A robot motion control method based on state perception and adaptive adjustment

Abstract

The application provides a robot motion control method based on state sensing and adaptive adjustment, and relates to the technical field of intelligent rehabilitation medical instruments. The method is used for obtaining an initial parameter set; based on multi-source sensing information, a multi-source sensing data set is obtained; based on wheel running state data, equivalent input disturbance received by each wheel is calculated to obtain a road surface friction coefficient estimation value of each wheel and a road surface sensing state in which the current device is located; based on human physiological feature data, pattern recognition is performed to obtain a muscle strength sensing state of the current user; based on device interaction mechanics data and an initial barycenter position, a dynamic barycenter compensation vector is calculated to obtain a real-time dynamic barycenter sensing state; based on a target motion control instruction, motion control is performed on a standing auxiliary robot, a coefficient is updated according to an execution feedback result, and a motion control result is obtained. The application solves the problems of insufficient adaptability, limited prediction accuracy and high risk of slipping and falling on complex road surfaces in the prior art.

Description

Technical Field

[0001] This invention relates to the field of intelligent rehabilitation medical device technology, and in particular to a robot motion control method based on state perception and adaptive adjustment. Background Technology

[0002] As a key device for improving the autonomy of the elderly and people with mobility impairments, standing-assistance robots have a core requirement: to maximize the user's active muscle engagement while ensuring safety, thus avoiding muscle weakness caused by long-term reliance on the device. Existing technologies typically employ an assistance mode based on a preset reference trajectory. This establishes a safe range of motion around the reference trajectory. When the user's posture is determined to be within the safe tolerance range, the user is allowed to move autonomously to a certain extent according to their wishes, thereby promoting active muscle engagement and achieving muscle strength maintenance and functional training. Conversely, if a deviation from the posture is detected or if the posture is about to exceed the safe boundary, active corrective force is applied to guide the user's posture back to the safe range, thus prioritizing the safety of the movement process and effectively preventing the risk of falls.

[0003] In terms of user posture perception, existing technologies employ low-cost sensors for posture estimation, primarily using laser rangefinders and force sensors as measurement elements. Laser rangefinders are used to acquire point cloud data of the user's body contour in the sagittal plane, simplifying the human body model into a multi-link model (typically including three link segments: torso, thigh, and lower leg) to estimate the body's center of gravity position in real time. The coordinates of the center of gravity P can be calculated by weighted averaging of the mass of each link segment and its center of gravity position.

[0004] Furthermore, by integrating force sensors on both sides of the robot's handrail, the system can detect the difference in force distribution applied by the user's left and right arms, thereby indirectly estimating the degree of lateral tilt of the torso, i.e., the lateral positional shift of the center of gravity. Studies have shown that as the lateral shift increases, the system's stability margin decreases, and the safety tolerance range decreases with increasing lateral shift, at which point the risk of lateral imbalance and falls increases. Only when the center of gravity position P is within the safety tolerance range and the body tilt is below a certain threshold is the current posture considered to meet the stability requirements for maintaining balance during movement, with no risk of lateral falls, and normal movement is permitted.

[0005] Existing research indicates that human motor control during walking relies on the central nervous system coordinating the activity of multiple muscles through a limited set of muscle synergists. When walking on complex or uneven surfaces, to overcome the effects of disturbances and maintain stability, the body activates a more robust synergistic mode, shifting the motor control strategy from a "precise control mode" to a "robust control mode." This process is accompanied by higher metabolic costs. During exercise, control parameters are preset and fixed based on different degrees of trunk extension. When the trunk is not fully extended, the system uses a smaller damping coefficient and a larger position control coefficient to limit the user's voluntary range of motion and enhance safety. Once the trunk is fully extended, the system uses a larger damping coefficient and a smaller position control coefficient to encourage active muscle engagement. This hybrid control strategy achieves a static balance between encouraging active participation and ensuring movement safety, making it suitable for exercise in standard environments.

[0006] Limitations of existing technology: Fixed control parameters lead to insufficient adaptability. Existing technologies use fixed values ​​for damping and position control coefficients only during the standing phase, failing to consider individual user differences. Significant differences in muscle strength exist among users; using fixed parameters may excessively restrict the voluntary movement of users with stronger muscles, while providing insufficient assistance to users with weaker muscles, potentially even posing safety risks.

[0007] The simplified center of gravity calculation model has limited accuracy. Existing technology treats the human body as a simple multi-link model, assuming that the center of gravity of the links is located at the midpoint, and performs attitude estimation through a limited number of sensors. In reality, the different limb mass distributions of different users will cause the center of gravity of the links to shift, affecting the accuracy of the judgment of "safe movement tolerance".

[0008] The existing technology relies on a flat and stable indoor environment, and its motion control only intervenes when it detects that the human posture exceeds the safety tolerance. It lacks the ability to perceive and adapt to complex outdoor road environments. When the equipment is used in non-standard environments, the dynamic changes in the coefficient of friction between the wheels and the ground may cause the drive wheels to slip, the equipment to shift laterally, or the support to become unstable, thereby increasing the risk of falls. The existing system cannot predict and respond to such risks caused by environmental factors, which poses obvious safety hazards and limits the applicable scenarios and all-weather reliability of the equipment. Summary of the Invention

[0009] In view of the above-mentioned shortcomings in the prior art, the present invention provides a robot motion control method based on state perception and adaptive adjustment, which solves the problems of insufficient adaptability, limited prediction accuracy, and the risk of slipping and falling on complex road surfaces in the prior art.

[0010] To achieve the aforementioned objectives, the present invention employs the following technical solution: a robot motion control method based on state perception and adaptive adjustment, comprising: S1: Based on the user individual feature data obtained in a static environment, perform static parameter calibration and initial posture calculation to obtain the initial parameter set; S2: Based on the multi-source sensor information collected in real time during system operation, obtain a multi-source sensor dataset; S3: Based on the wheel running status data in the multi-source sensor dataset, the equivalent input disturbance of each wheel is calculated using the equivalent input disturbance observer to obtain the road friction coefficient estimate of each wheel. Based on the road friction coefficient estimate, the current road perception state of the device is obtained. S4: Based on human physiological feature data in a multi-source sensor dataset, use feature extraction algorithms and neural network models to perform pattern recognition and obtain the current user's muscle strength perception state. S5: Based on the device interaction mechanical data in the multi-source sensor dataset and the initial center of gravity position in the initial parameter set, calculate the dynamic center of gravity compensation vector and correct it to obtain the real-time dynamic center of gravity perception state of the current user. S6: Based on the road surface perception state, muscle force perception state, real-time dynamic center of gravity perception state, and the initial damping control coefficient and initial position control coefficient in the initial parameter set, a decision model is used to perform safety assessment and adaptive parameter adjustment to generate target motion control commands. S7: Based on the target motion control command, execute motion control on the standing assistive robot, update the initial damping control coefficient and initial position control coefficient according to the execution feedback result of motion control, obtain the motion control result, and complete the motion control of the user.

[0011] The beneficial effects of this invention are as follows: This invention provides a robot motion control method based on state perception and adaptive adjustment. It acquires individual user characteristic data in a static environment for static parameter calibration, and combines this with human physiological characteristic data from a real-time multi-source sensor dataset to identify the current user's muscle strength perception state, achieving matching between control parameters and the user's individual and real-time state, thus reducing parameter adaptation errors. Based on the device interaction mechanics data and initial center of gravity position from the multi-source sensor dataset, a dynamic center of gravity compensation vector is calculated and corrected to obtain the real-time dynamic center of gravity perception state, reducing single center of gravity estimation errors and predicting posture imbalance risks in advance. Based on real-time state assessment, a decision model is used to generate target motion control commands, and the initial damping control coefficient and initial position control coefficient are updated according to the execution feedback results, achieving dynamic closed-loop updating of control parameters according to the user's motion state. An equivalent input disturbance observer is used to calculate the equivalent input disturbance experienced by each wheel, obtaining an estimate of the road friction coefficient for each wheel to determine the road surface perception state, enabling the system to make comprehensive decisions based on the current road surface perception state and generate corresponding motion control commands.

