Low-pressure drive-control intelligent servo control method for walking robot and driving device

Abstract

The invention relates to the technical field of robot servo control, in particular to a low-pressure drive control intelligent servo control method for a walking robot and a driving device. The method comprises the following steps of collecting a user behavior control instruction output by a decision unit, performing behavior semantic analysis and intelligent servo parameter regulation and control, and generating a bottom layer servo control instruction; a robot foot end sensing parameter set is collected, foot end pressure distribution calculation is carried out, terrain interaction state analysis is carried out, and a real-time terrain interaction state map is constructed; robot action behavior prediction is carried out based on a bottom layer servo control instruction and a real-time terrain interaction state map, and a full-joint load distribution sensing model is constructed; and performing minimum output power demand calculation based on a full-joint load distribution perception model, performing dynamic power distribution, and constructing a low-voltage driving control engine. Driving parameters are dynamically adjusted according to environment changes and load trends, and efficient and stable motion control of the walking robot is achieved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robot servo control, and in particular to a low-voltage driving control intelligent servo control method and driving device for a walking robot. BACKGROUND

[0002] The driving control system of a walking robot requires the system to respond quickly and maintain high-precision control performance when performing motion tasks, especially when performing complex gait control and environmental adaptation. Traditional high-voltage electric driving control systems often face problems such as high power consumption, long response time, and complex system, and once a system failure or power shortage occurs, it often leads to interruption or errors in robot motion control, affecting its execution efficiency and reliability. In addition, when a walking robot operates in complex terrain and dynamic environments, the servo system is required to accurately adjust the power output to ensure the stability and safety of the motion, which puts higher requirements on the driving and control system.

[0003] Compared with traditional high-voltage driving control systems, low-voltage driving control technology gradually becomes an important direction in the design of walking robots due to its low power consumption, small size, and light weight. Low-voltage driving control systems not only reduce the energy consumption of walking robots and improve their endurance, but also reduce the weight of the system, thereby improving the flexibility and motion efficiency of the robot in dynamic environments. However, low-voltage driving control technology also faces some technical challenges, especially in high-precision servo control. The current and power limitations of low-voltage power supply make it difficult for traditional servo control methods to meet the requirements of high load, high precision, and high-speed response. Therefore, in order to solve the above problems, a more intelligent servo control method is needed to achieve efficient and stable motion control of the walking robot. SUMMARY

[0004] To solve the above technical problems, the present application provides a low-voltage driving control intelligent servo control method and driving device for a walking robot to solve at least one of the above technical problems.

[0005] To achieve the above purpose, the present application provides a low-voltage driving control intelligent servo control method for a walking robot, comprising the following steps:

[0006] Step S1: Collect the user behavior control instructions output by the decision unit, perform behavior semantic analysis and intelligent servo parameter regulation, and generate bottom-layer servo control instructions;

[0007] Step S2: Collect the robot foot end sensing parameter set, perform foot end pressure distribution calculation, and perform terrain interaction state analysis to construct a real-time terrain interaction state atlas;

[0008] Step S3: Based on the underlying servo control instruction and real-time terrain interaction state map, the robot action behavior is predicted, and the joint load calculation of each action is performed to construct a full joint load distribution perception model;

[0009] Step S4: Based on the full joint load distribution perception model, the lowest output power demand is calculated, and dynamic power distribution is performed to construct a low-voltage drive control engine;

[0010] Step S5: Based on the real-time terrain interaction state map, multi-joint load trend analysis is performed, and adaptive parameter adjustment is performed on the low-voltage drive control engine to construct an adaptive drive control engine;

[0011] Step S6: Based on the adaptive drive control engine, the robot drive control job is executed, and joint failure gait reconstruction and multi-working-condition collaborative drive control are performed to construct an intelligent collaborative control model.

[0012] In the present specification, a drive device is provided for executing the low-voltage drive control intelligent servo control method for a walking robot as described above, comprising:

[0013] A servo control module is used to collect user behavior control instructions output by the decision unit, perform behavior semantic analysis and intelligent servo parameter control, and generate underlying servo control instructions;

[0014] A terrain interaction state module is used to collect robot foot end sensing parameter sets, perform foot end pressure distribution calculation, and perform terrain interaction state analysis to construct a real-time terrain interaction state map;

[0015] A joint load perception module is used to predict robot action behavior based on the underlying servo control instruction and real-time terrain interaction state map, and to perform joint load calculation of each action to construct a full joint load distribution perception model;

[0016] A low-voltage drive control module is used to calculate the lowest output power demand based on the full joint load distribution perception model, and to perform dynamic power distribution to construct a low-voltage drive control engine;

[0017] A load trend analysis module is used to perform multi-joint load trend analysis based on the real-time terrain interaction state map, and to perform adaptive parameter adjustment on the low-voltage drive control engine to construct an adaptive drive control engine;

[0018] An intelligent collaborative control module is used to execute robot drive control job based on the adaptive drive control engine, and to perform joint failure gait reconstruction and multi-working-condition collaborative drive control to construct an intelligent collaborative control model.

[0019] The beneficial effects of the present application are as follows: through semantic analysis of user behavior control instructions, the type and target of actions that the robot needs to perform can be accurately understood, avoiding misoperation. By automatically adjusting the servo parameters (such as speed, torque, acceleration, etc.) combined with behavior characteristics, the action is more in line with the actual demand, improving the execution efficiency and response speed. The high-level instruction is quickly converted into low-level servo control instruction, ensuring the continuity and real-time of the instruction chain, and improving the collaboration and precision of the overall control of the robot. Through the data collected by the foot end pressure sensor, the current foot-ground contact state of the robot can be accurately reflected. By calculating the foot end pressure distribution, different terrain characteristics such as hard and soft, slope, concave and convex can be identified, providing a basis for subsequent action adjustment. A real-time terrain interaction state map is constructed to help the robot understand the complexity of the current environment and improve the adaptability and safety of the robot in complex environments. Combined with the current control instruction and environmental state, the robot action trajectory is predicted in advance to reduce the risk of accidental action. The load of each joint is accurately calculated to timely discover possible overload or imbalance. Through the load distribution model, the load state of each joint of the robot is comprehensively mastered to provide a scientific basis for subsequent power distribution and drive control, prolonging the mechanical life. According to the load model, the minimum power requirement is calculated to avoid power waste and improve energy utilization. The drive force distribution of each joint is adjusted in real time to ensure smooth and efficient action, improving the endurance of the robot. A low-voltage drive engine is constructed to ensure the stable operation of the robot in low-voltage environments and reduce electrical safety risks. By analyzing the load change trend, possible load peaks or abnormal states are predicted in advance. According to the environmental changes and load trend, the drive parameters are dynamically adjusted to improve the control flexibility and robustness. An adaptive drive engine is constructed to enable the robot to better cope with variable terrains and complex working conditions, reducing human intervention. Through joint failure gait reconstruction, the robot can maintain basic walking ability when some joints fail, improving system reliability. The coordination between different joints and drive units is realized to improve the overall motion coordination and stability. An integrated intelligent drive control framework is formed to support complex task execution and diversified motion requirements, enhancing the autonomy and intelligence of the robot. BRIEF DESCRIPTION OF DRAWINGS

[0020] Figure 1 A step flowchart for a low-voltage drive control intelligent servo control method for a walking robot of the present application is shown in the figure.

[0021] Figure 2 A detailed implementation step flowchart for step S1 is shown in the figure.

[0022] Figure 3 A detailed implementation step flowchart for step S2 is shown in the figure.

[0023] Figure 4 A detailed implementation step flowchart for step S3 is shown in the figure. DETAILED DESCRIPTION

[0024] It is to be understood that the specific embodiments described herein are merely illustrative of the present application and do not limit the scope of the application.

[0025] The application provides a low-voltage drive control intelligent servo control method and a driving device for a walking robot.

[0026] Please refer to Figures 1 to 4 The application provides a low-voltage drive control intelligent servo control method for a walking robot, which comprises the following steps:

[0027] Step S1: Collecting user behavior control instructions output by a decision unit, performing behavior semantic analysis and intelligent servo parameter regulation, and generating bottom-layer servo control instructions;

[0028] Step S2: Collecting a set of robot foot end sensing parameters, performing foot end pressure distribution calculation, and performing terrain interaction state analysis to construct a real-time terrain interaction state atlas;

[0029] Step S3: Based on the bottom-layer servo control instructions and the real-time terrain interaction state atlas, performing robot action behavior prediction, and performing individual action joint load calculation to construct a full-joint load distribution perception model;

[0030] Step S4: Based on the full-joint load distribution perception model, performing lowest output power demand calculation, and performing dynamic power distribution to construct a low-voltage drive control engine;

[0031] Step S5: Based on the real-time terrain interaction state atlas, performing multi-joint load trend analysis, and performing adaptive parameter adjustment on the low-voltage drive control engine to construct an adaptive drive control engine;

[0032] Step S6: Based on the adaptive drive control engine, performing robot drive control work, and performing joint failure gait reconstruction and multi-working-condition collaborative drive control to construct an intelligent collaborative control model.

