A motion control method and system for intelligent robots

Abstract

The application provides a motion control method and system for an intelligent robot, comprising: collecting environment and state data through a humanoid robot intelligent sensor group, inputting a pre-trained first neural network model, obtaining motion prediction data and environment analysis results; constructing a motion planning model according to the motion prediction data and the environment analysis results, and adopting a preset first algorithm to perform energy consumption-stability multi-objective optimization on joint motion trajectories; a central controller generates joint position, speed and torque reference trajectories based on the optimization results; and each joint local second controller locally adjusts the reference trajectories within a prediction time domain according to real-time feedback. The application realizes intelligent perception of the environment and the state of the robot through a multi-sensor combined hybrid attention neural network, solves a time sequence dependence problem during multi-sensor data fusion, significantly improves motion efficiency in a complex environment through energy consumption-stability multi-objective optimization, and reduces communication delay and improves system response speed through collaborative work of the central controller and the local controller and combination of an edge computing architecture.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robots, in particular to a motion control method and system for intelligent robots. BACKGROUND

[0002] With the development of intelligent technology, intelligent robots are increasingly widely used in industrial manufacturing, medical rehabilitation and special task execution. However, due to the multi-degree-of-freedom joint system of intelligent robots, the existing robot motion control method has deficiencies in motion efficiency control, response speed and the like. SUMMARY

[0003] The present application is based on the above problems, and proposes a motion control method and system for intelligent robots, which realizes intelligent perception of the environment and its own state through a combination of multiple sensors and a hybrid attention neural network, solves the time sequence dependence problem of multi-sensor data fusion, performs energy-stability multi-objective optimization, significantly improves the motion efficiency in complex environments, and reduces communication delay and improves system response speed through the collaborative work of the central controller and the local controller combined with the edge computing architecture.

[0004] Therefore, one aspect of the present application proposes a motion control method for intelligent robots, comprising:

[0005] An intelligent sensor group arranged on a humanoid robot is used to collect environmental data and robot state data;

[0006] The environmental data and the robot state data are input into a pre-trained first neural network model to obtain motion prediction data of the robot and environmental analysis results;

[0007] A motion planning model is constructed according to the motion prediction data and the environmental analysis results;

[0008] A preset first algorithm is used to perform energy-stability multi-objective optimization on the joint motion trajectory in the motion planning model;

[0009] A central controller generates a reference trajectory of joint position, velocity and torque based on the optimization results;

[0010] A second controller configured locally at each joint adjusts the reference trajectory locally within a prediction time domain according to real-time feedback data.

[0011] Optionally, the step of inputting the environmental data and the robot state data into the pre-trained first neural network model to obtain the motion prediction data of the robot and the environmental analysis results comprises:

[0012] preprocessing the environment data and the robot state data;

[0013] inputting the preprocessed environment data and the robot state data into a spatio-temporal attention sub-model of a first neural network model;

[0014] the spatio-temporal attention sub-model establishes an attention weight matrix for a time dimension;

[0015] the time attention mechanism is used to calculate the correlation between the current time and the historical time data, and identify the time sequence dependence of the motion pattern;

[0016] the spatial attention mechanism is used to analyze the spatial correlation between different sensor data, and establish a mutual influence relationship diagram between sensors;

[0017] the time and spatial attention weights are fused to generate a comprehensive spatio-temporal attention distribution matrix;

[0018] based on the spatio-temporal attention distribution matrix, a sensor fusion adaptive weight sub-model is used to calculate the fusion weight coefficients of each sensor data;

[0019] based on the current environment complexity and task requirements, the importance weights of different sensor data are dynamically adjusted;

[0020] the multi-modal data after weight adjustment are weighted and fused to form a unified feature representation vector;

[0021] a gating mechanism is used to control the information flow, and the most critical sensor information for the current motion decision is selected;

[0022] the fused feature representation vector is input into a motion prediction sub-model to extract the current motion trend and target orientation of the robot;

[0023] the feature representation vector is simultaneously input into an environment analysis sub-model to analyze the geometric structure, obstacle distribution and dynamic changes of the surrounding environment;

[0024] the motion prediction data and the environment analysis results are generated by combining the motion intention and the environment understanding results.

[0025] Optionally, the step of constructing a motion planning model according to the motion prediction data and the environment analysis results comprises:

[0026] The first sub-model is constructed, including: according to the motion prediction data and the environment analysis result, combining the preset task target parameter, a global task decomposition tree is established; based on the obstacle distribution information and the terrain feature in the environment analysis result, a global path from the starting point to the target point is generated; according to the task complexity and the environment dynamics, the global path is segmented into a plurality of sub-path segments, each sub-path segment corresponding to a behavior primitive; an execution priority and a time window are assigned to each behavior primitive, and a time sequence execution sequence containing basic behaviors such as walking, turning, grabbing and obstacle avoidance is generated; an alternative scheme library of the global path is established, which can quickly switch to an alternative path when the main path is blocked;

[0027] The second sub-model is constructed, including: receiving the behavior sequence output by the first sub-model, calling the corresponding motion mode library for each behavior primitive; based on the kinematics and dynamics constraints of the robot, the behavior primitive is decomposed into a set of joint motion trajectories; considering the coupling relationship and coordination constraints between joints, the inverse kinematics algorithm is used to calculate the position, velocity and acceleration target values of each joint; according to the environment constraint information in the environment analysis result, the joint trajectory is subjected to feasibility test and collision detection; the trajectory that does not meet the constraint condition is subjected to local adjustment and optimization to ensure the smoothness and continuity of the trajectory; a detailed joint motion trajectory table containing time stamp is generated, and a checkpoint mechanism for trajectory execution is established;