[0012] Further, S1 includes: Based on the acquired user height, weight, limb length, and mass distribution data, biomechanical methods are used to construct user individual characteristic data. The initial center of gravity position is calculated based on user individual characteristic data and initial posture pressure distribution data obtained under static standing posture. Based on individual user characteristic data, the initial damping control coefficient and initial position control coefficient corresponding to the individual user are matched and obtained. The initial center of gravity position, initial damping control coefficient, and initial position control coefficient are integrated to obtain the initial parameter set.

[0013] By collecting real-time information on the robot's speed and position during operation, the pressure information from the rehabilitation trainer on both sides of the robot, and the electromyographic signals from the rehabilitation trainer during training, noise and random errors that may be generated by a single sensor can be filtered out, thereby reducing the uncertainty of the overall information and making the system more confident in its judgment of the rehabilitation ability and posture of the rehabilitation trainer.

[0014] Further, in step S3, the equivalent input disturbance experienced by each wheel is calculated using an equivalent input disturbance observer to obtain an estimated value of the road friction coefficient for each wheel, including: Based on the angular velocity of each wheel and the driving torque of each motor in the wheel running state data, combined with the wheel radius and moment of inertia parameters, an independent dynamic model of each wheel is constructed. Based on the independent dynamics model, the equivalent input disturbance values ​​acting on each wheel are calculated using the equivalent input disturbance observer; Based on the center of gravity offset of the device in the lateral and horizontal directions obtained from the multi-source sensor data, the change in normal load of each wheel caused by the center of gravity offset is calculated. The equivalent input disturbance value is corrected based on the change in normal load to obtain the corrected estimated value of the road surface friction coefficient for each wheel.

[0015] By observing the different road surface friction coefficients as "equivalent inputs," this method can provide an estimate of the current road surface friction coefficient. This allows the vehicle to know not only "the road is slippery" (qualitatively), but also "how slippery" (quantitatively). Based on this observation, the controller can use the estimated friction coefficient as a parameter to dynamically adjust the control strategy. Even if the road surface model is inaccurate, the control system can adapt to various complex road surface changes based on the real-time sensed "equivalent disturbances," significantly improving the vehicle's adaptability on roads with different adhesion coefficients.

[0016] Furthermore, the expression for the corrected estimated coefficient of friction of each wheel on the road surface is as follows: ; in, This represents the corrected estimated coefficient of friction for the i-th wheel. Let i represent the moment of inertia of the i-th wheel; Represents the radius of the i-th wheel; This represents the equivalent input disturbance observer gain for the i-th wheel; This represents the angular velocity observer error of the i-th wheel; This represents the corrected normal load on the i-th wheel; the value of i is 1, 2, 3, or 4. The The value of is obtained by solving the system of four static equilibrium equations simultaneously: ; ; ; ; in, This indicates the body mass of the intelligent standing-assisted walking system; This indicates the additional load mass added by the user during use; Represents gravitational acceleration; Indicates the lateral distance between the left and right wheels; This indicates the front-to-back coordinates of the position of the front wheel; This indicates the front-to-back coordinates of the rear wheel's position. This indicates the horizontal offset of the current user's center of gravity. This indicates the amount of the current user's center of gravity shifting in the forward / backward direction; , , and The normal loads correspond to the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively.

[0017] In rehabilitation training, the robot can dynamically adjust the magnitude of the assistive force by sensing the patient's muscle strength status in real time (such as whether the patient is actively exerting force, whether compensation has been achieved, or whether the patient is close to exhaustion), thus avoiding secondary injury.

[0018] Further, S4 includes: Based on the electromyographic signals and metabolic features of the main muscle groups of the lower limbs in the multi-source sensor dataset, a matrix factorization algorithm is used to extract muscle synergistic feature vectors; the electromyographic signals and metabolic features of the main muscle groups of the lower limbs are human physiological feature data. The muscle coordination feature vector is input into a neural network model for forward inference to obtain the control mode probability distribution; Based on the probability distribution of the control mode, it is determined whether the current user is in precise control mode or robust control mode, and the muscle strength perception state of the current user is obtained.

[0019] Initial center of gravity position is typically obtained based on a static standing posture or a general model. However, during movement, the center of gravity shifts in real time due to limb movements and changes in load. By introducing interactive mechanical data for dynamic compensation, deviations in initial parameters can be corrected in real time, making the center of gravity perception result closer to the actual physical state. This allows for real-time sensing of changes in the patient's center of gravity during rehabilitation, capturing transient imbalance trends, and preventing falls.

[0020] Further, S5 includes: Based on the device interaction mechanics data in the multi-source sensor dataset, gait feature data including step length, step width and support phase variability are extracted; The gait feature data and metabolic feature data are input into a temporal convolutional network model to calculate a dynamic center of gravity compensation vector that characterizes metabolic adaptation and neuromuscular adjustment. The dynamic center of gravity compensation vector is superimposed and corrected with the initial center of gravity position to obtain the real-time dynamic center of gravity perception state of the current user.

[0021] The decision-making model integrates information from three dimensions: the external environment (road friction coefficient), the body's internal state (muscle fatigue level), and dynamic stability (real-time center of gravity position), enabling a comprehensive assessment of the safety of the current movement. Based on the risk assessment results, the system can proactively adjust control parameters (such as increasing the damping coefficient to enhance movement stability, or adjusting the position control coefficient to limit joint range of motion), intervening early to generate robust movement commands and prevent falls and secondary injuries, rather than simply reacting after an accident occurs. Through muscle force perception and center of gravity status, the system can understand the user's movement intentions (such as acceleration, deceleration, and turning) in real time. Combined with road conditions, the decision-making model can dynamically adjust the weights of damping and position control: reducing damping and enhancing compliance when the user actively exerts force; strengthening position control when a precise trajectory is required, achieving seamless switching of auxiliary force.

[0022] Furthermore, the expression for the real-time dynamic center of gravity sensing state is: ; in, This indicates the current user's real-time dynamic center of gravity perception status, i.e., the real-time center of gravity estimate. This represents the dynamic centroid compensation vector calculated by the temporal convolutional network model. This indicates the initial center of gravity position calculated based on initial posture pressure distribution data and the user's individual biomechanical model.

[0023] As the device learns from the effectiveness of previous assistance and fine-tunes its next movement plan, patients experience an increasingly smooth and appropriate interactive experience. This gradual adaptation significantly reduces the harshness between human and machine. When patients perceive that the device can learn from past interactions and performs more reliably, their trust in the device increases significantly, and they are more willing to actively cooperate in completing the training.

[0024] Further, S6 includes: If the road surface perception state is a normal road surface and the real-time dynamic center of gravity perception state is within the preset safety tolerance range, then based on the muscle force perception state, the decision model calculates the parameter adjustment amount of the initial damping control coefficient and the initial position control coefficient, and generates a dynamic motion control command containing the parameter adjustment amount. If the road surface perception state is a slippery road surface and the estimated road surface friction coefficient of any wheel is lower than a preset safety threshold, the decision model generates a safety intervention response command that triggers the drive wheel self-locking and position retracement mechanism; among them, the dynamic motion control command and the safety intervention response command belong to the target motion control command.

