluation result of the environment complexity in combination with the initial evaluation result of the environment complexity.

[0222] The subset acquisition module is used to acquire gait subsets and corresponding historical subsets from user action data in response to safety hazards in the environment in front of the exoskeleton robot.

[0223] The temporary strategy generation module is used to respond to the mismatch between the gait subset and the historical subset, determine the user's walking intention based on the user's action data, and determine the temporary gait adjustment strategy based on the final evaluation result of environmental complexity.

[0224] The strategy correction module is used to acquire user action data again and correct the temporary gait adjustment strategy based on the acquired user action data to obtain the final gait adjustment strategy.

[0225] The model adjustment module is used to adjust the coefficient values ​​of at least some of the exoskeleton robot gait generation logic in the gait model in response to the matching of the gait adjustment ultimate strategy and the environmental complexity ultimate evaluation result, so as to obtain the adjusted gait model.

[0226] The description of the gait model adjustment system for exoskeleton robots can be found in the aforementioned gait model adjustment method for exoskeleton robots. Its functions will not be repeated here, but can be found in the detailed description of the above method embodiments.

[0227] In this embodiment, by fusing user motion data with terrain feature data to form a gait model, comprehensive information on ground slope changes and obstacle distribution is obtained, solving the problem of incomplete information from a single sensor. Simultaneously, this embodiment continuously monitors the frequency of changes in the motion data set. Once a fluctuation exceeds a preset threshold, a dynamic disturbance assessment is immediately triggered, incorporating environmental change factors into the evaluation result to generate a more accurate final evaluation of environmental complexity. This mechanism enables the exoskeleton robot to perceive changes during sudden environmental shifts, rather than reacting passively after gait errors occur, effectively overcoming the slow response issues inherent in related technologies.

[0228] Furthermore, in this embodiment, the final evaluation result of environmental complexity, gait subset, historical subset, and user intent information are fused together for judgment, so that the exoskeleton robot can make an overall judgment by combining the terrain in front and the user's current posture at the same time. This avoids gait misjudgment caused by insufficient information in a single dimension, and improves the gait detection accuracy, dynamic response speed, and walking safety of the exoskeleton robot in complex terrain.

[0229] Finally, in this embodiment, user action data is acquired again as feedback, and the temporary strategy is corrected through an adaptive parameter update mechanism to generate a more accurate final strategy. Furthermore, after confirming that the final strategy matches the environmental complexity evaluation results, the method adjusts the coefficient values ​​in the gait model (such as stride adjustment coefficient, gait frequency adjustment coefficient, joint torque distribution coefficient, center of gravity shift compensation coefficient, etc.) so that the gait model can directly call the optimized parameters in subsequent similar terrain conditions. This closed-loop feedback and continuous model optimization mechanism enables the exoskeleton robot to continuously and intelligently improve with changes in the environment and user state, rather than recalculating each time, thus achieving continuous adaptability in complex environments.

[0230] It is understood that the deep learning models (including but not limited to object detection models and intent recognition models) and data processing algorithms (including but not limited to principal component analysis algorithms and clustering algorithms) involved in the above embodiments can achieve real-time inference using lightweight methods and hardware acceleration schemes known in the art when deployed on an exoskeleton robot embedded platform. Specifically, the deep learning model can use a lightweight network structure to replace the standard network structure to reduce the forward computation and parameter data volume; the parameters of the deep learning model can be deployed on the embedded platform after INT8 quantization to reduce model memory usage and improve inference speed; the inference process of the deep learning model can be hardware-accelerated by the neural network processing unit on the embedded platform. The principal component analysis algorithm is a linear dimensionality reduction method with low computational complexity; the clustering algorithm is only executed at a low frequency in the point cloud preprocessing stage, and the amount of point cloud data is limited after voxel filtering and downsampling. Through the above optimization methods, the inference latency of each algorithm model can meet the real-time requirements of the exoskeleton robot motion control closed loop, achieving the technical effect of low latency and high safety.

[0231] This application also provides an electronic device, which includes one or more processors and memory resources represented by memory for storing instructions executable by the processor, such as application programs. The application programs stored in the memory may include one or more modules, each corresponding to a set of instructions.

[0232] The electronic device may also include a power supply component configured to perform power management of the electronic device, a wired or wireless network interface configured to connect the electronic device to a network, and an input / output (I / O) interface. The electronic device can be operated based on operating devices stored in memory, such as Windows Server™, Mac OS X™, Unix™, Linux™, FreeBSD™, or similar.

[0233] In one embodiment, a computer device, which may be a server, is also provided. The computer device includes a processor, memory, input / output interfaces (I / O), and a communication interface. The processor, memory, and I / O interfaces are connected via a system bus, and the communication interface is connected to the system bus via the I / O interfaces. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes non-volatile storage media and internal memory. The non-volatile storage media stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage media. The database of the computer device stores data. The I / O interfaces of the computer device are used for exchanging information between the processor and external devices. The communication interface of the computer device is used for communicating with external terminals via a network connection.