[0028] The third sub-model is constructed, including: establishing a multi-thread monitoring mechanism to detect environmental changes and sudden conditions in real time, including dynamic obstacle appearance, ground condition change and external interference; when a sudden condition is detected, an emergency response program is immediately started, the current trajectory execution is suspended and a reaction control mode is entered; based on the current sensor data and the robot state, the threat level and the influence range of the sudden condition are quickly evaluated; according to the threat level, a corresponding reaction strategy is selected: for slight interference, trajectory fine-tuning is performed, for medium threat, avoidance action is started, and for serious threat, emergency stop is executed; in the reaction control mode, a rule-based rapid decision algorithm is used to generate a local avoidance trajectory or a protection action; the execution effect of the reaction action is evaluated in real time, and when the threat is eliminated, a connecting trajectory from the current state to the original target is re-planned;

[0029] A coordination and information transmission mechanism between the sub-models is constructed;

[0030] The planning result is output and verified.

[0031] Optionally, the energy consumption-stability multi-objective optimization of the joint motion trajectory in the motion planning model by using the preset first algorithm includes:

[0032] A compound objective function containing an energy consumption term, a motion stability term and a task completion degree term is constructed;

[0033] The quantum annealing mechanism is used to quickly search in the solution space to achieve optimal motion control parameters in complex terrain and dynamic environment.

[0034] The joint stiffness and damping parameters are dynamically adjusted according to the optimization results, so that the robot minimizes energy consumption while maintaining stability.

[0035] Optionally, the step of generating a reference trajectory of joint position, velocity and torque based on the optimization results comprises:

[0036] Perform optimization result analysis and parameter extraction;

[0037] Generate a joint position reference trajectory according to the data obtained by optimization result analysis and parameter extraction;

[0038] Generate a joint velocity reference trajectory based on the joint position reference trajectory;

[0039] Generate a joint torque reference trajectory based on the robot dynamics model, joint position reference trajectory and joint velocity reference trajectory;

[0040] Perform multi-joint coordination and trajectory synchronization;

[0041] Integrate and distribute the joint position reference trajectory, joint velocity reference trajectory and joint torque reference trajectory.

[0042] Optionally, the step of locally adjusting the reference trajectory in the prediction time domain according to the real-time feedback data by the second controller configured locally at each joint comprises:

[0043] Collect the current joint state data of the joint in real time through the position encoder, velocity sensor and torque sensor built-in the joint;

[0044] Time-synchronize and align the current joint state data with the reference trajectory, and calculate the trajectory deviation between the current joint state and the reference trajectory;

[0045] Based on the current joint state data, establish a prediction time domain and perform state prediction;

[0046] Analyze the trajectory deviation and develop a corresponding current adjustment strategy;

[0047] Based on the current state data and the current adjustment strategy, re-plan and optimize the corresponding local trajectory in the parameter trajectory;

[0048] Perform adjacent joint coordination and conflict resolution;

[0049] Evaluate the adjustment effect and feedback.

[0050] Optionally, the sensor fusion adaptive weight sub-model of the first neural network model calculates the dynamic weight of each sensor data in the following manner:

[0051] wherein,

[0052] W i1 (t) represents the weight coefficient of the i1th sensor at time point t; R i1 (t) represents the reliability score of the i1th sensor data, which is calculated by the following formula:

[0053]

[0054] wherein is the sensor data variance, D i1 (t) is the data packet loss rate, Δ i1 (t) is the historical data consistency deviation, β1, β2 and β3 are balance coefficients and satisfy β1+β2+β3=1, λ i1 and ξ i1 are sensor characteristic parameters;

[0055] C i1 (t) represents the relevance score of the i1th sensor to the current task, which is generated by a task context encoder, and increases the environmental condition adaptability;

[0056] T i1 (t) represents the timeliness factor of the i1th sensor, which is defined as:

[0057] T i1 (t) = exp(-γ i1 (t-t last，i1 ))·(1-ν i1 exp(-κ i1 f i1 (t)));

[0058] wherein t last，i1 is the time stamp of the latest valid data of the i1th sensor, γ i1 is a time decay parameter, f i1 (t) is the sensor data sampling frequency, v i1 and κ i1 are adjustment parameters;

[0059] Q i (t) represents the newly added sensor data quality evaluation factor, which is defined as:

[0060]

[0061] wherein SNR i1 (t) is the sensor signal-to-noise ratio, SNRth，i1 is a threshold constant, p i1 and is an adjustment parameter, which can dynamically adjust the weight according to the sensor signal-to-noise ratio;

[0062] H i1 (t) represents the newly added sensor historical performance credibility, which is updated by an exponential moving average method:

[0063] H i1 (t) = a H H i1 (t-1) + (1-a H )P i1 (t);

[0064] where P i1 (t) is the current performance score, and a H is a smoothing coefficient;

[0065] N is the total number of multi-modal sensors.