[0025] Furthermore, the expressions for the damping control coefficient and the position control coefficient are as follows: ; in, Indicates the damping control coefficient; Indicates the position control coefficient; This represents the feature vector input into the neural network model; Indicates the weights of the first layer of the network; Indicates the first-layer bias parameters; Indicates the weights of the second layer network; Indicates the second-layer bias parameters; Indicates the activation function; This represents the column vector consisting of the damping control coefficient and the position control coefficient.

[0026] Further, S7 includes: Based on the safety intervention response command, the motor output torque is cut off and a braking torque is applied to lock the drive wheel, thereby obtaining locking information; Based on the locking information, the system continuously records historical safe and stable position data, and dynamically calculates the torque distribution ratio between the hub motor and the push rod motor based on the current body tilt angle and center of gravity lateral offset of the device. Based on the torque distribution ratio, the push rod motor is controlled to perform telescopic compensation and synchronously drive the hub motor to move backward until the equipment returns to the historical safe and stable position point, thus completing the position backtracking control and obtaining the execution feedback result of motion control. The initial damping control coefficient and initial position control coefficient are updated based on the execution feedback results of motion control to obtain the motion control results and complete the motion control of the user. Attached Figure Description

[0027] This specification will be further described by way of exemplary embodiments, which will be described in detail with reference to the accompanying drawings. These embodiments are not limiting; in these embodiments, the same reference numerals denote the same structures, wherein: Figure 1 This is an exemplary flowchart illustrating a robot motion control method based on state perception and adaptive adjustment, according to some embodiments of this specification. Detailed Implementation

[0028] The specific embodiments of the present invention are described below to enable those skilled in the art to understand the present invention. However, it should be understood that the present invention is not limited to the scope of the specific embodiments. For those skilled in the art, various changes are obvious as long as they are within the spirit and scope of the present invention as defined and determined by the appended claims. All inventions utilizing the concept of the present invention are protected.

[0029] Example Figure 1 This is an exemplary flowchart illustrating a robot motion control method based on state perception and adaptive adjustment, according to some embodiments of this specification. Figure 1 As shown, the process includes the following steps. In some embodiments, the process may be executed by a processor.

[0030] S1: Based on the user's individual feature data obtained in a static environment, perform static parameter calibration and initial attitude calculation to obtain the initial parameter set.

[0031] User individual characteristic data refers to a set of data used to characterize a user's physical attributes and biomechanical properties. User individual characteristic data may include a user's height, weight, length of each limb segment, and mass distribution ratio of each limb segment.

[0032] In some embodiments, the processor can guide the user to maintain a standard standing posture in a safe environment and acquire the user's individual characteristic data using a laser rangefinder, a weighing sensor, and an external input device.

[0033] The initial parameter set refers to a set of basic control variables and reference thresholds set by the system before motion control is initiated, based on the user's individual characteristics and initial state. The initial parameter set includes the initial center of gravity position, initial damping control coefficient, and initial position control coefficient.

[0034] In some embodiments, the processor can perform biomechanical modeling on individual user characteristic data and combine it with sensor data under the initial posture to calculate data within the initial parameter set.

[0035] In some embodiments, the processor can construct user individual characteristic data using biomechanical methods based on the acquired user height, weight, limb length, and mass distribution data; calculate the initial center of gravity position based on the user individual characteristic data and the initial posture pressure distribution data acquired in a static standing posture; match and obtain the initial damping control coefficient and initial position control coefficient corresponding to the user individual based on the user individual characteristic data; and integrate the initial center of gravity position, initial damping control coefficient, and initial position control coefficient to obtain the initial parameter set.

[0036] Biomechanical methods refer to the techniques used to analyze human structure and motor function by applying mechanical principles, and to construct mathematical models to describe human dynamics and kinematic characteristics. Biomechanical methods can include human center of gravity calculation models based on multi-link models and mass distribution estimation algorithms based on anthropometry databases.

[0037] In some embodiments, the processor can obtain the specific computational logic of the biomechanical method by calling a standard anthropometric database and a biomechanical modeling program pre-stored in memory.

[0038] Initial posture pressure distribution data refers to the distribution of forces exerted by different parts of the user's body on the supporting surface and auxiliary equipment when the user is in a static, standard standing position. Initial posture pressure distribution data can include the normal pressure values ​​of the user's feet on the ground and the vertical and horizontal moments exerted by the left and right hands on the handrails.

[0039] In some embodiments, the processor may acquire initial attitude pressure distribution data through an array of pressure sensors integrated on the ground or on the device pedal, and force sensors mounted on both sides of the device handrail.

[0040] S2: Based on the multi-source sensor information collected in real time during system operation, obtain a multi-source sensor dataset.

[0041] Multi-source sensor information refers to the collection of raw signals about human body status, equipment status, and environmental status synchronously acquired from multiple different types of sensors during system operation. Multi-source sensor information can include continuous electromyographic voltage signals, point cloud coordinate data, pulse count values, and analog voltage / current quantities.

[0042] In some embodiments, the processor can acquire multi-source sensing information by receiving electrical signals from surface electromyography sensors, laser rangefinders, wheel speed encoders, and force sensors in real time through a multi-channel data acquisition card and a bus interface.

[0043] A multi-source sensor dataset refers to a structured data collection formed after preprocessing multi-source sensor information such as filtering, analog-to-digital conversion, and time synchronization. Multi-source sensor datasets include wheel running status data, human physiological characteristic data, and equipment interaction mechanics data.

[0044] In some embodiments, the processor can obtain a multi-source sensor dataset by performing a preset signal processing algorithm on the multi-source sensor information and storing the processed data in a memory buffer according to a specific data structure.

[0045] S3: Based on the wheel running status data in the multi-source sensor dataset, the equivalent input disturbance of each wheel is calculated using the equivalent input disturbance observer to obtain the road friction coefficient estimate of each wheel. Based on the road friction coefficient estimate, the current road perception state of the device is obtained.

[0046] Wheel operating status data refers to data reflecting the dynamics and kinematics of each wheel of the equipment when it moves on the ground. Wheel operating status data includes the angular velocity of each wheel and the driving torque of each motor.

[0047] In some embodiments, the processor can obtain the angular velocity of each wheel by reading the pulse signal output by the high-precision wheel speed encoder installed on each wheel axle, and obtain the driving torque of each motor by detecting the real-time current of each drive motor and combining it with the motor torque constant.

[0048] An equivalent input disturbance observer is an algorithmic model based on control theory used to estimate unknown external disturbances acting on a system using the system's input and output data without directly measuring the external disturbance. An equivalent input disturbance observer can include state equations, an observer gain matrix, and error feedback calculation logic.

[0049] In some embodiments, the processor can obtain the operating structure of the equivalent input disturbance observer by instantiating the algorithm model in the control program and pre-tuning the observer parameters using a system identification method.

[0050] Equivalent input disturbance refers to a single disturbance variable that represents the equivalent reduction of complex external forces (such as road friction and drag caused by unevenness) acting on the wheel dynamics system to the system input. Equivalent input disturbance can include the reduced value of the frictional drag torque between the wheel and the ground.

[0051] In some embodiments, the processor can obtain the equivalent input disturbance value by comparing the measured value of the wheel angular velocity with the predicted value of the observer and using the error feedback calculation logic of the equivalent input disturbance observer.

[0052] The road surface friction coefficient estimate refers to the adhesion capability between the equipment wheels / tire and the road surface currently in contact, calculated in real time by the system. The road surface friction coefficient estimate can include the quantified values ​​of the dynamic friction coefficient between each individual wheel and the ground.