[0033] Please refer to Figure 1 The application provides a low-voltage drive control intelligent servo control method for a walking robot, which comprises the following steps:

[0034] Step S1: Collect the user behavior control instructions output by the decision unit, perform behavior semantic analysis and intelligent servo parameter regulation, and generate bottom layer servo control instructions;

[0035] In this embodiment, the control system of the robot needs to communicate with the upper decision unit to obtain user behavior control instructions. The decision unit is usually composed of artificial intelligence modules (such as deep learning, reinforcement learning, etc.) or sensor networks, and outputs high-level behavior instructions according to task requirements. For example, the user may instruct the robot to perform a specific action, such as "start", "turn", "accelerate" or "stop", etc. These control instructions are usually high-level semantic commands such as "forward" or "left turn", but these instructions themselves cannot directly drive the robot to perform actions. In this link, the output of the decision unit is transmitted to the low-level control system through network communication (such as CAN, Ethernet, Wi-Fi, etc.), and the stability and real-time performance of the signal are ensured. For example, the collection frequency of control instructions needs to reach a certain high frequency standard, usually 100Hz~200Hz, to ensure the timely response of the robot when performing dynamic tasks. A decision model based on rules or a deep learning model (such as natural language processing NLP) is used to analyze the instructions. Through pattern recognition, the system converts the user's semantic requirements into specific action parameters, and evaluates the current task requirements (such as the current speed of the robot, terrain conditions, etc.), further optimizing the accuracy and effect of instruction mapping. The user behavior control instructions are converted into motion parameters of the robot, which will directly affect the output of the servo controller of the robot, and then affect the motion of each joint of the robot. The purpose of intelligent servo parameter regulation is to ensure that the robot can reasonably adjust the motion parameters (such as speed, acceleration, attitude, torque, etc.) according to different task requirements. To achieve this regulation, the control system uses some intelligent optimization algorithms, such as fuzzy control, machine learning algorithms, etc., to dynamically adjust the servo parameters to ensure the smoothness, stability and response speed of the motion. For example, when executing the "accelerate" instruction, the system needs to adjust the speed and torque output of the servo motor to cope with the load changes during acceleration; when executing the "turn" instruction, the system may adjust the joint angle, torque distribution and the cooperative motion relationship between joints. Intelligent servo regulation also needs to consider the state feedback of the robot, such as real-time data provided by sensors (including speed, acceleration, torque, etc.), and environmental conditions (such as ground friction, slope, etc.). Through adaptive algorithms, the servo parameter regulation will be dynamically adjusted according to real-time feedback to ensure that the robot can stably complete complex tasks. After the above semantic analysis and intelligent servo parameter regulation, the final low-level servo control instructions of the robot are generated and delivered to the drive control system. The low-level servo control instructions are specific execution signals for controlling each joint of the robot, usually including motor current, voltage, speed, etc. For example, under the "start" instruction, the low-level servo control instructions may include: starting the front leg motor, setting the initial speed of the front leg motor to 10° / s, the speed of the rear leg motor to 5° / s, and the torque limit to 0.5Nm. The purpose of this instruction is to ensure that the robot starts smoothly from a stationary state, avoiding motion incoordination caused by excessive acceleration or deceleration.The process of generating the bottom servo control instruction needs to consider various factors such as the load of each joint, the battery power, the execution accuracy, etc. in real time. The control system will calculate these parameters through a dynamic optimization algorithm to ensure that the robot can complete the task while ensuring the movement effect and avoiding system overload or rapid battery consumption. The generation of the control instruction also needs to be coordinated with other systems of the robot (such as the energy management system, fault diagnosis system, etc.) to ensure the stability of the overall system.

[0036] Step S2: Collect the foot end sensing parameter set of the robot, calculate the foot end pressure distribution, and analyze the terrain interaction state to construct a real-time terrain interaction state atlas;

[0037] In this embodiment, the robot foot end is usually integrated with a variety of high-precision sensors for collecting real-time interaction information with the ground. Common sensing modules include flexible thin film pressure sensors, six-axis force / torque sensors, inertial measurement units (IMUs), capacitive tactile arrays, etc. These sensors are distributed in different areas of the robot foot bottom to capture information such as pressure changes, shear forces, contact point positions, attitude angles, etc. The foot bottom sensor array usually has a point density of 4-9 sensing units per square centimeter, which can provide a two-dimensional pressure distribution map with sufficient resolution. The data sampling frequency is set to 500 Hz or higher to ensure accurate contact data even at high robot gait changes. The system needs to collect, normalize, and filter the sensing data in real time (such as Kalman filtering, median filtering, etc.) to eliminate environmental noise and mechanical vibration interference. By spatially reconstructing the above raw sensing data, the current foot bottom pressure distribution map can be calculated. This map reflects the force state of different areas of the robot foot bottom contacting the ground, and is an important basis for judging the current support stability, landing angle, load distribution, etc. The system maps the data of each sensing unit according to its position in the foot bottom space to form a two-dimensional matrix, where each matrix element represents the pressure of that area. This pressure distribution map can be used to identify the shape and range of the foot end contact area (such as heel landing, full foot landing, or toe contact, etc.), and to analyze the current gait phase of the robot (such as support or swing phase) accordingly. In experimental verification, when the robot stands statically on a horizontal surface, the pressure distribution map is approximately symmetrical, with the center of gravity concentrated in the middle and rear palm area. When dynamically stepping, the pressure moves quickly to the front palm and presents a dynamic trajectory with gait changes. By continuously tracking the pressure change trend, the system can determine whether the landing point deviates from the expected, whether the contact is uniform, and whether there is an abnormal impact or sliding trend. Based on the foot bottom pressure distribution and foot end attitude information, the system further analyzes the terrain interaction state. The goal of this analysis is to identify the physical properties and topological features of the ground, including flatness, slope, hardness, slip risk, obstacle contact, etc. The analysis method combines rule-based modeling and data-driven algorithms: the former calculates the relationship between foot bottom pressure changes and terrain features through mechanical models, and the latter uses trained neural network models to classify and identify typical terrain patterns (such as flat ground, gravel, slope, sand).

[0038] In a specific implementation, the inclination angle or irregularity of the current terrain can be determined by comparing the degree of pressure imbalance in the left and right feet or the front and rear palm areas. For example, when the left rear foot area remains high pressure for a long time and the right front area has almost no contact, the system determines that there is a slope or non-horizontal protrusion in the terrain. When the sensor feedback shear force suddenly rises, it may indicate that the robot is in a sliding state, or the terrain has a low adhesion feature (such as wet ground or sandy ground). The above multi-modal sensing results of the foot end, pressure distribution data, and terrain state recognition results are integrated to construct a real-time terrain interaction state atlas. This atlas is a multi-dimensional space model continuously updated in time dimension. It takes the foot end contact area as the basic unit, records the spatial position, contact state, load data, terrain attribute, etc. of each foot-ground interaction, and continuously expands with the advancement of the walking path. This atlas not only records the current foot end state, but also provides forward prediction and backward tracing capabilities. For example, when the robot is walking forward, it can estimate the contact of the area to be reached. At the same time, the system also retains the foot-ground state data of the past 5-10 steps to generate path safety evaluation or judge whether to re-enter a high-risk area. The real-time terrain interaction atlas is an important output of the environment perception module in the overall system, which can be called by modules such as drive engine, gait planner, load distribution system, etc. to realize the high robustness of the robot in unstructured terrain. Its construction and update rely on high-speed data channels, low-latency data fusion framework (such as ROS2 DDS architecture), multi-thread concurrent processing mechanism, etc. to ensure that the system always responds to environmental changes with millisecond-level response speed. Through this mechanism, the robot can maintain motion stability under complex terrain conditions and provide real-time scene data basis for subsequent adaptive control and behavior strategy optimization.

[0039] Step S3: Based on the underlying servo control instructions and the real-time terrain interaction state atlas, the robot action behavior is predicted, and the joint load calculation is performed for each action, and the full-joint load distribution perception model is constructed;

[0040] In this embodiment, by fusing the bottom layer servo control instructions and real-time terrain interaction state atlas, the impending action behavior is accurately predicted, and the load of each joint participating in the action execution is calculated one by one, thereby constructing a set of high-time-efficiency, high-precision full-joint load distribution perception model. The model provides decision basis for subsequent energy optimization, control parameter dynamic scheduling and fault avoidance mechanism, and is a crucial perception and calculation module in the drive and control loop. The input data of this step includes two parts: one is the bottom layer servo control instruction, which records the target action of each joint in the current control period, such as expected angle change, output torque, voltage and current excitation, etc.; the other is the real-time terrain interaction state atlas from the previous stage, which is based on ground contact force, pressure distribution, friction state, terrain slope and local morphology, etc. It provides the constraint conditions of physical interaction between robot and environment. The combination of the two constitutes the key input for robot action behavior prediction. In the action behavior prediction stage, the system first infers the action trend of the next control period according to the current bottom layer servo instruction. Taking the gait cycle as the time window (for example, 500ms), the system uses time series modeling methods (such as time series regression model, LSTM recurrent neural network or state diagram automaton) to analyze the servo instruction sequence, predicts the motion mode of each limb of the robot in the future gait cycle, and this prediction is not a static interpolation, but a dynamic fusion of the output of the robot state and the environment state, such as the local slope angle and the hardness of the ground in the real-time terrain interaction state atlas. The system will predict whether the robot will have inertia deviation, abnormal center of gravity shift or landing point deviation in the next gait cycle, and perform forward adaptation processing on the control instruction. After completing the behavior prediction, the system enters the joint-by-joint load calculation stage. The core of this stage is to map the expected action to the physical response at the joint level, accurately estimate the mechanical load (such as joint torque, force direction, shear force distribution, etc.) that each joint needs to bear. The calculation process is usually based on dynamic modeling methods, such as Euler-Lagrange equation, spatial rigid body dynamics model or inverse dynamics reasoning based on simulation platform (such as Gazebo or MuJoCo). Considering the dynamic changes of terrain constraints, the load calculation model must introduce the contact force input from the foot, thereby establishing a multi-layer coupling relationship of "terrain-support force-inertia of trunk-joint response". For example, when stepping on a slope, the vertical component of the force borne by the hip joint of the back leg of the robot will be significantly higher than that on a horizontal terrain, and due to inertia deviation, the joint may bear a high lateral torque in the support phase, which must be compensated in advance in joint control. The real-time load, prediction trend and corresponding action state of each joint are dynamically integrated to form a structured full-joint load distribution perception model. The model takes time series as the skeleton and spatial joint structure as the carrier, and fully reflects the force state of each joint of the robot at present and in the short term.It not only supports monitoring of local joint load peaks, but also has certain load trend forecasting capabilities, such as early warning of joint overload risk in the next few cycles. Model output data can be used by subsequent modules for low-voltage current allocation, motor operating point selection, fault prediction redundancy control, and other subsystem calls, thereby building a closed-loop "perception-prediction-regulation" intelligent driving control mechanism.