[0066] Optionally, the first algorithm comprises:

[0067] S11: construct a quantum representation of energy consumption-stability multi-objective optimization, which is evaluated by the following composite objective function:

[0068] F(θ) = a E ·E(θ) + a S ·(1-S(θ)) + a T ·(1-T(θ)) + a F ·(1-F L (θ)) + a A ·A D (θ);

[0069] where θ represents a robot joint control parameter vector;

[0070] E(θ) represents an energy consumption function, which is improved and defined as:

[0071]

[0072] where P i2 (t, θ) is the power consumption of the i2th joint at time t, a i2 (t, θ) is the joint acceleration, ζ i2 is an acceleration energy consumption influence coefficient, M is the total number of joints, t f is the task completion time; the additional effect of acceleration on energy consumption is considered;

[0073] S(θ) represents a motion stability function, which is defined as:

[0074]

[0075] where ZMP dist (t, θ) is the minimum distance from the zero moment point to the support polygon boundary, A CoM (t, θ) is the center of mass acceleration, J ang (t, θ) is the angular momentum rate of change, μ1, μ2 and μ3 are weight coefficients; the angular momentum rate of change term is added, which can more comprehensively evaluate dynamic stability;

[0076] T(θ) represents a task completion degree function, quantifying the degree of task target realization;

[0077] F L (θ) represents a load adaptability function, defined as:

[0078]

[0079] where F ext (t, θ) is the external load force, F est (t, θ) is the estimated load force, T comp (t, θ) is the load compensation torque, ψ1 and ψ2 are weight coefficients; this function evaluates the adaptability of the robot to external load changes;

[0080] A D (θ) represents an action diversity index, defined as:

[0081]

[0082] where D KL is the KL divergence, p θ is the current action distribution, p ref，k is the reference action template distribution, w k is the weight coefficient, and K is the number of action templates; this index promotes the exploration of solution space diversity;

[0083] α E , α S , α T , α F and α A are adaptive weight coefficients, satisfying α E + α S + α T + α F + α A = 1;

[0084] S12: encode each candidate solution in the solution space using qubit probability amplitude, and update the solution space through the following quantum annealing probability function:

[0085]

[0086] where τ(t) = τ0·exp(-η·t / T max )·(1+δ·sin(π·t / t p )) represents the quantum annealing temperature function, τ0is the initial temperature parameter, η is the annealing rate, t max is the maximum iteration number, δ is the fluctuation amplitude coefficient, t p is the fluctuation period;

[0087] is the solution space distance adjustment factor, h(x j2 , x k ) is the Hamming distance, χ and are adjustment parameters;

[0088] The quantum search network is:

[0089]

[0090] where U is the unitary transformation matrix of the i3 layer, representing the quantum system evolution;

[0091] is the perturbation operator, and the parameter ε i3 controls the quantum exploration-exploitation balance;

[0092] is the newly added quantum measurement operator, σ i3 is the measurement intensity parameter, which realizes quantum collapse to prevent the solution space from over-diverging; and L is the number of quantum network layers.

[0093] Optionally, the method further comprises the steps of:

[0094] S21: Based on the robot dynamics model and the optimized trajectory, the central controller generates a feedforward control signal using the following state space prediction equation:

[0095]

[0096] where:

[0097] X (t + k1|t) represents the predicted system state vector at time t + k1, X (t) represents the actual system state vector at time t, containing position, velocity and acceleration information, U (t + j3|t) represents the planned control input at time t + j3, Φ (t) is a time-varying state transition matrix describing the dynamic characteristics of the system, which is updated as follows: Φ (t) = Φ0+ ΔΦ (t) ; wherein Φ0is a nominal model, and ΔΦ (t) is a model deviation identified online; Γ (t) is a time-varying control input matrix considering the change of actuator efficiency: Γ (t) = Γ0· E (t) ; wherein Γ0is a nominal control input matrix, and E (t) is an actuator efficiency diagonal matrix;

[0098] Ψ (t + k1|t) is an environmental disturbance prediction term, which is estimated by a high-order adaptive Kalman filter:

[0099]

[0100] wherein Ψi4is the i4th disturbance model, Ψi4j3is a disturbance interaction term, ω i4 (t) is a first-order adaptive weight, ωi4j3(t) is a second-order interaction weight, N d is the size of the disturbance model library;

[0101] Ω (t + k1|t) represents a contact dynamics prediction term, which is defined as:

[0102]

[0103] wherein N c is the number of possible contact points, p c (t + k1|t) is a contact probability prediction, J c (t + k1|t) is a contact Jacobian matrix, F c (t + k1|t) is a contact force prediction;

[0104] Λ (t, k1) represents a model uncertainty compensation term, which is defined as:

[0105] Λ (t, k1) = B (t) · tanh (W (t) · Z (t, k1)) ;

[0106] wherein B (t) and W (t) are adaptive parameter matrices, Z (t, k) is an eigenvector, and tanh is a hyperbolic tangent function;

[0107] S22: each joint second controller (local controller) adopts a sliding time domain model predictive control strategy, and precise tracking is realized by solving the following constrained optimization problem:

[0108]

[0109] satisfying the constraint condition:

[0110] Umin(t)≤U(t+j4|t)≤Umax(t), j4=0, 1,..., H c -1;

[0111] ΔUmin(t)≤ΔU(t+j4|t)≤ΔUmax(t), j4=0, 1,..., H c -1;

[0112]

[0113] where:

[0114] Y ref (t+j4) represents the reference trajectory; ΔU(t+j4|t) represents the control increment; H p represents the prediction horizon length; H c represents the control horizon length, satisfying H c ≤H p ; Q(t, j4) and R(t, j4) are time-varying output tracking error and control increment weight matrices, respectively; Y tar get is the terminal target state, S is the terminal weight matrix, which enhances control stability; Γ smooth is the smoothness penalty function, which promotes smooth transition of control signals; C safety is the safety constraint function, which ensures that the robot avoids dangerous states;

[0115] S23: The edge computing architecture dynamically schedules computing load through an adaptive resource allocation algorithm, defined as:

[0116]

[0117] satisfying the constraint condition:

[0118]

[0119] R reliability (A(t))≥R min ;

[0120] where:

[0121] A(t) represents the task allocation matrix, element A i5j5 (t) represents the indication variable of the i5th control task allocated to the j5th edge computing node; C i5j5 (t) represents the computing cost of the i5th task on the j5th node; L i5j5 (t) represents the communication delay cost; E i5j5 (t) represents the additional energy consumption cost; Ii5j5 (t) represents an added information security risk indicator; D i5 (t) represents the computing resource requirement of the i5th task; K j5 (t) represents the available computing resource of the j5th node; P i5 (t) represents the power requirement of the i5th task; P max，j5 (t) represents the maximum power limit of the j5th node; R reliability (A(t)) represents the system overall reliability function; R min represents the minimum reliability requirement; M1 is the total number of control tasks; N e is the number of edge computing nodes; w1(t), w2(t), w3(t) and w4(t) are time-varying weight coefficients, satisfying w1(t)+w2(t)+w3(t)+w4(t)=1.

[0122] Another aspect of the present application provides a motion control system for a smart robot, for executing a motion control method for a smart robot, comprising: a humanoid robot; a central controller; a smart sensor group and a second controller arranged on the humanoid robot; and a server;

[0123] The smart sensor group is configured to: collect environmental data and robot state data;

[0124] The server is configured to:

[0125] input the environmental data and the robot state data into a pre-trained first neural network model to obtain motion prediction data of the robot and environmental analysis results;

[0126] According to the motion prediction data and the environmental analysis results, a motion planning model is constructed, which includes:

[0127] A preset first algorithm is used to perform energy consumption-stability multi-objective optimization on the joint motion trajectory in the motion planning model;

[0128] The central controller is configured to: based on the optimization result, generate a reference trajectory of joint position, velocity and torque;

[0129] The second controller arranged locally at each joint is configured to: according to real-time feedback data, locally adjust the reference trajectory within a prediction time domain.

[0130] The technical scheme of the present application is used for a motion control method of an intelligent robot, which comprises: collecting environment data and robot state data by an intelligent sensor group arranged on a humanoid robot; inputting the environment data and the robot state data into a pre-trained first neural network model to obtain motion prediction data of the robot and environment analysis results; constructing a motion planning model according to the motion prediction data and the environment analysis results; performing energy consumption-stability multi-objective optimization on joint motion trajectories in the motion planning model by using a pre-set first algorithm; generating reference trajectories of joint positions, velocities and torques by a central controller based on the optimization results; and performing local adjustment on the reference trajectories within a prediction time domain by a second controller arranged locally at each joint based on real-time feedback data. The intelligent perception of the environment and the state of the robot is realized by combining a mixed attention neural network with multiple sensors, the time sequence dependence problem of multi-sensor data fusion is solved, energy consumption-stability multi-objective optimization is performed, the motion efficiency in a complex environment is significantly improved, and the communication delay is reduced and the system response speed is improved by the collaborative work of the central controller and the local controller in combination with an edge computing architecture. BRIEF DESCRIPTION OF DRAWINGS

[0131] Fig. 1 is a flowchart of a motion control method for an intelligent robot provided by an embodiment of the present application;

[0132] Fig. 2 is a schematic block diagram of a motion control system for an intelligent robot provided by an embodiment of the present application. DETAILED DESCRIPTION

[0133] In order to more clearly understand the above-mentioned purposes, features and advantages of the present application, the present application will be further described in detail below with reference to the drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0134] In the following description, many specific details are set forth in order to provide a thorough understanding of the present application, however, the present application can also be implemented in other different ways from those described herein, therefore, the protection scope of the present application is not limited by the specific embodiments disclosed below.

[0135] The terms "first", "second", etc. in the specification and claims of the present application and the above-mentioned drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed or can optionally include other steps or units inherent to the process, method, product or device.

[0136] Reference to“an embodiment” herein means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the application. The appearances of the phrase“in one embodiment” in various places in the specification are not necessarily referring to the same embodiment, nor are they necessarily mutually exclusive of one another. It is expressly understood that any of the embodiments described herein can be combined with any of the other embodiments unless specifically noted otherwise.

[0137] A motion control method and system for a smart robot are described below with reference to Figs. 1-2 some embodiments of the application.

[0138] As shown in Fig. 1 , one embodiment of the application provides a motion control method for a smart robot, comprising:

[0139] collecting environment data and robot state data through a set of intelligent sensors arranged on the humanoid robot;

[0140] The set of intelligent sensors includes a depth camera, an inertial measurement unit, a torque sensor, a tactile sensor, a thermal imaging camera, and the like.

[0141] inputting the environment data and the robot state data into a pre-trained first neural network model to obtain motion prediction data of the robot and environment analysis results;

[0142] The first neural network model includes a spatio-temporal attention sub-model, a sensor fusion adaptive weight sub-model, a motion prediction sub-model, an environment analysis sub-model, and the like. The spatio-temporal attention mechanism dynamically adjusts the attention degree to different sensor data, solves the time sequence dependence and spatial correlation problem in multi-sensor data fusion, and outputs the motion prediction data of the humanoid robot and the environment analysis results.

[0143] According to the motion prediction data and the environment analysis results, a motion planning model is constructed, which includes: a first sub-model for generating a global path and a behavior sequence according to a task target and environment analysis results; a second sub-model for decomposing the planning of the first sub-model into a specific set of joint motion trajectories; and a third sub-model for real-time motion adjustment for unexpected situations.