[0053] In some embodiments, the processor can obtain an estimated value of the road surface friction coefficient by substituting the calculated equivalent input disturbance value, combined with physical parameters such as wheel normal load and wheel radius, into the friction calculation formula.

[0054] In some embodiments, the processor can construct an independent dynamic model for each wheel based on the angular velocity of each wheel and the driving torque of each motor in the wheel operating state data, combined with the wheel radius and moment of inertia parameters; based on the independent dynamic model, the processor can calculate the equivalent input disturbance value acting on each wheel using an equivalent input disturbance observer; based on the center of gravity offset of the device in the lateral and horizontal directions obtained from the multi-source sensor data, the processor can calculate the change in normal load of each wheel caused by the center of gravity offset; and based on the change in normal load, the processor can perform a correction calculation on the equivalent input disturbance value to obtain the corrected road friction coefficient estimate for each wheel.

[0055] An independent dynamics model refers to a mathematical equation established separately for each wheel in a system, describing its torque balance and motion evolution. An independent dynamics model can include parameters such as wheel rotational inertia, bearing viscous friction coefficient, motor driving force, and external resistance.

[0056] In some embodiments, the processor can read the mechanical parameters preset by the system at the factory and construct an independent dynamic model in the software based on the Newton-Euler equations or the Lagrange equations.

[0057] Center of gravity offset refers to the distance difference in a spatial coordinate system between the current user's real-time body center of gravity position and a safe and stable reference point (such as the initial static center of gravity position or the geometric center of the device). Center of gravity offset can include the displacement component of the center of gravity in the horizontal direction of movement (forward and backward offset) and the displacement component in the vertical direction of movement (lateral offset).

[0058] In some embodiments, the processor can obtain the centroid offset by performing a vector subtraction operation between the dynamically calculated centroid coordinates and the reference coordinates.

[0059] The change in normal load on each wheel refers to the increase or decrease in the vertical force exerted by each wheel on the ground compared to the static equilibrium state, caused by changes in the user's body posture, shift in center of gravity, or changes in additional load. The change in normal load on each wheel can include the additional or reduced vertical pressure on the left front wheel, right front wheel, left rear wheel, and right rear wheel.

[0060] In some embodiments, the processor can obtain the change in normal load of each wheel by acquiring the center of gravity offset and solving the static equilibrium equations and moment equilibrium equations constructed based on the device geometry.

[0061] In some embodiments, the expression for the corrected estimated coefficient of friction of each wheel on the road surface is: ; in, This represents the corrected estimated coefficient of friction for the i-th wheel. Let i represent the moment of inertia of the i-th wheel; Represents the radius of the i-th wheel; This represents the equivalent input disturbance observer gain for the i-th wheel; This represents the angular velocity observer error of the i-th wheel; This represents the corrected normal load on the i-th wheel; the value of i is 1, 2, 3, or 4. The The value of is obtained by solving the system of four static equilibrium equations simultaneously: ; ; ; ; in, This indicates the body mass of the intelligent standing-assisted walking system; This indicates the additional load mass added by the user during use; Represents gravitational acceleration; Indicates the lateral distance between the left and right wheels; This indicates the front-to-back coordinates of the position of the front wheel; This indicates the front-to-back coordinates of the rear wheel's position. This indicates the horizontal offset of the current user's center of gravity. This indicates the amount of the current user's center of gravity shifting in the forward / backward direction; , , and The normal loads correspond to the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively.

[0062] Road surface perception status refers to the discrete classification judgment made by the system based on the estimated road surface friction coefficients of each wheel, assessing the safety and adhesion of the current ground environment in which the equipment is located. Road surface perception status can include classification labels such as "normal road surface" and "slippery road surface".

[0063] In some embodiments, the processor can obtain the road surface perception state by comparing the estimated road surface friction coefficient of each wheel with a preset safe adhesion threshold, and obtaining the state label based on the logical combination of the comparison results.

[0064] In some embodiments, the derivation process of the corrected road surface friction coefficient estimates for each wheel is as follows: The following dynamic model is established for wheel 1: ; in, For the moment of inertia of the wheel, The angular velocity of the wheel. For motor driving torque, The bearing friction coefficient, For the wheel radius, The normal force acting on the wheel. The force exerted on wheel 1 by this force in the dynamics of wheel 1. Equivalent to an input disturbance Its definition is: ; An EID observer is defined as: ; in Let be the observer gain, and let be the observer error. The observer error equation is: ; Therefore, the equivalent input disturbance estimate can be derived as follows: The above formula is used to calculate... The estimated value :

[0065] ; The following dynamic model is established for wheel 3: ; The force acting on the dynamics of wheel 3 Equivalent to an input disturbance Its definition is: ; An EID observer is defined as: ; in The observer gain for round 3 is given by [value], and the observer error is given by [value]. The observer error equation is: ; Therefore, the equivalent input disturbance estimate can be derived as follows: The above formula is used to calculate... The estimated value :

[0066] ; Since the wheels on both sides of this four-wheel model are symmetrically distributed, that is, wheels 1 and 2 are of equal size, and wheels 3 and 4 are of equal size, the dynamic model of wheel 2 can be obtained similarly as follows: ; Its EID observer is defined as: ; The force applied to wheel 2 Equivalent to an input disturbance Finally, we got The estimated value : ; The dynamic model of wheel 4 is as follows: ; An EID observer is defined as: ; The force applied to wheel 4 Equivalent to an input disturbance Finally obtained The estimated value : ; Therefore, for any wheel Its dynamic model can be uniformly expressed as: ; The force acting on the wheel Equivalent to input disturbance Defined as: ; wheel Constructing an EID observer: ; in Let be the observer gain, and let be the observer error. The observer error equation is: ; Therefore, the equivalent input disturbance estimate can be derived as follows: , obtained through calculation The estimated value.

[0067] In some embodiments, the process of establishing the static equilibrium equations includes obtaining a smoothed equivalent input through a low-pass filter. However, in reality, the shift in the center of gravity will lead to... Changes occur (when the center of gravity shifts to one side, the weight of that side changes) It will increase), thus leading to The estimated values ​​are distorted; therefore, this method uses sensors to acquire the offset of the center of gravity in the lateral and horizontal directions in real time to correct the EID observer. A coordinate system is defined based on the device: The axis represents the horizontal direction (the direction in which the trolley moves forward). The axis is in the vertical direction. The axis is horizontal (left and right of the cart), and the projection of the human body's center of gravity onto this coordinate system is ( , In low-speed standing assisted motion, dynamic inertial forces can be neglected, resulting in static equilibrium equations.

[0068] For the revised A suitable threshold is set to distinguish between normal and slippery road surfaces. Based on the load and adhesion status of each wheel, a distributed collaborative control strategy is adopted to dynamically adjust the driving torque of each wheel. On normal road surfaces, indicating sufficient adhesion between the equipment's wheel surface and the ground, only the user's posture is considered to be within a safe range. Preset control parameters are adjusted to assist the user and maximize muscle force contribution. On slippery road surfaces, indicating reduced adhesion between the equipment's wheel surface and the ground, there may be a potential risk of slippage. In this case, the position control coefficient is increased, the allowable range of motion is reduced, and the user's freedom of movement is restricted. When slippage occurs, the drive wheel self-locking is triggered, and the equipment will be fixed in position to prevent excessive user movement from causing slippage and reduce the risk of falling.

[0069] S4: Based on human physiological feature data from multi-source sensor datasets, feature extraction algorithms and neural network models are used to perform pattern recognition to obtain the current user's muscle strength perception state.