[0234] In one embodiment, a computer device is provided, which may be a terminal. The computer device includes a processor, memory, an input / output interface, a communication interface, a display unit, and an input device. The processor, memory, and input / output interface are connected via a system bus, and the communication interface, display unit, and input device are also connected to the system bus via the input / output interface. The processor of the computer device provides computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The input / output interface of the computer device is used for exchanging information between the processor and external devices. The communication interface of the computer device is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. The display unit of the computer device is used to form a visually visible image and may be a display screen, a projection device, or a virtual reality imaging device. The display screen can be an LCD screen or an e-ink screen. The input device of the computer device can be a touch layer covering the display screen, or buttons, trackballs, or touchpads set on the casing of the computer device, or external keyboards, touchpads, or mice, etc.

[0235] This application also provides a non-transitory computer-readable storage medium, which, when the instructions in the storage medium are executed by the processor of the electronic device, enables the electronic device to perform the gait model adjustment of the exoskeleton robot.

[0236] This application may take the form of a computer program product implemented on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing program code. Computer-readable storage media include permanent and non-permanent, removable and non-removable media, and information storage can be implemented by any method or technology. Information may be computer-readable instructions, data structures, program modules, or other data. Examples of computer storage media include, but are not limited to: phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic magnetic disk storage or other magnetic storage devices, or any other non-transfer medium that can be used to store information accessible by a computing device.

[0237] It should be noted that although the steps of the method in this application are described in a specific order in the accompanying drawings, this does not require or imply that these steps must be performed in that specific order, or that all the steps shown must be performed to achieve the desired result. Additional or alternative steps, such as omitting certain steps, combining multiple steps into one step, and / or breaking down one step into multiple steps, should all be considered part of this application.

[0238] It should be understood that this application is not limited to the detailed data of the methods presented in this specification. This application is capable of other implementations and can be implemented and performed in various ways. The foregoing variations and modifications fall within the scope of this application. It should be understood that the disclosure and definition of this application extend to all alternative combinations of two or more individual features mentioned or apparent in the text and / or drawings. All these different combinations constitute multiple alternative aspects of this application. The embodiments described in this specification illustrate the best known mode for implementing this application and will enable those skilled in the art to utilize this application.

Claims

1. A method for adjusting the gait model of an exoskeleton robot, characterized in that, The method includes: Acquire a set of motion data for the exoskeleton robot, wherein the set of motion data includes at least user motion data and terrain feature data; The user action data and the terrain feature data are input into the gait model to obtain the slope data and obstacle density data of the area in front of the exoskeleton robot, and an initial evaluation result of environmental complexity is generated based on the slope data and obstacle density data. The frequency of change of the action data set is detected, and in response to fluctuations in the frequency of change of the action data set, a final evaluation result of the environmental complexity is generated by combining the initial evaluation result of the environmental complexity. In response to a safety hazard in the environment in front of the exoskeleton robot, a gait subset and a corresponding historical subset are obtained from the user's motion data. In response to a mismatch between the gait subset and the historical subset, the user's walking intention is determined based on the user action data, and a temporary gait adjustment strategy is determined in conjunction with the final environmental complexity evaluation result. The user action data is acquired again, and the gait adjustment temporary strategy is corrected based on the acquired user action data to obtain the gait adjustment final strategy; In response to the matching between the ultimate gait adjustment strategy and the ultimate environmental complexity evaluation result, the coefficient values ​​of at least some of the components in the gait model used to control the gait generation logic of the exoskeleton robot are adjusted according to the ultimate gait adjustment strategy to obtain the adjusted gait model.

2. The gait model adjustment method for the exoskeleton robot according to claim 1, characterized in that, The step of responding to the matching between the gait adjustment ultimate strategy and the environmental complexity ultimate evaluation result, adjusting at least some of the coefficient values ​​in the gait model used to control the gait generation logic of the exoskeleton robot according to the gait adjustment ultimate strategy, to obtain the adjusted gait model, includes: Based on the gait adjustment final strategy and the environmental complexity final evaluation results, confirm whether the safety hazards have been eliminated; In response to the elimination of the safety hazard, it is determined that the ultimate gait adjustment strategy matches the ultimate environmental complexity evaluation result, and at least some of the coefficient values ​​in the gait model used to control the gait generation logic of the exoskeleton robot are adjusted according to the ultimate gait adjustment strategy to obtain the adjusted gait model.

3. The gait model adjustment method for the exoskeleton robot according to claim 2, characterized in that, The step of confirming whether safety hazards have been eliminated based on the gait adjustment final strategy and the environmental complexity final evaluation result includes: Calculate the gait stability improvement rate based on the gait stability data before and after implementing the ultimate gait adjustment strategy; Calculate the risk mitigation rate based on the risk index before and after implementing the gait adjustment ultimate strategy; When both the gait stability improvement rate and the risk mitigation rate exceed the first threshold, it is confirmed that the safety hazard has been eliminated.