[0144] A preset first algorithm (such as a quantum heuristic optimization algorithm) is used to perform energy-stability multi-objective optimization on the joint motion trajectories in the motion planning model.

[0145] A central controller generates a reference trajectory of joint position, velocity, and torque based on the optimization results.

[0146] In this step, the central controller generates a reference trajectory of joint position, velocity and torque based on the optimization result of the joint trajectory in the motion planning model by energy-stability multi-objective optimization based on the first algorithm.

[0147] The second controller configured locally at each joint adjusts the reference trajectory locally in the prediction time domain according to real-time feedback data.

[0148] By using the technical scheme of this embodiment, intelligent perception of the environment and the state of the robot is realized through a combination of multiple sensors and a hybrid attention neural network, the time sequence dependence problem of multi-sensor data fusion is solved, energy-stability multi-objective optimization is performed, and the motion efficiency in a complex environment is significantly improved, and through the collaborative work of the central controller and the local controller, combined with the edge computing architecture, the communication delay is reduced and the system response speed is improved.

[0149] In some possible embodiments of the application, the step of performing an adaptive learning process is further included, which comprises:

[0150] During the motion execution process, the control strategy is continuously updated through the reinforcement learning algorithm to adapt to different environmental conditions and task requirements, a motion experience memory bank is constructed to record successful and failed motion cases and their environmental context, and based on the data in the memory bank, a meta-learning strategy is used to quickly adapt to new environments and tasks.

[0151] In some possible embodiments of the application, the step of realizing human-robot collaborative motion control is further included, which comprises:

[0152] The motion intention of the human collaborator is perceived through visual and force sensing sensors, the motion trajectory and force of the robot are adjusted according to the actions of the human collaborator, and the motion state information of the robot is transmitted to the human collaborator through tactile feedback to realize bidirectional information exchange.

[0153] In some possible embodiments of the application, the step of inputting the environment data and the robot state data into a pre-trained first neural network model to obtain motion prediction data and environment analysis results of the robot comprises:

[0154] The preprocessing of the environment data and the robot state data comprises: performing size standardization and noise filtering processing on image data from a depth camera; performing coordinate system unification and time synchronization calibration on acceleration and angular velocity data of an inertial measurement unit; performing amplitude normalization and outlier detection on data of torque sensors, tactile sensors and thermal imaging cameras respectively; aligning sensor data of different modalities according to a unified timestamp to form a time sequence data stream;

[0155] inputting the preprocessed environmental data and the robot state data into a spatio-temporal attention submodel of a first neural network model;

[0156] The spatio-temporal attention submodel establishes an attention weight matrix for the time dimension;

[0157] The time attention mechanism is used to calculate the correlation between the current time and the historical time data, and identify the time sequence dependence of the motion pattern;

[0158] The spatial attention mechanism is used to analyze the spatial correlation between different sensor data, and establish a mutual influence relationship diagram between sensors;

[0159] The time and space attention weights are fused to generate a comprehensive spatio-temporal attention distribution matrix;

[0160] According to the spatio-temporal attention distribution matrix, the sensor fusion adaptive weight submodel calculates the fusion weight coefficients of each sensor data;

[0161] Based on the current environmental complexity and task requirements, the importance weights of different sensor data are dynamically adjusted;

[0162] The multi-modal data after weight adjustment are weighted and fused to form a unified feature representation vector;

[0163] The gating mechanism is used to control the information flow and filter out the most critical sensor information for the current motion decision;

[0164] The fused feature representation vector is input into a motion prediction submodel to extract the current motion trend and target orientation of the robot;

[0165] The feature representation vector is simultaneously input into an environmental analysis submodel to analyze the geometric structure, obstacle distribution and dynamic changes of the surrounding environment;

[0166] The motion prediction data (including motion direction, speed preference, etc.) and the environmental analysis results (including environmental constraint information) are generated by combining the motion intention and the environmental understanding results.

[0167] In the embodiment, through dynamic weight adjustment of the spatio-temporal attention mechanism, the fusion difficulty of different sensor data in time synchronization and space correlation can be effectively solved, and compared with the traditional fixed weight fusion method, the data fusion accuracy can be improved; the adaptive weight sub-model can adjust the fusion strategy in real time according to the environmental change and the sensor state, when a certain sensor fails or the data quality decreases, the system can automatically reduce the weight of the sensor and enhance the contribution of other sensors, thereby improving the fault tolerance and stability of the system; through modeling of the time sequence dependence relationship, the motion trend of the robot can be more accurately predicted, the misjudgment caused by single-time data fluctuation can be avoided, and the motion intention recognition accuracy is improved; the spatial correlation analysis enables the robot to better understand the internal relationship between the sensor data, and can more accurately identify the environmental characteristics and potential risks in a complex environment, thereby improving the environmental adaptability; through the information screening of the gating mechanism, the processing amount of irrelevant information is reduced, the calculation complexity is reduced while ensuring the decision quality, and real-time control is possible. The hybrid attention neural network architecture provides high-quality input data for subsequent multi-level motion planning, and is a key technical basis for realizing accurate, stable and adaptive robot motion control.