[0041] Step S4: Perform minimum output power demand calculation based on the full-joint load distribution perception model, and perform dynamic power distribution to build a low-voltage drive control engine;

[0042] In this embodiment, the basis of the minimum output power requirement calculation is the high-precision joint load data. The full-joint load distribution perception model from step S3 has identified the expected force, torque requirements, and motion amplitude of each joint at different time slices. The system periodically calculates the instantaneous power of each joint using the inverse estimation method combined with the angular velocity and torque output of the joint, and then obtains the minimum effective power interval required for each motion phase. In this process, dynamic factors such as non-working loads (such as static maintenance loads), inertial interference loads (such as motor response losses caused by high-frequency oscillations), and additional loads caused by terrain feedback forces need to be considered. To improve estimation accuracy, the system introduces a regression strategy based on the energy consumption mapping model to establish a nonlinear mapping between power output behavior and joint actual posture, load state, and current input. For example, in a continuous climbing experiment on a quadruped robot, the hip joint of the hind leg is the main load joint in most cycles, with a minimum effective power fluctuation range of 3.2-5.8W; while the shoulder joint of the front leg is dominated by inertial traction load, with a power requirement fluctuation of 1.1-3.5W. In flat gait mode, the total minimum power requirement of all joints is about 18.7W, which is more than 14% less power redundancy than the static estimation model. After obtaining the minimum power requirement of each joint, the system enters the dynamic power distribution phase. This phase needs to consider several key constraints: including battery voltage platform, current discharge capacity, local current limit, power supply stability index, heat dissipation condition, etc. The system builds a multi-constraint optimization model (such as linear programming, QP, or reinforcement learning scheduling model) to prioritize power output for key execution joints within the total power budget, and adjust the load for redundant tasks. Especially in low-voltage power supply mode (such as 24V platform or below), the dynamic power distribution mechanism needs to compare the deviation of each joint's current power consumption and minimum requirement in real time, and adjust the voltage, current, and duty cycle accurately, to avoid energy waste and power overload. To maximize drive efficiency, the system also introduces an optimal operating point matching mechanism based on the motor efficiency map, that is, in the power distribution process, not only the minimum output power requirement is met, but also the running state of each joint motor is guided to the efficient interval of its efficiency curve. For example, when the minimum power requirement of a joint is 3.8W, in the corresponding motor efficiency map, its efficiency is highest (about 88%) in the 0.9Nm@18rpm interval, the system will adjust the PWM frequency, current peak value, etc. to make it run stably at this point, both meet the output and control the loss. The dynamic power scheduling results of all joints are integrated into a unified low-voltage drive control engine, which serves as the bottom-layer execution control core, periodically (such as every 10ms) reads power requirement changes, drive state, and feedback parameters, and performs closed-loop adjustment. The engine also includes a burst power scheduling channel for temporary high-power requests in scenarios such as obstacle avoidance, steep slope starting, imbalance transient, etc.Through the scheduling buffer management mechanism and the dynamic duty cycle adjustment algorithm, short-term energy support is given to the key joints under the premise of not breaking through the upper limit of the voltage platform.

[0043] Step S5: Based on the real-time terrain interaction state atlas, the multi-joint load trend is analyzed, and the adaptive parameter adjustment of the low-voltage driving control engine is performed to construct the adaptive driving control engine.

[0044] In this embodiment, the atlas records the current robot foot end contact mode, ground characteristics (such as slope, adhesion coefficient, stiffness, contact area, etc.) and historical contact change trajectory. The system jointly models the atlas and the full-joint load distribution perception model, and constructs a load trend time series for each joint, including torque fluctuation interval, force direction change, load wave frequency, etc. in a unit gait cycle. In the analysis process, sliding window filtering and trend fitting method is introduced, using first-order incremental analysis, nonlinear regression and abnormal fluctuation judgment model based on Bayesian inference to predict and model the load evolution trend of each joint under the current terrain condition. When the robot walks on the gravel and sand transition terrain, the system identifies the load trend of the right front knee joint for 7 consecutive gait cycles, identifies that the torque peak gradually rises (from 0.65 Nm to 0.92 Nm), and the periodic fluctuation amplitude of torque increases to ± 0.13 Nm, and the system judges that there is a terrain soft sinking trend in this area, and marks it as a "high load risk area". Similar trend identification process is executed in the system for all joints simultaneously, and multi-point collaborative modeling is performed according to the spatial distribution of terrain changes to construct a cross-cycle, multi-joint, multi-terrain correlated load dynamic evolution model.

[0045] In the second stage, the system uses the above load trend model as input to adaptively adjust multiple key control parameters of the low-voltage drive control engine. The most critical parameters include: voltage output window of each joint drive motor, maximum output current threshold, PWM duty cycle adjustment frequency, current response time constant, gait state trigger power boundary, etc. The parameter adjustment mechanism is based on dynamic priority sorting and minimum energy consumption strategy: for joints with obvious load rising trend, the system will automatically relax their voltage and current output limits, improve their motor response ability and torque margin, to prevent gait from sliding or energy compensation lagging in the middle of the way; while for joints currently in stable or low load state, their voltage window and response frequency will be compressed to further suppress redundant energy consumption and dynamic jitter risk. The adaptive mechanism is based on real-time control parameter mapping table and neural network optimization model. The system will call the corresponding terrain load parameter template according to the load trend analysis result, and fine-tune it combined with the current control state. For example, in the experiment, the system uses 3 types of typical terrain state templates (high damping type, low adhesion type, slope sliding type) to compare and analyze the joint drive adjustment effect, among which in the low adhesion type terrain, the adaptive engine increases the response voltage of the front leg joint by about 12%, while reducing the response frequency of the rear leg control current peak, the gait stability is improved by nearly 22%, and the energy consumption is reduced by about 11.5%. The results show that the dynamic parameter adjustment mechanism driven by load trend not only improves the motion robustness, but also optimizes the driving efficiency. The system integrates the results of parameter adaptive adjustment and writes them to the drive control bus, updating to the execution layer of the low-voltage drive control engine, so that it has the ability to drive and control based on the evolution of terrain changes and load trends, forming an adaptive drive control engine. The engine has three characteristics of "real-time update, trend guidance, self-adjustment", which can continuously update the control strategy in the control cycle, and make feedback correction at the end of the cycle based on the execution result, finally realizing efficient, stable and energy-saving multi-joint collaborative control of robots in unstructured environments.

[0046] Step S6: Perform robot drive control tasks based on the adaptive drive control engine, and reconstruct gait with joint failure and multi-working-condition collaborative drive control, to build an intelligent collaborative control model.

[0047] In this embodiment, the adaptive drive control engine issues control instructions to complete the precise driving execution of each joint. According to the load trend prediction results and the terrain interaction state atlas in the previous control cycle, the engine dynamically adjusts key parameters such as voltage, current, and duty cycle to achieve optimal control point driving under low voltage conditions. During the control execution process, the system continuously monitors the running state of each joint, including joint angle, actual output torque, input energy consumption, response delay, and heat accumulation, forming a complete running state closed-loop feedback. In the gait cycle, the system can automatically adjust the output strategy of each joint according to different terrain feedback. For example, in soft soil or slippery terrain, the system actively reduces the driving strength to avoid torque overshoot, while extending the stable output time of the support phase joints to improve gait stability performance. The second stage is joint failure detection and gait reconstruction. In actual application, joint motors, drivers, or force sensing components may have different degrees of functional degradation, such as insufficient output torque, abnormal current response, persistent overheating, or feedback loss. The system compares with the running baseline under normal conditions and uses a multi-dimensional anomaly detection mechanism for real-time fault identification. For example, when a joint continuously fails to reach the expected angle for three cycles and the motor current output remains high, the system determines that the joint has a failure trend. Taking the left rear leg knee joint of a certain quadruped robot as an example, in the gravel ground test, the joint had a torque output anomaly (decrease amplitude exceeding 35%) for 4 consecutive gait cycles, and the system triggered the limp gait switching mechanism accordingly. Gait reconstruction is based on the existing gait template library and motion phase compensation model. After failure occurs, the system reconstructs a minimum gait unit based on existing healthy joints. By adjusting the gait rhythm, support phase duration, and swing angle of adjacent joints (such as the opposite front leg and the same side hip joint), the overall center of gravity of the robot is maintained stable and the basic motion function is maintained. At this time, the adaptive drive control engine redistributes power and control parameters according to the new load demand. For example, in the limp mode, the power demand of the rear leg hip joint increases by 23%, and the system automatically increases the upper limit of its voltage output interval and adjusts the PWM frequency to maintain response sensitivity. The third stage is multi-condition collaborative driving control. In unstructured environments, robots often face tasks under complex conditions, such as climbing steep slopes and stepping over obstacles, or jumping short distances in sandy terrain. In such complex scenarios, the system must have the ability to quickly coordinate between multiple joints and multiple target driving tasks. The system constructs a dynamic coordination mapping relationship graph between joints to analyze and synchronize the control parameters between key joints. For example, during continuous high gait stepping motion (step length > 350mm, ground slope > 18°), the system simultaneously schedules the hip-knee linkage control strategy of the double front legs and the diagonal rear legs, uses a pulse power allocation mechanism to quickly complete the center of gravity transition, and uses a pulse current envelope adjustment strategy in the high load section to avoid power peak overshoot.