4. The gait model adjustment method for the exoskeleton robot according to claim 1, characterized in that, The user action data is collected through an inertial measurement unit; The terrain feature data was collected using lidar and depth cameras; The step of inputting the user action data and the terrain feature data into the gait model to obtain the slope data and obstacle density data of the area in front of the exoskeleton robot includes: The Kalman filter algorithm is used to perform time synchronization and spatial registration of the user action data and the terrain feature data; A deep learning model is used to identify the locations of obstacles in the terrain feature data; The point cloud data in the terrain feature data is converted into a three-dimensional grid map, and the slope value of each grid is calculated to obtain the slope data. Clustering algorithms are used to identify obstacle boundaries in the terrain feature data. The number of obstacles per unit area is counted based on the obstacle location and the obstacle boundary to obtain the obstacle density data.

5. The gait model adjustment method for the exoskeleton robot according to claim 1, characterized in that, The process involves detecting the frequency of changes in the action data set, and responding to fluctuations in the frequency of changes in the action data set. Combined with the initial environmental complexity evaluation result, a final environmental complexity evaluation result is generated, including: A time-series anomaly detection method based on a sliding window is used to calculate the change frequency of the action data set; When the change frequency exceeds the second threshold, it is determined that the change frequency of the action data set is fluctuating, and a Bayesian probability model is used to calculate the probability of interference in combination with the change frequency, and a random forest algorithm is used to calculate the risk index in the future time period. Based on the probability of interference and the risk index, the final evaluation result of environmental complexity is generated.

6. The gait model adjustment method for the exoskeleton robot according to claim 1, characterized in that, The step of obtaining the gait subset and the corresponding historical subset from the user action data includes: By combining principal component analysis algorithm with the user action data, key gait features are extracted to obtain the gait subset; Historical action data corresponding to the current time interval is obtained from the historical database, and the key features of historical gait in the historical action data are extracted using the principal component analysis algorithm to obtain the historical subset.

7. The gait model adjustment method for the exoskeleton robot according to claim 6, characterized in that, The response to a mismatch between the gait subset and the history subset includes: The feature distance between the gait subset and the history subset is calculated using the Euclidean distance algorithm; The matching degree between the gait subset and the history subset is determined based on the feature distance; In response to the matching degree being lower than a third threshold, it is determined that the gait subset does not match the history subset.

8. The gait model adjustment method for the exoskeleton robot according to claim 1, characterized in that, Determining the user's walking intention based on the user's action data includes: Collect signals from the user's lower limb muscles; The Kalman filter algorithm was used to reduce noise in the user's lower limb muscle signals. Based on the analysis of the noise-reduced lower limb muscle signals and user movement data using a deep learning model, the user's walking intention is predicted.

9. The gait model adjustment method for the exoskeleton robot according to claim 1, characterized in that, The step of revising the temporary gait adjustment strategy based on the re-acquired user action data to obtain the final gait adjustment strategy includes: Calculate the deviation between the predicted gait sequence in the temporary gait adjustment strategy and the reacquired user action data; In response to the deviation value exceeding the fourth threshold, the Kalman filter algorithm is used to smooth the user action data acquired again; The control parameters in the temporary gait adjustment strategy are adjusted using the gradient descent algorithm. Based on the inverse kinematics algorithm, smoothed user motion data, and the adjusted gait adjustment temporary strategy, the target angles of each joint of the exoskeleton robot are recalculated to obtain the ultimate gait adjustment strategy.

10. A gait model adjustment system for an exoskeleton robot, characterized in that, include: The data acquisition module is used to acquire a set of motion data of the exoskeleton robot, which includes at least user motion data and terrain feature data; The initial evaluation module is used to input the user action data and the terrain feature data into the gait model, obtain the slope data and obstacle density data of the area in front of the exoskeleton robot, and generate an initial evaluation result of environmental complexity based on the slope data and obstacle density data. The final evaluation module is used to detect the frequency of change of the action data set, and in response to fluctuations in the frequency of change of the action data set, it generates a final evaluation result of the environmental complexity based on the initial evaluation result of the environmental complexity. The subset acquisition module is used to acquire a gait subset and a historical subset corresponding to the gait subset in response to a safety hazard in the environment in front of the exoskeleton robot. A temporary strategy generation module is used to determine the user's walking intention based on the user action data and to determine a temporary gait adjustment strategy in response to a mismatch between the gait subset and the historical subset, and in conjunction with the final environmental complexity evaluation result. The strategy correction module is used to acquire user action data again and correct the gait adjustment temporary strategy based on the acquired user action data to obtain the gait adjustment final strategy. The model adjustment module is used to adjust the coefficient values ​​of at least some of the gait models used to control the gait generation logic of the exoskeleton robot according to the gait adjustment final strategy and the environmental complexity final evaluation result in response to the gait adjustment final strategy, so as to obtain the adjusted gait model.

Images