[0168] In some possible embodiments of the present application, the step of constructing a motion planning model according to the motion prediction data and the environment analysis result comprises:

[0169] The first sub-model is constructed, comprising: establishing a global task decomposition tree according to the motion prediction data and the environment analysis result in combination with a preset task target parameter; generating a global path from a starting point to a target point based on obstacle distribution information and terrain characteristics in the environment analysis result; segmenting the global path into a plurality of sub-path segments according to task complexity and environmental dynamics, each sub-path segment corresponding to a behavior primitive; assigning an execution priority and a time window to each behavior primitive to generate a time sequence execution sequence containing basic behaviors such as walking, turning, grabbing and obstacle avoidance; and establishing an alternative path library for the global path, which can quickly switch to an alternative path when the main path is blocked;

[0170] The second sub-model is constructed, comprising: receiving the behavior sequence output by the first sub-model, and calling a corresponding motion mode library for each behavior primitive; decomposing the behavior primitive into a set of joint motion trajectories based on kinematic and dynamic constraints of the robot; considering the coupling relationship and coordination constraints between joints, and calculating the position, velocity and acceleration target values of each joint by using an inverse kinematics algorithm; performing feasibility test and collision detection on the joint trajectories according to the environmental constraint information in the environment analysis result; locally adjusting and optimizing the trajectories that do not meet the constraint conditions to ensure the smoothness and continuity of the trajectories; generating a detailed joint motion trajectory table containing a time stamp, and establishing a checkpoint mechanism for trajectory execution;

[0171] The third sub-model is constructed, including: establishing a multi-thread monitoring mechanism, detecting environmental changes and sudden conditions in real time, including the appearance of dynamic obstacles, changes in ground conditions and external interference; when a sudden condition is detected, an emergency response program is immediately started, the current trajectory execution is suspended and the reaction control mode is entered; based on the current sensor data and the state of the robot, the threat level and the impact range of the sudden condition are quickly evaluated; according to the threat level, the corresponding reaction strategy is selected: for slight interference, the trajectory is fine-tuned, for medium threat, the avoidance action is started, and for serious threat, the emergency stop is executed; in the reaction control mode, a rule-based fast decision algorithm is used to generate a local avoidance trajectory or a protection action; the execution effect of the reaction action is evaluated in real time, and when the threat is eliminated, the connecting trajectory from the current state to the original target is re-planned;

[0172] A coordination and information transmission mechanism between sub-models is constructed, including: establishing an information transmission channel between sub-models, the first sub-model transmits behavior instructions and constraint conditions to the second sub-model, the second sub-model transmits trajectory parameters and execution status to the third sub-model; the third sub-model feeds back sudden condition information and processing results to the second sub-model, and the second sub-model adjusts subsequent trajectory planning according to the feedback information; when the third sub-model is frequently triggered, the second sub-model reports abnormal information to the first sub-model, and the first sub-model re-evaluates the feasibility of the global path; a priority arbitration mechanism is established between sub-models to ensure that when multiple sub-models generate control instructions at the same time, the safety priority principle is followed for selection;

[0173] Planning result output and verification, including: integrating the planning results of each sub-model to generate a complete set of multi-level motion control instructions; performing consistency checking on the integrated instruction set to ensure that there is no conflict between the instructions of each sub-model and that the instructions meet the physical constraints of the robot; establishing a planning quality evaluation mechanism to evaluate the safety, efficiency and executability of the generated motion scheme; outputting the motion planning result that passes the verification to the first algorithm for processing.

[0174] The scheme of the embodiment realizes hierarchical control from macro task planning to micro action execution through cooperation of the three motion planning sub-models, so that complex motion tasks can be orderly decomposed and efficiently executed, and the maintainability and expansibility of the system are improved; the first sub-model guarantees the optimality of the global path and the feasibility of task completion, while the local adjustment mechanism of the second sub-model and the third sub-model ensures rapid adaptation to environmental changes, realizing the organic combination of global optimization and local flexibility; the rapid response mechanism of the third sub-model can handle sudden conditions within milliseconds, greatly improving the safety of the robot in a dynamic environment; by assigning planning tasks of different time scales and complexities to different sub-models, the computational bottleneck of a single algorithm processing all problems is avoided, and the overall calculation efficiency is improved; the multi-model hierarchical planning ensures the continuity and smoothness of motion, reduces frequent trajectory re-planning and sharp action changes, and makes the robot motion more natural and smooth, in line with human movement habits and expectations.

[0175] In some possible embodiments of the application, the first preset algorithm is used to perform energy consumption-stability multi-objective optimization on the joint motion trajectory in the motion planning model, including:

[0176] A composite objective function including an energy consumption term, a motion stability term and a task completion degree term is constructed;

[0177] A quantum annealing mechanism is used to quickly search in the solution space to achieve optimal motion control parameters in complex terrain and dynamic environment;

[0178] According to the optimization result, the joint stiffness and damping parameters are dynamically adjusted to minimize energy consumption while maintaining stability.

[0179] In some possible embodiments of the application, the central controller generates a reference trajectory of joint position, velocity and torque based on the optimization result, including:

[0180] Optimization result analysis and parameter extraction are performed, including: the central controller extracts the target position sequence, expected velocity curve and required torque distribution information of each joint from the optimization result; wherein the optimization result includes the optimal motion control parameter set and the dynamically adjusted joint stiffness and damping parameters; the time allocation scheme in the optimization result is analyzed to determine the duration of each motion stage and the transition time window; environmental constraint parameters and safety boundary conditions are extracted to provide a constraint basis for subsequent trajectory generation; a mapping relationship table of optimization parameters and robot physical parameters is established to ensure the accuracy of parameter conversion;

[0181] According to the data obtained by analyzing the optimization results and extracting parameters, a joint position reference trajectory is generated, including: based on the extracted target position sequence and time allocation scheme, a spline interpolation algorithm is used to generate continuous position trajectories of each joint; the generated position trajectory is subjected to smoothness test to ensure that the trajectory is within the joint limit range and has no abrupt points; according to the kinematic constraints of the robot, the reachability of the position trajectory is verified and the singular point avoidance processing is performed; considering the coupling relationship between joints, the synchronization of the position trajectory of multi-joint cooperative motion is corrected; the time stamp index of the position trajectory is established to ensure the accurate alignment of each joint trajectory in time; the position trajectory extension containing the predicted time domain is generated to provide future state reference for predictive control;