[0070] Human physiological characteristic data refers to biological signals and derived indicators that reflect a user's neuromuscular activity level, energy consumption, and fatigue level during exercise. Human physiological characteristic data includes electromyographic signals and metabolic characteristics of the main muscle groups of the lower limbs.

[0071] In some embodiments, the processor can receive and acquire human physiological characteristic data via Bluetooth or radio frequency communication protocols through wireless surface electromyography sensor arrays worn on the user's lower limbs, as well as wristband heart rate monitors, breathing masks, and other devices.

[0072] Feature extraction algorithms refer to the data processing flow that separates low-dimensional key information representing a user's core movement patterns and intentions from raw, high-dimensional human physiological feature data. Feature extraction algorithms can include specific algorithms such as bandpass filtering, full-wave rectification, smooth envelope calculation, non-negative matrix factorization (NMF), and principal component analysis (PCA).

[0073] In some embodiments, the processor can execute the above algorithm flow on the input physiological feature data and obtain the processing result by calling a pre-compiled signal processing function library.

[0074] A neural network model is a computational model that simulates the structure and function of biological neural networks. It is used to establish a complex nonlinear mapping relationship between input features and output states. In S4, it specifically refers to a network used for motion pattern classification. Neural network models can include the topology, number of layers, number of nodes, activation function type, and pre-trained weight matrices and bias vectors of multilayer perceptrons (MLPs) and lightweight convolutional neural networks.

[0075] In some embodiments, the processor can obtain a neural network model by loading a model file stored in non-volatile memory and instantiating the network structure in runtime memory.

[0076] Muscle strength perception status refers to the system's comprehensive assessment of the user's current muscle strength output pattern, fatigue level, and active control ability based on the user's physiological characteristics. Muscle strength perception status can be classified into states such as "precise control mode" (indicating sufficient muscle strength and precise control) and "robust control mode" (indicating decreased muscle strength and reliance on compensatory mechanisms).

[0077] In some embodiments, the processor can obtain muscle strength perception state based on the probability distribution of the model output by inputting the extracted muscle coordination feature vector into a trained neural network model.

[0078] In some embodiments, the processor can extract muscle coordination feature vectors based on the electromyographic signals and metabolic feature data of the main muscle groups of the lower limbs in the multi-source sensor dataset using a matrix factorization algorithm; the electromyographic signals and metabolic feature data of the main muscle groups of the lower limbs are human physiological feature data; the muscle coordination feature vectors are input into a neural network model for forward inference to obtain a control mode probability distribution; based on the control mode probability distribution, it is determined whether the current user is in a precise control mode or a robust control mode, and the muscle strength perception state of the current user is obtained.

[0079] Matrix factorization algorithms are mathematical methods that decompose a high-dimensional data matrix into a product of multiple low-dimensional matrices. In this context, they are used to extract potential muscle coordination control signals from multi-channel electromyography (EMG) signals. Matrix factorization algorithms can include non-negative matrix factorization algorithms, whose outputs include a muscle coordination weight matrix and an activation time coefficient matrix.

[0080] In some embodiments, the processor can obtain a matrix factorization algorithm by applying an iterative optimization algorithm to the preprocessed multi-channel electromyography signal envelope matrix.

[0081] Muscle coordination feature vectors are numerical sequences obtained through feature extraction algorithms that quantify the pattern features of coordinated activity of multiple muscles in the central nervous system during specific movements. Muscle coordination feature vectors can include the number of coordinating modules, the full width at half maximum (FWHM) of the co-activation curves, and the weight sparsity index.

[0082] In some embodiments, the processor can obtain the muscle synergy feature vector by performing statistical feature calculations on the weight matrix and activation matrix output by the matrix factorization algorithm.

[0083] The control mode probability distribution refers to the statistical probability that a user is in various preset motion control modes given the current input features, as output by the neural network model. The control mode probability distribution can include a probability value of P1 for the user being in "precise control mode" and a probability value of P2 for the user being in "robust control mode," with P1 + P2 = 1.

[0084] In some embodiments, the processor can obtain the control mode probability distribution by acquiring the output node values ​​of the last layer (such as the Softmax layer) of the neural network model.

[0085] S5: Based on the device interaction mechanics data in the multi-source sensor dataset and the initial center of gravity position in the initial parameter set, calculate the dynamic center of gravity compensation vector and correct it to obtain the current user's real-time dynamic center of gravity perception state.

[0086] Device interaction mechanical data refers to the dynamic mechanical signals generated between the user's body and the contact points of the device during the operation and use of the standing assistive robot. Device interaction mechanical data can include the magnitude and direction of the three-dimensional pressure on both sides of the handrail, as well as the pushing and pulling torque applied by the user.

[0087] In some embodiments, the processor can acquire device interaction mechanical data by reading the data registers of multi-axis force / torque sensors distributed on the device handle and armrest support surface.

[0088] The initial center of gravity position refers to the three-dimensional spatial coordinates of the user's overall body gravity point in the reference coordinate system, calculated during the system initialization phase. The initial center of gravity position can include coordinates (X0, Y0, Z0).

[0089] In some embodiments, the processor can obtain the initial center of gravity position by substituting user individual feature data obtained from static measurements into a multi-link center of gravity calculation model.

[0090] The dynamic center of gravity compensation vector refers to the correction amount made when the actual center of gravity deviates from the calculation results based on a pure kinematic model due to metabolic adaptation, fatigue compensation, and adjustments in neuromuscular control strategies during user movement. The dynamic center of gravity compensation vector can include displacement compensation values ​​(Δx, Δy, Δz) in a three-dimensional coordinate system.

[0091] In some embodiments, the processor can obtain a dynamic centroid compensation vector by inputting gait features, electromyographic features, and metabolic indicators into a pre-trained temporal convolutional network model and then using the output of the network model.

[0092] Real-time dynamic center of gravity perception refers to the spatial position information that is closest to the user's current true center of gravity after compensation and correction during system operation. Real-time dynamic center of gravity perception can include the corrected real-time three-dimensional coordinates (Xt, Yt, Zt).

[0093] In some embodiments, the processor can obtain the real-time dynamic center of gravity sensing state by performing a vector addition operation on the center of gravity coordinates calculated in real time based on the kinematic model and the dynamic center of gravity compensation vector.

[0094] In some embodiments, to address the problems of weak environmental perception and difficulty in adapting to complex road surfaces in existing technologies, a road surface state perception method based on equivalent input disturbance (EID) is proposed. Compared with existing technologies, in multi-wheel models, the shift of the user's center of gravity during movement causes a dynamic distribution of loads on the front and rear wheels and left and right wheels, resulting in the failure of traditional EID observers and their inability to accurately reflect the true grounding state of each wheel. Therefore, this application treats the four-wheel system in the standing assistive robot as four independent drive units affected by dynamic loads, designs an independent EID observer for each wheel, estimates the adhesion coefficient between each wheel and the ground in real time, and uses a distributed cooperative control algorithm based on the slip state of each wheel. When a decrease in the adhesion of a single wheel is detected, the driving force distribution of each wheel is dynamically adjusted, the user's range of motion is limited, and a wheel hub locking and safety retreat mechanism is triggered when necessary. This significantly improves the overall adaptability, motion stability, and safety of the device under complex road surface conditions such as wet, slippery, and uneven surfaces.

[0095] Let the four wheels be wheel 1, wheel 2, wheel 3, and wheel 4 (corresponding to front left, front right, rear left, and rear right), with different radii for the front and rear wheels, and symmetrical on both sides. If we only consider the longitudinal motion of the vehicle and neglect lateral forces, then the dynamic equations of the car body are: ; in, This refers to the wheel load (the force is measured by a sensor). The speed of the vehicle when it is moving. For time The coefficient of friction between the wheel and the surface. This is the normal force between the wheel and the ground.