[0048] In this embodiment, referring to Figure 2 For the detailed implementation steps of step S1, in this embodiment, the detailed implementation steps of step S1 include:

[0049] Collecting user behavior control instructions output by the decision unit;

[0050] Performing power spectral density calculation on the user behavior control instructions, performing real-time scene noise recognition of the robot, and extracting real-time scene noise signals;

[0051] Performing adaptive digital filtering on the real-time scene noise signals, and performing instruction signal waveform dynamic gain processing to construct dynamic gain control instructions;

[0052] Performing instruction decoding and behavior semantic analysis on the dynamic gain control instructions to obtain a plurality of key motion intention factors;

[0053] Performing robot state demand analysis on the plurality of key motion intention factors to generate motion state demand;

[0054] Performing intelligent servo parameter regulation according to the motion state demand to generate bottom servo control instructions.

[0055] In this embodiment, the user behavior control instructions are obtained from the decision unit. These control instructions are calculated based on sensor feedback and environmental perception in the upper control system, and involve various aspects of robot behavior, such as gait, speed, turning, obstacle avoidance, etc. The sensor system inside the robot, such as the inertial measurement unit (IMU), vision system, radar sensor, etc., collects and processes environmental data, combines with the current state of the robot (such as position, attitude, speed, etc.), and generates user behavior control instructions. These instructions will be passed to the lower execution system to indicate what kind of motion behavior the robot should perform. In order to ensure the accuracy and timeliness of the instructions, the control instructions are usually transmitted in real time with high frequency (for example, 50 Hz), and are transmitted to the low-level control unit of the robot through a reliable communication protocol (such as CAN bus or Ethernet). In a noisy environment, the original control instructions may be disturbed by external interference, affecting the accuracy of motion. Power spectral density analysis converts the instruction signal from time domain to frequency domain through Fourier transform or fast Fourier transform (FFT) technology, revealing the frequency components of the signal. Through this method, the effective frequency components and noise components in the control instructions can be distinguished. Noise components may come from environmental noise (such as ground friction, mechanical noise, etc.) or sensor errors, and these noise signals will interfere with the motion execution of the robot. Therefore, the system identifies the noise characteristics in different scenes by comparing the noise model, and further extracts the noise signals for subsequent processing.

[0056] Adaptive digital filtering algorithms are used to denoise real-time scene noise signals. Adaptive filters, such as minimum mean square error (LMS) filters or recursive least squares (RLS) filters, can dynamically adjust filter parameters based on real-time noise feedback, eliminating the effects of noise in real time. These filtering algorithms enable the robot to adjust control commands in real time in a changing environment and suppress the influence of interfering signals. The filtered signal then enters the next step: dynamic gain processing of the command signal waveform. Because noise signals can attenuate the command amplitude, the system requires dynamic gain adjustment to compensate for this effect. Dynamic gain control adjusts the command signal gain based on real-time environmental changes, ensuring that the command signal remains within a reasonable amplitude range and that the robot accurately executes the intended action. The robot's motion intention factors are extracted from the command signal and then converted into specific actions. These factors typically include information such as the desired speed, posture change, step length, and steering angle. Behavioral semantic parsing uses specialized parsing algorithms (such as deep learning-based models) to translate these control commands into specific motion instructions for the robot. For example, when a command includes "walk forward," the decoding process generates corresponding parameters such as gait, step frequency, and leg movement pattern. These parameters are the core data required for the robot to execute "walk forward." By analyzing the current motion intention factor, environmental factors, and the robot's own state (such as battery level, joint angles, and load), the robot determines the specific motion pattern and motion characteristics required. For example, if the command requires the robot to accelerate, the system calculates the required acceleration, step length, and other motion requirements based on the current ground friction, load conditions, and the robot's stability requirements. The robot's current state requirements are compared with the desired motion goals to ensure the optimal motion plan is generated under the given conditions. High-level motion requirements are converted into low-level servo control signals, driving the robot's servo system for precise movement. The intelligent servo control system dynamically adjusts the servo system's operating parameters (such as motor speed, torque, and acceleration) based on sensor feedback to ensure the robot can efficiently and accurately execute motion commands. In this case, the servo system typically uses a closed-loop control algorithm to adjust the motion of each joint in real time to achieve a stable gait or turn.

[0057] In this embodiment, refer to Figure 3 , is a flowchart of the detailed implementation steps of step S2. In this embodiment, the detailed implementation steps of step S2 include:

[0058] The robot foot end sensing parameter set is collected through an integrated multimodal terrain sensing array;

[0059] Identify and eliminate abnormal outlier parameters of the robot foot sensor parameter set to obtain an outlier-eliminated sensor parameter set;

[0060] Outlier rejection of the sensor parameter set for slip detection and torque feedback calculation to obtain the terrain friction coefficient;

[0061] Based on the outlier rejection of the sensor parameter set, the foot end pressure distribution analysis is performed, and the foot end pressure distribution map is extracted;

[0062] According to the foot end pressure distribution map and the terrain friction coefficient, the terrain interaction state analysis is performed, and the real-time terrain interaction state atlas is constructed.

[0063] In this embodiment, by deploying a multi-modal terrain perception array at the foot end of the robot, real-time acquisition and summarization of multi-source sensing information when the foot end contacts the terrain are realized. The perception array usually includes flexible pressure sensors, six-axis force sensors (for measuring three-axis force and three-axis torque), accelerometers, gyroscopes, and high-frequency vibration sensors. Among them, the flexible sensor is used to perceive the unit pressure distribution of the foot end contact surface, the force sensor captures the impact and shear force change at the moment of contact, and the accelerometer and gyroscope jointly monitor the micro-displacement and foot end attitude. In the test system, the data sampling frequency is set to 1 kHz to ensure that the terrain information during dynamic walking can be recorded completely. Through CAN or RS485 high-speed bus, the signals of each module are synchronized and summarized to the central data fusion module, forming the complete foot end sensing parameter set of the robot at each contact moment. The content of the parameter set includes the force in each direction, the pressure matrix, the vibration amplitude, the contact duration, etc., which provides high timeliness of the original data for subsequent outlier detection, friction estimation and contact analysis. Because complex terrain (such as gravel, uneven grassland, soft mud and sand) will cause the data collected by the foot end sensor to have abnormal jump values or outlier data points for a short time, therefore, the original parameter set must be preprocessed to improve the data quality. First, based on the joint discrimination method of IQR (interquartile range) and z-score, the outlier values in the parameter set are identified. The z-score method judges the extreme points by standard deviation multiples (usually set the threshold to ±3), and the IQR method is suitable for processing non-normal distribution of offset values. Second, in the identified outlier data area, sliding window interpolation correction method (such as Savitzky-Golay filter) or linear interpolation method is used for rejection and completion to maintain continuity. In the actual test, for the gravel road test scene, the proportion of abnormal data is about 3.6%, and after rejection processing, the mean value of the sensor data is reduced by about 29%, and the noise signal amplitude is reduced to 0.72 times of the original, thereby improving the stability of subsequent slip detection and friction calculation.

[0064] Based on the real-time change rate of the transverse and longitudinal forces at the foot end, a slip index value (e.g., slip ratio, force vector offset angle) is calculated. When a significant deviation (more than 15°) of the transverse force from the support direction is detected at the moment of contact, accompanied by a momentary drop in pressure, it is determined that a slip event has occurred. Then, in combination with the normal force (Fz) and tangential force (Fx, Fy) applied by the foot end, the friction coefficient of the current contact point is calculated using μ = √(Fx² +Fy²) / Fz. In the simulation and real test fusion test, the friction coefficient obtained on the dry concrete floor is about 0.82±0.04, and on the wet and slippery grassland it is 0.48±0.06. To improve robustness, the friction coefficient estimate is smoothed using a time sliding window (window length 100ms), removing jitter values. The friction parameter generated in this step is the core input for evaluating ground stability and robot gait adjustment strategy. Using the flexible pressure sensing array (usually a 16x16 array) at the foot bottom, the pressure distribution information is mapped into a two-dimensional image, where each pixel point corresponds to the pressure value on a unit area. By using image analysis algorithms (such as edge detection, center projection analysis), key features such as the center of gravity position, the maximum pressure concentration area, and the force range area are extracted. The pressure distribution map shows obvious differences in different terrain contacts: on hard ground, the pressure is centrally symmetrically distributed, and the center of gravity is concentrated, while on soft terrain, the pressure distribution is diffuse, and the edge force points increase significantly. In the experiment, the image entropy index is used to evaluate the quality of each frame of pressure distribution map, and the average value of image entropy is improved from 0.86 (without processing) to 1.32 (after filtering and interpolation processing), and the readability and stability of the image atlas are significantly enhanced. The pressure map is an important visualization basis for terrain contact mode analysis, and provides key graphical input for the final terrain state modeling.