[0182] Based on the joint position reference trajectory, a joint speed reference trajectory is generated, including: based on the time domain differentiation of the joint position reference trajectory, the instantaneous speed and average speed distribution of each joint are calculated; the speed obtained by differentiation is corrected and optimized in combination with the expected speed curve in the optimization result; the generated speed trajectory is ensured not to exceed the maximum speed limit of each joint by applying the speed constraint condition; the continuity and differentiability of the speed trajectory are tested to eliminate speed jumps and discontinuous points; according to the energy consumption optimization requirement, the energy efficiency of the speed trajectory is adjusted to minimize the speed change rate on the premise of meeting the task demand; a consistency verification mechanism of the speed trajectory and the position trajectory is established to ensure the mathematical matching of the two;

[0183] Based on the robot dynamics model, the joint position reference trajectory and the joint speed reference trajectory, a joint torque reference trajectory is generated, including: based on the robot dynamics model and the generated position and speed reference trajectories, the required driving torque of each joint is calculated; the calculated torque is adjusted and optimized in combination with the torque distribution information in the optimization result; considering gravity compensation, friction compensation and external disturbance compensation, a complete torque trajectory containing compensation terms is generated; the generated torque trajectory is ensured to be within the output capacity range of each joint by applying the torque constraint condition; according to the joint stiffness and damping parameters in the optimization result, the torque trajectory is adjusted to adapt to different control mode requirements; a safety monitoring mechanism of the torque trajectory is established to set torque abnormality detection and protection thresholds;

[0184] Multi-joint coordination and trajectory synchronization are executed, including: analyzing the mutual influence relationship between the joint trajectories, establishing a joint coordination matrix and a coupling coefficient table; trajectory coordination optimization is performed on joint groups with strong coupling relationship to ensure the stability of cooperative motion; dynamic balance constraints are established to adjust the balance of joint trajectories affecting the center of gravity; the coordination of upper limb and lower limb motion is realized to ensure the harmony and efficiency of the overall motion; a time sequence synchronization mechanism of trajectory execution is established to ensure that each joint executes motion according to a unified time rhythm;

[0185] The joint position reference trajectory, joint speed reference trajectory and joint torque reference trajectory are integrated and distributed, including: the generated position, speed and torque reference trajectory are formatted and packaged to form a standardized control instruction package; a dedicated trajectory data package is prepared for the second controller (local controller) of each joint, containing complete reference information of the joint; a real-time transmission channel of trajectory data is established to ensure that the reference trajectory can be timely and accurately transmitted to each second controller (local controller); a feedback receiving mechanism of trajectory execution status is established to provide data basis for trajectory adjustment and update; a trajectory execution monitoring report is generated to record the trajectory generation process and key parameter information; and complete reference trajectory information is output to the second controller (local controller) for distributed execution.

[0186] The scheme of the embodiment ensures the complete consistency of the position, speed and torque reference trajectories in mathematics through a systematic trajectory generation process and a multiple verification mechanism, eliminates the control errors caused by trajectory mismatch in traditional methods, and improves the trajectory tracking accuracy; through the establishment of the joint coordination matrix and the coupling coefficient, the precise synchronization and coordinated cooperation of multi-joint motion are realized, the mutual interference and conflict between joints are avoided, and the coordination of the overall motion is improved; the energy consumption optimal parameter of the quantum heuristic optimization result is combined to fully consider the energy efficiency requirement in the trajectory generation process, compared with the traditional trajectory generation method, the energy consumption can be reduced, and the working endurance time of the robot is prolonged; through the trajectory extension in the prediction time domain and the standardized data packaging, an efficient data transmission and processing basis is provided for the distributed control system, the system response delay is reduced, and the strict requirements of real-time control are met. The central controller reference trajectory generation method provides a high-quality control benchmark for the distributed predictive control system, and is a key technical link to realize accurate, stable and efficient humanoid robot motion control.

[0187] In some possible embodiments of the application, the second controller arranged locally at each joint adjusts the reference trajectory locally in the prediction time domain according to the real-time feedback data, including:

[0188] The current joint state data of the joint are collected in real time by the position encoder, speed sensor and torque sensor built in the joint;

[0189] The current joint state data are time-synchronized and aligned with the reference trajectory, and the trajectory deviation (the deviation includes position deviation, speed deviation and torque deviation) between the current joint state and the reference trajectory is calculated; a deviation history record database is established to track the change trend and periodic characteristics of the deviation; the stability and execution quality of the current control state are evaluated according to the deviation size and change rate;

[0190] Based on the current joint state data, a prediction time domain is established and state prediction is performed, including: based on the current joint state data and historical motion data, the prediction time domain length suitable for the joint is established; the possible development trajectory of the future state is deduced within the prediction time domain by using the local dynamics model of the joint; the accuracy of the predicted state is corrected by considering the coupling effect of the motion of adjacent joints on the current joint; the disturbance factors that may occur in the prediction time domain are analyzed, including load change, friction change and external disturbance; a plurality of prediction scenarios are established corresponding to different disturbance degrees and environmental changes; a state confidence interval in the prediction time domain is generated to quantify the uncertainty of the prediction result;