[0096] In some embodiments, the processor can extract gait feature data including stride length, stride width, and support phase variability based on the device interaction mechanics data in the multi-source sensor dataset; input the gait feature data and metabolic feature data into a temporal convolutional network model to calculate a dynamic center of gravity compensation vector characterizing metabolic adaptation and neuromuscular adjustment; and superimpose and correct the dynamic center of gravity compensation vector with the initial center of gravity position to obtain the real-time dynamic center of gravity perception state of the current user.

[0097] Gait characteristic data refers to quantitative parameters that describe the periodicity, spatial and temporal patterns of lower limb movements during walking or movement. Gait characteristic data may include a user's stride length, stride width, cadence, stance phase duty cycle, and stance phase time variation coefficient.

[0098] In some embodiments, the processor can obtain gait feature data by analyzing the changes in the lower limb contour point cloud sequence obtained by the laser rangefinder, or by combining wheel speed encoder data with the center of gravity trajectory fluctuation law.

[0099] Metabolic characteristic data refers to physiological parameters that reflect a user's energy metabolism level and cardiovascular system load during exercise. Metabolic characteristic data may include heart rate (HR) values, estimated oxygen uptake (VO2), and perceived fatigue level (RPE) grades based on questionnaires or heart rate variability.

[0100] In some embodiments, the processor can obtain metabolic characteristic data by parsing data packets sent by a wearable heart rate wristband or chest strap and using relevant physiological estimation models.

[0101] Temporal convolutional network models are deep learning network structures specifically designed to process data with time-series dependencies. They are used to capture the dynamic changes of multimodal sensor data over time to predict centroid shifts. Temporal convolutional network models can include one-dimensional convolutional layers, dilated causal convolutional structures, residual connections, and model parameter weights.

[0102] In some embodiments, the processor can obtain a temporal convolutional network model by reading a network configuration file pre-trained using historical dynamic data from memory and constructing a network computation graph.

[0103] In some embodiments, to address the problem that existing center of gravity models are simplified and do not consider the influence of metabolic adaptation, a method is proposed that integrates laser rangefinder data, IMU inertial data, handrail force sensor data, surface electromyography signals, ground reaction force data, and metabolic indicators, and further optimizes the center of gravity estimation method by incorporating a compensation strategy. The system inputs metabolic features, electromyographic coordination features, and gait features (step length, step width, support phase variability, etc.) extracted from the sensors into a lightweight temporal convolutional network (TCN). The network outputs a dynamic center of gravity compensation vector, which represents the real-time center of gravity shift caused by metabolic adaptation and neuromuscular adjustment.

[0104] In some embodiments, the expression for the real-time dynamic center of gravity sensing state is: ; in, This indicates the current user's real-time dynamic center of gravity perception status, i.e., the real-time center of gravity estimate. This represents the dynamic centroid compensation vector calculated by the temporal convolutional network model. This indicates the initial center of gravity position calculated based on initial posture pressure distribution data and the user's individual biomechanical model.

[0105] S6: Based on the road surface perception state, muscle force perception state, real-time dynamic center of gravity perception state, and the initial damping control coefficient and initial position control coefficient in the initial parameter set, a decision model is used to perform safety assessment and adaptive parameter adjustment to generate target motion control commands.

[0106] The initial damping control coefficient refers to the baseline value of the control parameter set during the system initialization phase based on the individual characteristics of the user, used to limit the speed of the equipment and provide speed resistance. The initial damping control coefficient may include a specific proportional coefficient value B0.

[0107] In some embodiments, the processor can obtain the initial damping control coefficient by querying a pre-set user individual characteristic parameter library and matching the user's weight with the assessed initial muscle strength level.

[0108] The initial position control factor refers to the baseline value of the control parameters set by the system during the initialization phase, based on the individual characteristics of the user, to guide the equipment to maintain a preset safe reference trajectory or position. The initial position control factor may include a specific stiffness or proportional coefficient value K0.

[0109] In some embodiments, the processor can obtain the initial position control coefficient by querying a pre-set user individual characteristic parameter library and matching it with the user's limb length and balance ability indicators.

[0110] A decision model refers to a set of logical rules or algorithmic structure embedded in a system that comprehensively judges the current motion risk based on multi-dimensional state perception results and outputs the optimal control strategy. A decision model may include a series of "IF-THEN" conditional statements, fuzzy control rule tables, or feedforward neural network computation logic.

[0111] In some embodiments, the processor can execute decision control code stored in the program memory to input the perceived state data as an input variable into the decision model to obtain a judgment result.

[0112] Target motion control commands refer to the set of control commands output by the decision model that directly guide the actions of the underlying actuators of a standing assistive robot. Target motion control commands can include dynamic motion control commands and safety intervention response commands.

[0113] In some embodiments, the processor can generate and format the target motion control command by the output of the logical decision branch of the decision model into a data frame format recognizable by the motor driver.

[0114] In some embodiments, if the road surface perception state is a normal road surface and the real-time dynamic center of gravity perception state is within a preset safety tolerance range, then based on the muscle force perception state, the decision model calculates the parameter adjustment amount of the initial damping control coefficient and the initial position control coefficient, and generates a dynamic motion control command containing the parameter adjustment amount; if the road surface perception state is a slippery road surface and the estimated road surface friction coefficient of any wheel is lower than a preset safety threshold satisfying any one of the conditions, then the decision model generates a safety intervention response command that triggers the drive wheel self-locking and position retracement mechanism; wherein, the dynamic motion control command and the safety intervention response command belong to the target motion control command.

[0115] The safety tolerance range refers to a virtual boundary set in three-dimensional space. When the user's center of gravity or posture is within this boundary, the system determines that the risk of falling is at an acceptable level. The safety tolerance range can include a three-dimensional spatial envelope centered on a preset reference trajectory, or a planar region defined by the maximum lateral offset threshold and the maximum forward and backward offset threshold of the center of gravity.

[0116] In some embodiments, the processor can obtain the safety tolerance range by generating the boundary parameter through system factory settings or by dynamically calculating based on user initial calibration data.

[0117] Parameter adjustment refers to the specific value by which the decision model increases or decreases the original control parameters based on the user's current muscle strength perception state. Parameter adjustment can include the increment ΔB of the damping control coefficient and the increment ΔK of the position control coefficient.

[0118] In some embodiments, the processor can input muscle strength sensing feature vectors into a neural network model, output new target parameters, and calculate the parameter adjustment amount by subtracting the new parameters from the current parameters.

[0119] Dynamic motion control commands are control commands issued by the system to adjust the assistance level to the user when the system determines that the current motion is within a safe range. Dynamic motion control commands may include data packets containing newly calculated damping control coefficients and position control coefficients.

[0120] In some embodiments, the processor can generate a control message and send it to the motor controller to obtain dynamic motion control instructions by appending the parameter adjustment amount to the current control parameters.

[0121] In some embodiments, the processor performs static parameter calibration in a safe environment to measure user parameters and obtain precise limb length and mass distribution, which are used to construct an individual-based biomechanical model. Based on this model, initial control parameters are adjusted. During exercise, surface electromyography (EMG) sensors are deployed on the main muscle groups of the user's lower limbs (rectus femoris, vastus lateralis, vastus medialis, gastrocnemius, etc.) to acquire EMG signals. Simultaneously, a wristband-type heart rate monitor is worn to acquire metabolic characteristics such as heart rate (HR), oxygen uptake (VO2), and perceived fatigue level (RPE). Several key muscles are extracted from the EMG signals using a non-negative matrix factorization algorithm. Collaborative features (number of collaborative modules, FWHM, weight sparsity, etc.) are input into a lightweight neural network to determine whether the user is currently in "precise control mode" or "robust control mode," and dynamically adjust accordingly: when increased metabolic load, muscle synergy tending towards robustness (widening of FWHM) and increased gait variability are detected, the assist force is automatically increased, i.e., the position control coefficient is increased, the damping control coefficient is decreased, and the range of motion is reduced, prioritizing stability; when the user's muscle output is enhanced, metabolism is stable, and the collaborative mode tends to be precise, the range of motion is gradually widened, the assist intensity is reduced, and active muscle force participation is encouraged.