[0065] The above two key indicators, foot end pressure distribution map and terrain friction coefficient, are fused to analyze and model the interaction state between the robot and the terrain in the current gait cycle, forming a terrain interaction state atlas. Using a multi-dimensional state vector (pressure center of gravity coordinates, maximum pressure point, friction coefficient, slip occurrence probability), a contact state vector is constructed, and then combined with the historical multi-step cycle state sequence, an LSTM neural network is used to construct a terrain classification and state reasoning model. This model supports online labeling of common terrains (such as hard, soft, slippery, and granular), and forms a real-time terrain atlas. The atlas is presented in a two-dimensional grid + color gradient visualization, with color representing the terrain stability level (green for high stability, red for low stability). In continuous terrain transformation tests, the terrain state atlas updates at a frequency of 20Hz, with an identification accuracy of over 90%. This atlas not only provides data support for robot foot end motion adjustment, but also provides terrain planning and walking mode switching basis for high-level navigation modules, and is a key data basis for intelligent servo drive control linkage.

[0066] In this embodiment, reference is made to Figure 4The detailed implementation steps of step S3 include:

[0067] Based on the distributed joint monitoring unit, multi-dimensional joint monitoring data of the robot is collected, and the multi-dimensional joint monitoring data includes joint angle, angular velocity, current, and torque.

[0068] The multi-dimensional joint monitoring data is subjected to time sequence joint feature change evolution to generate a time sequence joint feature sequence.

[0069] Based on the bottom layer servo control instruction and the real-time terrain interaction state atlas, robot action behavior prediction is performed to obtain a robot action sequence.

[0070] Based on the time sequence joint feature sequence, action joint load calculation is performed on the robot action sequence to generate action load data of each joint.

[0071] According to the action load data, inter-joint load transfer analysis is performed to construct a full-joint load distribution perception model.

[0072] In this embodiment, the distributed monitoring unit (embedded MCU + multi-channel A / D) installed on each rotating joint of the whole machine acquires four types of original signals, including angle (absolute encoder 16bit resolution), angular velocity (Hall speed measurement or MEMS gyroscope), driving current (low resistance Hall current sensor), and output torque (miniature torque meter based on strain gauge), at a synchronous sampling frequency of 1 kHz. Subsequently, time stamp alignment and frame header checking are performed on the edge end using the CAN-FD bus, and the joint monitoring data frame is uniformly packaged as. Local noise (high-speed pulse, brush interference) is denoised by a 4th order Butterworth digital filter and a wavelet soft threshold joint method. Finally, a "multi-dimensional joint monitoring data" matrix is collected for all joints, laying a time-consistent and amplitude-reliable data foundation for subsequent evolution calculation. After the controller receives the uniformly filtered data matrix, a five-dimensional tensor D(j, t, k) is constructed according to the joint number and time axis. Seven types of statistics, including angle first difference, angular velocity instantaneous gradient, current RMS, and torque peak value, are calculated in the tensor using a sliding window S=50ms. LSTM-AE (32 hidden units) is used for short-term dynamic embedding learning to extract a "health-load-motion" three-domain comprehensive feature vector f_t. The vector flow is stacked by time to form a "time sequence joint feature sequence" F={f_1, f_2, …}. This sequence captures the continuous evolution trajectory of the joint micro-electromechanical state, and can give an abnormal premonition 1.8s in advance of the subtle fluctuations caused by wear and driving hysteresis in a 10min walking experiment.

[0073] The bottom layer servo control instructions (PWM duty cycle, speed given, position expectation) are fused with the real-time terrain interaction state atlas through a double-flow Transformer encoder. The system uses a Masked-Transformer-Predictor to output a "robot action sequence" A={a_1,…,a_6} of N=6 gait phases within a 200ms prediction window, where each a_i contains 12 parameters such as support leg switching time, swing amplitude, and expected foot impact force. The model adjusts the attention weight through the terrain friction confidence to ensure that the swing period is automatically shortened and the joint stiffness instruction is increased when on a slippery surface, thereby ensuring foot stability. For each action segment in A, the system calls the Denavit-Hartenberg-based dynamics inverse module and the joint driver rated curve to convert the action kinematics into the theoretical torque-speed demand of each joint; then it maps point-to-point with the timing joint feature sequence F, calculates the actual-theoretical difference ΔT, Δω on the GPU using fast parallel matrix multiplication, and corrects the rolling friction and backlash compensation items to obtain the "action load data" L(j,t)={T_real,T_theory,ΔT,…}. In the experiment, the calculation time is 4ms under the condition of 1kN·mm peak torque, which can cover the 20Hz servo refresh cycle in real time. L(j,t) is input into a graph convolution network (GCN, 4 layers, 64 hidden units) according to the joint topology and mechanical coupling order. The network edge weight is normalized in advance with the coupling stiffness matrix K_ij. The GCN outputs the "load transfer vector" p_j of each joint, including direct load, coupled load, redundancy, and stress allowance. Finally, the system forms a "full-joint load distribution perception model" with the vector set P={p_hip,p_knee,…}. This model can be visualized as a thermal skeleton diagram in real time, providing joint overload warning, energy optimization, and compliance control basis for the high-level strategy module, and realizing intelligent, safe, and energy-saving operation of the walking robot under the low-voltage direct current driving architecture.

[0074] In the embodiment, step S4 includes the following steps:

[0075] Identifying the operating parameters of the robot driving battery;

[0076] Calculating the real-time voltage output curve, speed, and torque based on the operating parameters of the robot driving battery;

[0077] Calculating the output power according to the speed and torque, and calculating the energy conversion efficiency according to the real-time voltage output curve, to construct a three-dimensional motor efficiency atlas;

[0078] Based on the full joint load distribution perception model, the minimum output power requirement is calculated to obtain the minimum output power requirement value under the operation of each joint;

[0079] Based on the three-dimensional motor efficiency atlas, the minimum output power requirement value is matched with the optimal current-voltage working point, and dynamic power distribution is performed to construct a low-voltage driving control engine.

[0080] In this embodiment, the operating parameters of the battery are comprehensively identified, including real-time voltage (V), current (I), temperature (T), remaining capacity (SOC, State of Charge), total energy capacity (Wh), charging and discharging rate (C-rate), etc. The main purpose of this step is to ensure that the subsequent control model has accurate power supply characteristic input when it is established. To obtain the above parameters, a high-precision power management system (BMS, Battery Management System) integrated voltage / current sampling chip, such as INA226 or MAX17320, can be used in combination with an STM32 or ESP32 main control board to collect data.