[0122] In some embodiments, the expressions for the damping control coefficient and the position control coefficient are: ; in, Indicates the damping control coefficient; Indicates the position control coefficient; This represents the feature vector input into the neural network model; Indicates the weights of the first layer of the network; Indicates the first-layer bias parameters; Indicates the weights of the second layer network; Indicates the second-layer bias parameters; Indicates the activation function; This represents the column vector consisting of the damping control coefficient and the position control coefficient.

[0123] S7: Based on the target motion control command, execute motion control on the standing assistive robot, update the initial damping control coefficient and initial position control coefficient according to the execution feedback result of motion control, obtain the motion control result, and complete the motion control of the user.

[0124] A standing-assisted robot refers to the physical hardware platform used in the method of this invention, which supports and assists users in standing and walking, and performs specific control actions. A standing-assisted robot may include a frame, drive wheels, casters, hub motors, auxiliary support robotic arms, push rod motors, a control board, and various sensors.

[0125] In some embodiments, the device exists as a physical entity as a whole, and the processor obtains its hardware status and sends instructions to it through the physical bus interface.

[0126] Execution feedback results refer to the action completion status and actual motion parameters returned by the underlying actuator to the system's main control unit after receiving control commands and executing actions. Execution feedback results may include the actual speed reached by the motor, the actual output torque, the actual extension and retraction displacement of the robotic arm, and the brake lock-up confirmation signal.

[0127] In some embodiments, the processor can obtain execution feedback results by reading real-time data registers returned by the motor driver, encoder, and position sensor.

[0128] Motion control results refer to the new combined motion state achieved by the device and the user, as well as the updated system parameters, at the end of a control cycle. Motion control results may include the device's new coordinates in space, the user's new posture, and the updated damping control coefficients and position control coefficients written to memory.

[0129] In some embodiments, the processor can obtain motion control results by aggregating the latest frame data from each sensor and the current values ​​of the parameter variables.

[0130] In some embodiments, the processor can, based on the safety intervention response command, cut off the motor output torque and apply braking torque to lock the drive wheel, thereby obtaining locking information; based on the locking information, acquire historical safe and stable position point data continuously recorded by the system, and dynamically calculate the torque distribution ratio between the hub motor and the push rod motor based on the current torso tilt angle and lateral offset of the center of gravity of the device; based on the torque distribution ratio, control the push rod motor to perform telescopic compensation and synchronously drive the hub motor to move backward until the device returns to the historical safe and stable position point, completing the position backtracking control and obtaining the motion control execution feedback result; update the initial damping control coefficient and the initial position control coefficient according to the motion control execution feedback result to obtain the motion control result and complete the motion control for the user.

[0131] Locking information refers to the status indication data used by the system to confirm that the actuator has successfully braked and limited wheel rotation after the drive wheel self-locking mechanism is triggered. Locking information may include the electromagnetic brake closure confirmation position and the drive wheel zero-speed holding status flag position.

[0132] In some embodiments, the processor can obtain locking information by monitoring the feedback level signal of the brake control loop and reading the wheel speed encoder output as zero for multiple consecutive sampling cycles.

[0133] Historical safe and stable position data refers to the sequence of position coordinates and attitude information continuously recorded by the system during operation, before any slippage or instability alarms occur, when the equipment is in a safe state. Historical safe and stable position data may include the equipment's (X,Y) coordinates in the ground coordinate system, as well as the corresponding heading angle.

[0134] In some embodiments, the processor can obtain historical safe and stable location point data by writing the current odometry data and inertial navigation data into a ring buffer storage each time a safe state is determined.

[0135] Torque distribution ratio refers to the ratio calculated by the system to control the output force of different actuators (such as hub motors and push rod motors) when performing complex actions such as position retracement. Torque distribution ratio can include the ratio of the output torque of the hub motor to the output torque of the push rod motor, or the differential ratio of the output torque of the left and right push rod motors.

[0136] In some embodiments, the processor can calculate the torque distribution ratio by reading the current torso tilt angle and the lateral offset of the center of gravity, and substituting them into a preset multi-axis cooperative control allocation matrix.

[0137] If any wheel slips, the system will immediately trigger the following safety response mechanism: 1) Drive wheel self-locking: The system immediately cuts off the motor output torque and applies a reverse braking torque, quickly locking the drive wheel via an electromagnetic brake. It also actively applies braking torque to the support wheel to prevent free slippage. If only a single wheel slips, braking is applied only to the other side to prevent the equipment from rotating. During braking, a wheel speed encoder verifies in real time whether each wheel has entered a zero-speed locking state, ensuring that all wheels effectively stop on wet surfaces and do not slip after locking. 2) Buzzer warning: The device will sound an alarm through the built-in buzzer to remind the user that the current state is unstable (slipping) and that the user should avoid changing posture drastically and stop moving forward.

[0138] 3) Position Retrospection: During each control cycle, the system continuously records information such as the device's position in the ground coordinate system and the user's posture. When the slippage alarm triggers self-locking, after the user's posture has been adjusted and stabilized, the system reads the most recent stable position point before slippage and simultaneously activates the auxiliary support robotic arms on both sides of the handrail. Using a low-speed, high-rigidity position control mode, the system smoothly pushes the user and the entire device back to the previous safe position where slippage did not occur, and maintains stable support in that position until the user readjusts their posture or changes their direction of movement.

[0139] A coordinated control strategy between the wheel hub motor and the armrest push rod motor is implemented to ensure that the user maintains torso stability and avoids forward leaning or lateral imbalance during the reversing process. The specific control logic is as follows: During the reversing initiation, the system dynamically distributes the output torque of the hub motor and the push rod motor based on real-time estimated torso tilt angle and lateral center of gravity offset. If the user's body is detected to be leaning forward, the upward auxiliary torque of the push rod motor is increased while limiting the backward movement speed of the hub. If lateral tilt is detected, active anti-rollover control is achieved through the differential extension and retraction of the left and right push rods. During the reversing process, the hub motor drives the wheels to move backward in a low-speed, high-torque mode, while the armrest push rod motor extends and retracts to compensate for changes in the vehicle's posture: when the wheels move backward, causing the user's center of gravity to lean backward, the push rod extends forward appropriately to maintain the relative stability of the torso and the support surface; conversely, it retracts backward to prevent the user from leaning forward excessively. The system continuously monitors the torso posture and center of gravity projection position throughout the reversing process. If the detected posture deviation exceeds the safety threshold, the backward movement of the hub is paused, and the push rod adjustment is prioritized to bring the torso back to a stable state before continuing the reversing process. Once the device and the user have retreated to the target safe position, the hub motor enters a zero-speed lock state, while the push rod motor applies a continuous but flexible supporting torque based on the current posture, allowing the user to readjust their posture or choose a subsequent direction of action on a stable basis.

[0140] This mechanism enables a closed-loop safety intervention throughout the entire process, from risk detection, active braking, warning reminders to safe retreat, in the event of slippery road surfaces or sudden skidding, thereby improving the reliability of the equipment and the safety of users in complex outdoor environments.