[0081] The sampling frequency is generally set to 10-100 Hz to ensure real-time data. Temperature sensing can be achieved by placing a thermistor (such as NTC 10K) on the surface of the battery to monitor the temperature. For the high dynamic load characteristics in the walking condition, a transient response model should be established to identify the voltage drop response curve of the battery under sudden load (such as from 0.5C to 2C in 0.3s). In addition, the relationship between open circuit voltage (OCV) and SOC needs to be recorded to establish a battery state mapping model for subsequent energy prediction and efficiency calculation. The key to this step is sampling accuracy and system timeliness, especially in scenarios with frequent voltage fluctuations, which puts higher requirements on the anti-interference ability of the sampling system. Based on real-time sampling data, the output response characteristics of the motor are calculated. The voltage output curve is constructed based on the above battery voltage and current data, and the continuous voltage output curve is obtained after processing by time series fitting method (such as sliding window average filtering). This curve needs to be matched with the robot walking stage, i.e. time synchronization with the gait cycle (e.g. 0.8s-1.2s cycle), so as to analyze the voltage response in different action stages. The motor speed can be obtained by sampling the encoder (such as AMT102 series), and the sampling resolution is recommended to be no less than 12-bit accuracy to ensure the detection accuracy of small amplitude acceleration. For torque, if the motor supports built-in Hall or current feedback, the torque T = Kt × I can be calculated using the known motor constant (Torque Constant, Kt) and current I. For example, if the drive motor Kt=0.12 Nm / A and the drive current is 1.5A, the output torque is 0.18Nm. The output mechanical power P_out can be calculated by the formula P_out = T ×ω, and the units need to be unified: torque in Nm, speed needs to be converted to rad / s (i.e. rpm × 2π / 60). The motor electrical input power P_in = U×I, with units in W. The energy conversion efficiency η = P_out / P_in is a key indicator to measure the performance of the drive control system. In the experimental design, different weights of external load (such as 0.5kg, 1kg, 1.5kg) can be loaded and tested at different walking speeds (0.3 m / s ~ 1.2 m / s) to construct a complete data space. Then, with speed (ω), torque (T) as the horizontal and vertical axes, and efficiency η as the Z-axis, a three-dimensional efficiency map of the motor is drawn. Each point in the map represents the efficiency level of a stable working interval. For example, the efficiency of a certain motor may be as high as 85% at 120rpm and 0.15Nm, but it may drop to 65% at low speed and high torque (60rpm, 0.3Nm). In the process of multi-joint collaborative walking of the robot, the minimum drive power required by each joint depends on its current load state and action target. Therefore, a load distribution-based perception model needs to be established, mainly through joint sensors (current, stress, angle sensors, etc.) to monitor the joint force and angle changes in real time.For example, joint force is inferred using joint current sampling, combined with IMU (such as MPU9250) to obtain angular velocity and angular acceleration data, and to infer the inertial effect. Through dynamic modeling (Lagrangian or Newton-Euler method), the minimum power output required by each joint of the robot in static and dynamic states can be accurately calculated. For example, the power required by the rear swing leg in the support state can be reduced to 0.5W, while the swing leg may need to output 1.8W during the lifting process. In the experiment, the power variation law under different gaits can be tested in the simulation platform (such as Gazebo or Webots) through specific trajectory planning (such as step length 10cm, cycle 1s), and the minimum power requirement of each motion unit is extracted. On the basis of the constructed motor efficiency map and the minimum power requirement value of the joint, the optimal working point matching of voltage-current needs to be realized. This process realizes energy distribution optimization by searching for the point (I, U) combination with the highest efficiency in the efficiency map under the premise of meeting the target power (P = T x ω). For example, if the current minimum power requirement of the joint is 1.2W, and the working point efficiency corresponding to 120rpm and 0.1Nm in the efficiency map is 87%, the input power of this point is 1.38W, the current is 1.15A, and the voltage is about 1.2V. The control system sets the target to this point and dynamically adjusts the PWM duty cycle and drive ratio. In terms of control strategy, fuzzy control combined with model predictive control (MPC) can be used to predict and adjust the energy consumption of each joint, realizing power coordination and redundancy adjustment among multiple joints. The low-voltage driving control engine also needs to have real-time fault perception function (such as current overload protection) to realize safe and stable operation. The low-voltage control algorithm (such as 48V or below DC servo drive) can be deployed on the experimental platform and integrated with the CAN bus communication protocol, and the control command is updated every 5ms. Through simulation test and real machine walking experiment (gait cycle 1s, walking speed 0.5m / s), the comprehensive optimization effect of dynamic power distribution in overall power consumption, thermal efficiency and response speed is verified.

[0082] In this embodiment, the specific steps of step S5 are:

[0083] Based on the real-time terrain interaction state map, the terrain state fluctuation is analyzed, and the dynamic walking working condition is identified to generate dynamic walking working condition information;

[0084] Based on the dynamic walking working condition information, the multi-joint load trend analysis is performed on the full-joint load distribution perception model, and the front prediction is performed to obtain the multi-joint load situation prediction map;

[0085] A load situation fluctuation modification threshold is defined; the multi-joint load situation prediction map is identified according to the threshold value, and the threshold triggering point is marked;

[0086] According to the threshold trigger point and the multi-joint load situation prediction map, adaptive parameter adjustment is performed on the low-voltage driving control engine, and an adaptive driving control engine is constructed.

[0087] In this embodiment, the core of the terrain fluctuation analysis is to track and model the changes in terrain characteristics in real time. A sliding window algorithm or Kalman filter is usually used to smooth the terrain state data, so as to accurately identify the fluctuations of the terrain. According to these changes, the system can dynamically identify the walking conditions. For example, when the robot walks on uneven ground, the system can determine whether to adjust the gait or adjust the joint load distribution according to the real-time terrain information. Finally, dynamic walking condition information will be generated and passed to the subsequent control module, which contains terrain fluctuation amplitude, expected load distribution, stability evaluation of the current condition, etc., providing necessary data support for subsequent joint load distribution and low-voltage driving control. According to these information, the load distribution of each joint under the current walking condition is analyzed. In order to achieve this goal, the robot full-joint load distribution perception model needs to be combined, which combines the mechanical properties of each joint of the robot (such as moment of inertia, maximum bearing capacity, motion mode, etc.) with the actual load data. Through real-time acquisition of terrain information and robot action mode, the system can calculate the instantaneous load of each joint, and on this basis, trend analysis is carried out. Trend analysis uses machine learning methods (such as time series prediction, long short-term memory network LSTM, etc.) to make pre-prediction on joint load data. By training the historical data of joint load changes, the system can predict the load change trend of each joint in the future period of time. For complex dynamic load, methods such as dynamic time warping (DTW) can effectively capture the time sequence relationship of load pattern. Finally, the system will generate a multi-joint load situation prediction graph, which depicts the load change trend of each joint in the future period of time, and can show the load fluctuation range under different conditions. This atlas provides data support for subsequent dynamic control and power distribution, helping the system to adjust motor output and joint control strategy in real time. The definition of load situation fluctuation modification threshold is to realize the effective monitoring and control of joint load changes. When the robot walks in complex environment, the load of each joint will fluctuate, especially when encountering rugged terrain, the load fluctuation amplitude is large. In order to avoid the performance problems caused by overload or insufficient load, a threshold needs to be set as the change range of load situation fluctuation. The threshold is usually set based on the design limits of each joint of the robot, historical load data, and real-time terrain characteristics. The threshold triggering mechanism monitors the multi-joint load situation prediction graph in real time, and when the predicted load fluctuation amplitude exceeds the set threshold, the system will trigger an alarm or adjust the action. The marking of threshold trigger point is crucial for subsequent control strategy. Once the threshold is triggered, the system will dynamically adjust the joint load through real-time calculation, avoiding the instability or damage of the robot caused by joint overload or unstable load.

[0088] In practical implementation, the setting of the threshold value can adopt an adaptive algorithm to adjust according to the real-time load condition of the robot. For example, a fuzzy control algorithm is used to dynamically adjust the threshold value according to real-time environmental conditions and load conditions, improving the robustness and flexibility of the system. By marking the threshold trigger point of the multi-joint load situation prediction graph in real time, the system can react in advance and adjust the robot's movement mode to ensure that the robot can stably complete the task. At the threshold trigger point and the multi-joint load situation prediction graph, the system performs adaptive parameter adjustment on the low-voltage drive control engine. When the load fluctuation exceeds the preset threshold value, the system quickly adjusts the control parameters, including current, voltage, PWM duty cycle, etc., through a feedback mechanism to ensure that the motor output matches the joint load demand. The low-voltage drive control engine adjusts the current-voltage operating point through real-time calculation and optimization algorithms to avoid excessive discharge of the battery while ensuring that the robot does not lose power performance when the load changes. The core of the adaptive control strategy is to dynamically adjust the drive parameters through real-time feedback, and this process uses adaptive filtering algorithms (such as Kalman filtering, LMS algorithm, etc.) and machine learning methods (such as reinforcement learning) to optimize control effect. The system will determine whether to adjust the motor output power based on the threshold trigger point and load prediction information. The implementation of the dynamic adjustment control strategy will adjust the drive current and voltage in real time according to the load change to ensure that the robot operates with the lowest power consumption and best efficiency. The adaptive control engine can make decisions and flexibly adjust the power output of the motor based on load fluctuations, terrain changes, and the state of the robot itself, ensuring the stability of the robot's movement and maximizing energy utilization efficiency. Through this process, the robot can maintain high-efficiency movement ability and long-lasting endurance in different terrains and different task requirements.

[0089] In this embodiment, the specific steps of step S6 are:

[0090] Performing robot drive control tasks based on the adaptive drive control engine and collecting joint motion posture parameters;

[0091] Analyzing joint motion characteristics based on joint motion posture parameters, extracting runtime sequence logic and phase relationship between joints;

[0092] Performing smooth gait optimization based on runtime sequence logic and phase relationship to obtain smooth gait optimization parameters;

[0093] Performing joint failure analysis and limp gait switching decision based on joint motion posture parameters, and constructing load redistribution fault emergency control mode;

[0094] Performing multi-condition collaborative drive control based on smooth gait optimization parameters and load redistribution fault emergency control mode, and constructing intelligent collaborative control model.

[0095] In this embodiment, during the control process, the angle, angular velocity, angular acceleration, torque, and other motion parameters of each joint are first monitored in real time by embedded sensors and encoders. These data are dynamically changing, especially during different stages of robot walking, such as starting, accelerating, decelerating, turning, etc. In order to accurately grasp the motion state of each joint, sensor data need to be collected in real time at a high sampling frequency (such as 100 Hz or higher). The collected data are first subjected to data filtering and denoising to eliminate sensor errors or environmental interference. Then, the data are transmitted to the central control unit for further analysis and processing. The goal of this process is to ensure that the drive control engine obtains accurate state information of each joint in real time, so as to adjust the control strategy and optimize the stability and efficiency of robot walking. Real-time acquisition of motion posture parameters provides an important basis for subsequent gait optimization, load distribution, and fault detection. Frequency domain analysis methods (such as Fourier transform) or time domain analysis (such as sliding window method) are usually used to analyze the motion data. For example, the gait cycle between the front and rear legs may exhibit a fixed phase difference, while the alternating motion of the left and right legs may be synchronized at certain times. By analyzing these motion characteristics, the synchronicity, alternation, and phase shift between joints can be identified, which provides basic data for subsequent gait optimization and fault detection. In addition, nonlinear characteristics of gait changes and joint motion need to be considered during the motion feature analysis process, so machine learning methods (such as clustering analysis, regression analysis) are often combined to further optimize the extraction of joint motion patterns. Through this process, the system can accurately identify the behavior characteristics of each joint in different motion stages and provide a basis for subsequent gait optimization. Gait optimization ensures smooth and coordinated motion of joints during robot walking. The core goal of smooth gait optimization is to eliminate sudden changes, oscillations, or uncoordinated phenomena between joints, so that the robot has higher stability and comfort when walking on various complex terrains. A gait generation model is built, which considers the phase relationship and motion characteristics between joints. Gait optimization usually uses time-space coordination optimization algorithms, such as dynamic programming methods based on Bellman equation, particle swarm optimization (PSO), or genetic algorithms, to minimize the motion difference and energy loss between robot joints. During the optimization process, the system considers the maximum rotation range, rotation speed, acceleration limit, and other hard constraints of the joint, and integrates these constraints into the optimization objective function.