Claims

1. A robot motion control method based on state perception and adaptive adjustment, characterized in that, include: S1: Based on the user individual feature data obtained in a static environment, perform static parameter calibration and initial posture calculation to obtain the initial parameter set; S2: Based on the multi-source sensor information collected in real time during system operation, obtain a multi-source sensor dataset; S3: Based on the wheel running status data in the multi-source sensor dataset, the equivalent input disturbance of each wheel is calculated using the equivalent input disturbance observer to obtain the road friction coefficient estimate of each wheel. Based on the road friction coefficient estimate, the current road perception state of the device is obtained. S4: Based on human physiological feature data in a multi-source sensor dataset, use feature extraction algorithms and neural network models to perform pattern recognition and obtain the current user's muscle strength perception state. S5: Based on the device interaction mechanical data in the multi-source sensor dataset and the initial center of gravity position in the initial parameter set, calculate the dynamic center of gravity compensation vector and correct it to obtain the real-time dynamic center of gravity perception state of the current user. S6: Based on the road surface perception state, muscle force perception state, real-time dynamic center of gravity perception state, and the initial damping control coefficient and initial position control coefficient in the initial parameter set, a decision model is used to perform safety assessment and adaptive parameter adjustment to generate target motion control commands. S7: Based on the target motion control command, execute motion control on the standing assistive robot, update the initial damping control coefficient and initial position control coefficient according to the execution feedback result of motion control, obtain the motion control result, and complete the motion control of the user.

2. The robot motion control method based on state perception and adaptive adjustment according to claim 1, characterized in that, S1 includes: Based on the acquired user height, weight, limb length, and mass distribution data, biomechanical methods are used to construct user individual characteristic data. The initial center of gravity position is calculated based on user individual characteristic data and initial posture pressure distribution data obtained under static standing posture. Based on individual user characteristic data, the initial damping control coefficient and initial position control coefficient corresponding to the individual user are matched and obtained. The initial center of gravity position, initial damping control coefficient, and initial position control coefficient are integrated to obtain the initial parameter set.

3. The robot motion control method based on state perception and adaptive adjustment according to claim 1, characterized in that, In step S3, the equivalent input disturbance experienced by each wheel is calculated using an equivalent input disturbance observer to obtain an estimated value of the road friction coefficient for each wheel, including: Based on the angular velocity of each wheel and the driving torque of each motor in the wheel running state data, combined with the wheel radius and moment of inertia parameters, an independent dynamic model of each wheel is constructed. Based on the independent dynamics model, the equivalent input disturbance values ​​acting on each wheel are calculated using the equivalent input disturbance observer; Based on the center of gravity offset of the device in the lateral and horizontal directions obtained from the multi-source sensor data, the change in normal load of each wheel caused by the center of gravity offset is calculated. The equivalent input disturbance value is corrected based on the change in normal load to obtain the corrected estimated value of the road surface friction coefficient for each wheel.

4. The robot motion control method based on state perception and adaptive adjustment according to claim 3, characterized in that, The expression for the corrected estimated coefficient of friction of each wheel on the road surface is as follows: ； in, This represents the corrected estimated coefficient of friction for the i-th wheel. Let i represent the moment of inertia of the i-th wheel; Represents the radius of the i-th wheel; This represents the equivalent input disturbance observer gain for the i-th wheel; This represents the angular velocity observer error of the i-th wheel; This represents the corrected normal load on the i-th wheel; the value of i is 1, 2, 3, or 4. The The value of is obtained by solving the system of four static equilibrium equations simultaneously: ； ； ； ； in, This indicates the body mass of the intelligent standing-assisted walking system; This indicates the additional load mass added by the user during use; Represents gravitational acceleration; Indicates the lateral distance between the left and right wheels; This indicates the front-to-back coordinates of the position of the front wheel; This indicates the front-to-back coordinates of the rear wheel's position. This indicates the horizontal offset of the current user's center of gravity. This indicates the amount of the current user's center of gravity shifting in the forward / backward direction; , , and The normal loads correspond to the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively.

5. The robot motion control method based on state perception and adaptive adjustment according to claim 1, characterized in that, S4 includes: Based on the electromyographic signals and metabolic features of the main muscle groups of the lower limbs in the multi-source sensor dataset, a matrix factorization algorithm is used to extract muscle synergistic feature vectors; the electromyographic signals and metabolic features of the main muscle groups of the lower limbs are human physiological feature data. The muscle coordination feature vector is input into a neural network model for forward inference to obtain the control mode probability distribution; Based on the probability distribution of the control mode, it is determined whether the current user is in precise control mode or robust control mode, and the muscle strength perception state of the current user is obtained.

6. The robot motion control method based on state perception and adaptive adjustment according to claim 1, characterized in that, S5 includes: Based on the device interaction mechanics data in the multi-source sensor dataset, gait feature data including step length, step width and support phase variability are extracted; The gait feature data and metabolic feature data are input into a temporal convolutional network model to calculate a dynamic center of gravity compensation vector that characterizes metabolic adaptation and neuromuscular adjustment. The dynamic center of gravity compensation vector is superimposed and corrected with the initial center of gravity position to obtain the real-time dynamic center of gravity perception state of the current user.

7. The robot motion control method based on state perception and adaptive adjustment according to claim 6, characterized in that, The expression for the real-time dynamic center of gravity sensing state is: ； in, This indicates the current user's real-time dynamic center of gravity perception status, i.e., the real-time center of gravity estimate. This represents the dynamic centroid compensation vector calculated by the temporal convolutional network model. This indicates the initial center of gravity position calculated based on initial posture pressure distribution data and the user's individual biomechanical model.

8. The robot motion control method based on state perception and adaptive adjustment according to claim 1, characterized in that, S6 includes: If the road surface perception state is a normal road surface and the real-time dynamic center of gravity perception state is within the preset safety tolerance range, then based on the muscle force perception state, the decision model calculates the parameter adjustment amount of the initial damping control coefficient and the initial position control coefficient, and generates a dynamic motion control command containing the parameter adjustment amount. If the road surface perception state is a slippery road surface and the estimated road surface friction coefficient of any wheel is lower than a preset safety threshold, the decision model generates a safety intervention response command that triggers the drive wheel self-locking and position retracement mechanism; among them, the dynamic motion control command and the safety intervention response command belong to the target motion control command.

9. A robot motion control method based on state perception and adaptive adjustment according to claim 8, characterized in that, The expressions for the damping control coefficient and the position control coefficient are as follows: ； in, Indicates the damping control coefficient; Indicates the position control coefficient; This represents the feature vector input into the neural network model; Indicates the weights of the first layer of the network; Indicates the first-layer bias parameters; Indicates the weights of the second layer network; Indicates the second-layer bias parameters; Indicates the activation function; This represents the column vector consisting of the damping control coefficient and the position control coefficient.

10. The robot motion control method based on state perception and adaptive adjustment according to claim 8, characterized in that, S7 includes: Based on the safety intervention response command, the motor output torque is cut off and a braking torque is applied to lock the drive wheel, thereby obtaining locking information; Based on the locking information, the system continuously records historical safe and stable position data, and dynamically calculates the torque distribution ratio between the hub motor and the push rod motor based on the current body tilt angle and center of gravity lateral offset of the device. Based on the torque distribution ratio, the push rod motor is controlled to perform telescopic compensation and synchronously drive the hub motor to move backward until the equipment returns to the historical safe and stable position point, thus completing the position backtracking control and obtaining the execution feedback result of motion control. The initial damping control coefficient and initial position control coefficient are updated based on the execution feedback results of motion control to obtain the motion control results and complete the motion control of the user.

Images