[0096] The output of the optimization parameters includes the optimal acceleration, rotation speed, gait transition time of each joint, etc. These parameters will help the control system to control the movement of each joint more accurately in actual operation, making the robot's gait more smooth and stable. For example, the system can appropriately slow down the support force of the hind legs when the front legs are lifted, thereby avoiding the shock caused by sudden movement. Finally, the smooth gait optimization parameters can be passed to the drive system through the adaptive control engine for real-time adjustment of the drive signals of each joint of the robot. During the walking process of the robot, some joints may fail due to overload, damage or other reasons, so joint failure analysis is crucial. Through the real-time acquisition of joint posture parameters, the system can detect abnormal behavior of the joint, such as motion range exceeding limit, temperature being too high or torque fluctuation, etc. These abnormal signals can be monitored in real time through threshold detection method or model prediction. Once the joint fails, the system will deal with the failure through the limp gait switching decision. In this process, the control system first identifies the failed joint and analyzes the current motion state and load condition of the robot. Then, the system will select the appropriate emergency control mode according to the fault type, such as adjusting the load of other healthy joints, or adopting double-leg coordinated gait strategy, etc., to maintain the smooth movement of the robot.

[0097] The load redistribution failure emergency control mode is the key to this decision. The system uses real-time load prediction and optimization algorithms, such as the load adjustment method based on multivariate optimization, to dynamically adjust the load distribution of each joint. For example, when the current leg fails, the load of the rear leg can be appropriately increased, and the lateral stability can be achieved by adjusting the motion trajectory of the supporting leg. Through the redistribution of load, the robot can continue to walk under the condition of joint failure, without stopping or losing balance. Walking robots need to complete tasks under various working conditions, including different terrains, speed requirements, load conditions, etc. In order to cope with these changes, the construction of an intelligent collaborative control model is crucial. This model achieves the coordinated work between joints by fusing smooth gait optimization parameters and load redistribution mechanisms. The core of the collaborative control model is to dynamically adjust the motion state of each joint of the robot according to the current working condition. For example, on relatively flat ground, the robot can use a higher speed and lower energy consumption mode, while on steep slopes or uneven ground, the system will automatically adjust the gait optimization parameters and load distribution strategy to ensure that the robot can move steadily. In this process, the intelligent collaborative control system will sense environmental changes in real time and adjust through adaptive control algorithms such as reinforcement learning, fuzzy control, etc. Specifically, the system will adjust control parameters such as current, voltage, gait cycle in real time to ensure that each joint is always in the optimal working state. The ultimate goal of the multi-condition collaborative driving control model is to enable the robot to flexibly adjust according to different tasks and terrain conditions, and always maintain stability and efficiency. Through this control model, the robot not only can perform actions under a single task, but also can handle complex and variable working environments, enhancing the overall task adaptability.

[0098] In this embodiment, the specific steps of the joint failure analysis and limp gait switching decision based on the joint motion posture parameters, and the construction of the load redistribution failure emergency control mode are as follows:

[0099] Perform joint failure analysis on the joint motion posture parameters, and mark the failure joints;

[0100] Identify the location of the failure joint and perform failure diffusion impact identification to mark the failure related joints;

[0101] Perform failure risk assessment based on the failure related joints and failure joints to obtain a failure risk assessment value;

[0102] Make a limp gait switching decision based on the failure risk assessment value, and perform load redistribution of the failure joints to construct a load redistribution failure emergency control mode.

[0103] In this embodiment, in the robot complex field task, the joint functional degradation or even failure due to overload, motor loss, sensor drift and other problems belongs to a high probability event. The system establishes a complete "motion posture abnormality identification mechanism" to realize real-time failure analysis and fault joint marking. The mechanism is based on three key parameters for discrimination: expected angle trajectory (θ_des), actual angle trajectory (θ_act) and joint current (I_act). Under normal circumstances, the error between θ_des and θ_act should be maintained within the threshold δθ_max (such as ±3°), if the time frame of continuous error threshold exceeds N=15 (sampling frequency 100 Hz), the system is initially judged that the joint has "motion execution abnormality".

[0104] Further combine current characteristics to judge fault type: ① If the actual current is much higher than the load prediction value (such as >2.5A, the standard value is only 1.2A) and the output torque does not increase obviously, it is initially judged that the motor mechanical jam or transmission wear; ② If the current is obviously low (such as <0.2A) and the execution trajectory deviates seriously, it is suspected that the electrical circuit is broken or the control signal is interrupted. In addition, the vibration amplitude of the joint is also detected by the redundant IMU (an independent six-axis gyroscope is installed at the upper end of each limb) to assist the judgment.

[0105] For example, in a broken stone slope test, the left rear knee joint has θ_act lagging behind, torque output suddenly dropping and current reaching 2.8A, the system completes abnormality judgment within 180ms and marks it as "fault joint (FailJoint)". This kind of judgment mechanism realizes 96.3% recognition accuracy and less than 0.5% misjudgment rate in 1000km real machine task cumulative test. According to the robot limb kinematic chain, the fault joint belongs to the link (such as the rear leg knee joint belongs to the rear leg chain), and then the influence of the upstream and downstream joints is evaluated according to the dynamic stiffness coefficient (K_couple, unit Nm / rad) and the control correlation coefficient (ρ_control, ∈[0,1]). The system uses empirical modeling and offline simulation data to mark the joints with K_couple>1.5Nm / rad and ρ_control>0.65 as "highly sensitive linked joints". Taking the knee joint as an example, its abnormality will cause the ankle joint to bear additional shear load, and force the hip joint to make additional lifting compensation, so both of them are marked as "fault linked joints (Linked Joint) by the system. The system uses a weighted risk function model to consider the fault position, the number of affected joints, the key degree and the current environmental complexity, and finally outputs a continuous "fault risk assessment value (Fault Risk Score, FRS)", which is in the range of [0,1]. The calculation formula of RS is composed of the following parts:

[0106] F1 (fault severity): determined by the degree of joint out-of-control (e.g. angle error, output current), the more severe the score is higher (e.g. single knee joint out-of-control may be assigned 0.4~0.5);

[0107] F2 (propagation impact factor): depends on the number of related joints and their importance level, for example, the hip joint is high priority, the ankle joint is medium priority, the wider the propagation, the higher the score (e.g. 0.2~0.3);

[0108] F3 (external working condition factor): based on the current working condition label to adjust the weight, increase the risk score when on rough road or uphill (e.g. additional weighting 0.1~0.2);

[0109] F4 (historical state memory): if the same joint has a micro-fault record in the past several gait cycles (e.g. high vibration, high current), the score of this item is increased (typical value 0.05~0.1). For example, on a combination of gravel + steps road surface, the back leg knee joint completely fails, affecting the hip and ankle joints, the system finally calculates FRS=0.81, which exceeds the threshold value 0.75, triggering the "limp control mode" switching logic. This mechanism has successfully avoided the robot from falling down and effectively ensured the continuity of the task in multiple field missions. When the fault risk assessment value exceeds the preset threshold value (e.g. FRS > 0.75), the system will immediately switch to the limp mode and redistribute the load. The limp mode is not simply masking the faulty joint, but rather by assigning it a "minimum excitation" strategy (e.g. maintaining neutral position + weak stiffness output), and transferring its functions to other healthy joints as much as possible, to ensure that the robot still has basic walking ability in an asymmetric state. First, in the trajectory planning layer, a set of "asymmetric gait reconstruction model" is introduced to adjust the support and swing time allocation (e.g. extend the healthy side step length by 15%, and the support period to 65%), while shrinking the action amplitude of the damaged side (e.g. limit the knee joint angle to ±10°). Then, in the control allocation layer, the output upper limit of the healthy joints is adjusted, such as the ankle joint allowing the maximum torque output to increase from 1.5 Nm to 2.1 Nm. If the original drive current is 0.9 A, it is temporarily increased to 1.25 A to compensate for the lack of support force. In the experiment, under the condition of left rear knee joint failure, the system still maintains a slow walking speed of 0.3 m / s after switching to the limp gait, the average power consumption increases by 18.4%, but the stability is significantly improved, and the robot does not fall down in a 5 km task segment. The controller has a thermal protection and current redundancy judgment to ensure that even after load transfer, the emergency scheduling can be completed within the acceptable range of low voltage power supply (e.g. 24 V).

[0110] Finally, the "load redistribution fault emergency control mode" realizes a dynamic switching mechanism from "full symmetric control" to "partial redundant fault tolerance", providing a highly robust running guarantee for low-voltage driving robots under sudden structural failures.

[0111] A driving device for executing the low-voltage driving intelligent servo control method for a walking robot as described above, comprising:

[0112] A servo control module for collecting user behavior control instructions output by the decision unit, performing behavior semantic analysis and intelligent servo parameter control, and generating bottom-layer servo control instructions;

[0113] A terrain interaction state module for collecting robot foot end sensing parameter sets, performing foot end pressure distribution calculation, and performing terrain interaction state analysis to construct a real-time terrain interaction state atlas;

[0114] A joint load sensing module for predicting robot motion behavior based on bottom-layer servo control instructions and real-time terrain interaction state atlas, and performing joint load calculation for each motion to construct a full-joint load distribution sensing model;

[0115] A low-voltage driving control module for performing minimum output power demand calculation based on the full-joint load distribution sensing model, and performing dynamic power distribution to construct a low-voltage driving control engine;

[0116] A load trend analysis module for performing multi-joint load trend analysis based on the real-time terrain interaction state atlas, and performing adaptive parameter adjustment on the low-voltage driving control engine to construct an adaptive driving control engine;

[0117] An intelligent collaborative control module for performing robot driving control tasks based on the adaptive driving control engine, and performing joint failure gait reconstruction and multi-working-condition collaborative driving control to construct an intelligent collaborative control model.

[0118] Therefore, embodiments should be considered in all respects as illustrative and non-restrictive, the scope of the present application being defined by the appended claims and not by the above description, and all changes falling within the meaning and range of equivalence of the essential features of the application are therefore intended to be embraced therein.

[0119] The above description is merely that of specific embodiments of the present application, enabling a person skilled in the art to understand or implement the present application. Various modifications to these embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application is not to be limited to the embodiments shown herein, but is to accord with the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A low-voltage drive intelligent servo control method for a walking robot, characterized in that: The following steps are involved: Step S1: Collect user behavior control instructions output by the decision unit, perform behavior semantic analysis and intelligent servo parameter adjustment, and generate low-level servo control instructions; Step S2: Collect the sensor parameter set of the robot foot end, calculate the foot end pressure distribution, and perform terrain interaction state analysis to construct a real-time terrain interaction state map; Step S3: Based on the underlying servo control instructions and the real-time terrain interaction state map, the robot's motion behavior is predicted, and the load of each joint is calculated to build a full-joint load distribution perception model; Step S4: Calculate the minimum output power requirement based on the full-joint load distribution perception model, perform dynamic power allocation, and build a low-voltage drive control engine; Step S5: performing multi-joint load trend analysis based on the real-time terrain interaction state map, and adaptively adjusting the parameters of the low-voltage drive control engine to construct an adaptive drive control engine; Step S6: Execute the robot drive control operation based on the adaptive drive control engine, perform joint failure gait reconstruction and multi-condition collaborative drive control, and build an intelligent collaborative control model.

2. The low-voltage drive intelligent servo control method for a walking robot according to claim 1, characterized in that: The specific steps of step S1 are: Collect user behavior control instructions output by the decision-making unit; Performing power spectral density calculation on the user behavior control instruction, performing real-time scene noise recognition on the robot, and extracting real-time scene noise signals; Adaptive digital filtering is performed on real-time scene noise signals, and dynamic gain processing of command signal waveforms is performed to construct dynamic gain control commands; Perform instruction decoding and behavioral semantic analysis on dynamic gain control instructions to obtain multiple key motion intention factors; Conduct robot state requirement analysis on multiple key motion intention factors and generate motion state requirements; Intelligent servo parameter adjustment is performed according to motion state requirements to generate underlying servo control instructions.

3. The low-voltage drive intelligent servo control method for a walking robot according to claim 1, characterized in that: The specific steps of step S2 are: The robot foot end sensing parameter set is collected through an integrated multimodal terrain sensing array; Identify and eliminate abnormal outlier parameters of the robot foot sensor parameter set to obtain an outlier-eliminated sensor parameter set; Slip detection and torque feedback calculation are performed on the outlier-removed sensor parameter set to obtain the terrain friction coefficient; The foot end pressure distribution is analyzed based on the sensor parameter set with outlier elimination, and the foot end pressure distribution map is extracted; The terrain interaction state is analyzed based on the foot pressure distribution map and the terrain friction coefficient, and a real-time terrain interaction state map is constructed.

4. The low-voltage drive intelligent servo control method for a walking robot according to claim 1, characterized in that: The specific steps of step S3 are: Collect multi-dimensional joint monitoring data of the robot based on the distributed joint monitoring unit, wherein the multi-dimensional joint monitoring data includes joint angle, angular velocity, current, and torque; Performing time-series joint feature change evolution on the multi-dimensional joint monitoring data to generate a time-series joint feature sequence; The robot's motion behavior is predicted based on the underlying servo control instructions and the real-time terrain interaction state map, thereby obtaining the robot's motion sequence; Calculate the load of each joint in the robot motion sequence based on the time-series joint feature sequence to generate the motion load data of each joint; The load transfer between joints is analyzed based on the motion load data, and a full-joint load distribution perception model is constructed.

5. The low-voltage drive intelligent servo control method for a walking robot according to claim 1, characterized in that: The specific steps of step S4 are: Identify the robot drive battery operating parameters; Calculate real-time voltage output curve, speed and torque based on the robot drive battery operating parameters; Calculate output power based on speed and torque, and calculate energy conversion efficiency based on real-time voltage output curve to construct a three-dimensional motor efficiency map; The minimum output power requirement is calculated based on the full-joint load distribution perception model to obtain the minimum output power requirement value under the operation of each joint; Based on the three-dimensional motor efficiency map, the optimal current-voltage operating point matching is performed for the minimum output power demand value, and dynamic power allocation is performed to build a low-voltage drive control engine.

6. The low-voltage drive intelligent servo control method for a walking robot according to claim 1, characterized in that: The specific steps of step S5 are: Based on the real-time terrain interaction state map, terrain state fluctuation analysis is performed, dynamic walking condition identification is performed, and dynamic walking condition information is generated; Based on dynamic walking condition information, the multi-joint load distribution perception model is used to analyze the multi-joint load trend and make advance predictions to obtain a multi-joint load situation prediction diagram; Define the load situation fluctuation modification threshold; perform threshold trigger identification on the multi-joint load situation prediction graph based on the load situation fluctuation modification threshold, and mark the threshold trigger point; According to the threshold trigger point and the multi-joint load situation prediction diagram, the low-voltage drive control engine is adaptively adjusted in parameters to construct an adaptive drive control engine.

7. The low-voltage drive intelligent servo control method for a walking robot according to claim 1, characterized in that: The specific steps of step S6 are: Execute robot drive control operations based on the adaptive drive control engine and collect the motion posture parameters of each joint; Analyze joint motion characteristics based on the motion posture parameters of each joint and extract the running timing logic and phase relationship between joints; Perform smooth gait optimization based on runtime timing logic and phase relationship to obtain smooth gait optimization parameters; Based on the motion posture parameters of each joint, joint failure analysis and limp gait switching decision are carried out to build a load redistribution fault emergency control mode; Multi-condition collaborative drive control is performed based on the smooth gait optimization parameters and load redistribution fault emergency control mode, and an intelligent collaborative control model is constructed.

8. The low-voltage drive intelligent servo control method for a walking robot according to claim 7, characterized in that: The specific steps of performing joint failure analysis and limp gait switching decision based on the motion posture parameters of each joint and constructing a load redistribution fault emergency control mode are as follows: Perform joint failure analysis on the motion posture parameters of each joint and mark the faulty joints; Identify the location of the faulty joint, identify the fault spreading impact, and mark the fault-involved joints; Performing fault risk assessment based on the fault-involved joints and the fault joints to obtain a fault risk assessment value; The limp gait switching decision is made based on the fault risk assessment value, and the load of the faulty joint is redistributed to construct a load redistribution fault emergency control mode.

9. A driving device, characterized in that: The method for executing the low-voltage drive intelligent servo control method for a walking robot according to claim 1 comprises: The servo control module is used to collect user behavior control instructions output by the decision-making unit, perform behavior semantic analysis and intelligent servo parameter control, and generate low-level servo control instructions; The terrain interaction state module is used to collect the robot's foot-end sensor parameter set, calculate the foot-end pressure distribution, analyze the terrain interaction state, and build a real-time terrain interaction state map; The joint load sensing module is used to predict the robot's motion behavior based on the underlying servo control instructions and the real-time terrain interaction state map, calculate the load of each joint, and build a full-joint load distribution sensing model; The low-voltage drive control module is used to calculate the minimum output power requirement based on the full-joint load distribution perception model, and perform dynamic power allocation to build a low-voltage drive control engine; The load trend analysis module is used to perform multi-joint load trend analysis based on the real-time terrain interaction state map, and to adaptively adjust the parameters of the low-voltage drive control engine to build an adaptive drive control engine; The intelligent collaborative control module is used to execute robot drive control operations based on the adaptive drive control engine, perform joint failure gait reconstruction and multi-working condition collaborative drive control, and build an intelligent collaborative control model.

Images