[0191] The trajectory deviation amount is analyzed, and a corresponding current adjustment strategy is formulated, including: the detected trajectory deviation amount is classified and identified to distinguish systematic deviation, random deviation and periodic deviation; the root cause of the deviation is analyzed to determine whether it is caused by model error, parameter drift or external disturbance; according to the type and severity of the deviation, the corresponding adjustment strategy is selected: fine tuning, re-planning or emergency correction; the potential impact of the adjustment operation on the overall motion performance is evaluated to ensure that local adjustment does not damage global coordination; the priority of the adjustment strategy is established to make a reasonable choice when multiple adjustment requirements coexist; the trigger threshold and execution conditions of the adjustment operation are set to avoid unnecessary trajectory modification frequently;

[0192] Based on the current state data and the current adjustment strategy, the corresponding local trajectory in the parameter trajectory is re-planned and optimized, including: within the prediction time domain, a new local trajectory segment is generated based on the current state data, the current adjustment strategy and the target reference point; a fast trajectory planning algorithm is used to ensure that the re-planned trajectory meets the physical constraint conditions of the joint; the re-planned trajectory is smoothed to ensure that there is no mutation at the connection with the original reference trajectory; the balance between energy consumption and execution time is considered to perform multi-objective optimization on the local trajectory; the executability and safety of the adjusted trajectory are verified to ensure that the joint is not overloaded or abnormal; a rollback mechanism for trajectory adjustment is established to quickly recover to the original trajectory when the adjustment effect is not good;

[0193] Adjacent joint coordination and conflict resolution are performed, including: detecting the potential impact of the trajectory adjustment of the current joint on the motion of adjacent joints; communicating with the second controller (local controller) of the adjacent joint to exchange trajectory adjustment information and coordination requirements; establishing a negotiation mechanism between joints to prioritize arbitration when there is a conflict in trajectory adjustment; for motion modes involving multi-joint coordination, cluster-type trajectory coordination adjustment is performed; the coupling effect between joints is monitored in real time to discover and handle coordination problems in time; an emergency handling mechanism is established for coordination failure to ensure that the system can still operate safely when coordination is difficult;

[0194] The adjustment effect is evaluated and fed back, including: real-time monitoring of the execution effect after trajectory adjustment, evaluating the reduction degree of deviation and the improvement of control performance; establishing quantitative evaluation indexes of adjustment effect, including tracking accuracy, energy consumption change and execution time; feeding back the adjustment process and effect information to the central controller to provide data support for subsequent global optimization; updating the adaptive parameters of the second controller (local controller), improving the subsequent control strategy based on adjustment experience; establishing a learning mechanism for adjustment history, accumulating typical adjustment cases for quick response to similar situations; generating a log record of adjustment operations to provide detailed information for system maintenance and fault diagnosis;

[0195] Continuous optimization and adaptive adjustment, including: continuously optimizing the accuracy of the prediction model and the length of the prediction time domain based on continuous feedback data; adjusting the detection threshold of trajectory deviation and response sensitivity based on long-term running data; learning and memorizing frequent adjustment patterns to establish preset adjustment schemes for quick response; periodically evaluating and updating the parameters of the joint dynamics model to adapt to equipment aging and performance changes; optimizing the overall control strategy of the local controller based on the success rate and quality evaluation of task execution; outputting the optimization results to the edge computing architecture for distributed processing and execution.

[0196] The scheme of the embodiment can timely find and correct trajectory deviation through the combination of real-time feedback and predictive control, can improve trajectory tracking accuracy, and can significantly improve the accuracy of motion control; the local adjustment mechanism can effectively cope with uncertain factors such as external disturbance, model error and parameter change, so that the system can maintain stable control performance when facing various disturbances, and the anti-interference ability is improved; through the forward-looking control of the prediction time domain and the fast trajectory re-planning algorithm, the response speed is improved, and the strict requirements of real-time control are met; the multi-joint coordination mechanism ensures that local adjustment will not damage the coordination of the overall motion, maintains the fluency and naturalness of the motion through information exchange and negotiation between joints, and improves the overall motion quality; through intelligent adjustment strategy selection and multi-objective optimization, the control performance is guaranteed while the additional energy consumption is minimized, and the energy consumption is reduced. This distributed local controller trajectory local adjustment method realizes the organic combination of global optimization and local precise control, and is a key technical means for building a high-performance humanoid robot motion control system.

[0197] In some possible embodiments of the application, the sensor fusion adaptive weight sub-model of the first neural network model calculates the dynamic weight of each sensor data in the following manner:

[0198] wherein,

[0199] w i1 (t) represents the weight coefficient of the i1th sensor at the time point t;

[0200] R i1 (t) represents the reliability score of the i1th sensor data, which is calculated by the following formula:

[0201]

[0202] where is the sensor data variance, D i1 (t) is the data packet loss rate, Δ i1 (t) is the historical data consistency deviation, β1, β2 and β3 are balance coefficients and satisfy β1+β2+β3=1, λ i1 and ξ i1 are sensor characteristic parameters;

[0203] C i1 (t) represents the relevance score of the i1th sensor to the current task, which is generated by the task context encoder, increasing the environmental condition adaptability;

[0204] T i1 (t) represents the timeliness factor of the i1th sensor, which is defined as:

[0205] T i1 (t) = exp(-γ i1 (t-t last，i1 ))·(1-v i1 exp(-κ i1 f i1 (t)));

[0206] where t last，i1 is the timestamp of the latest valid data of the i1th sensor, γ i1 is the time decay parameter, f i1 (t) is the sensor data sampling frequency, ν i1 and κ i1 are adjustment parameters;

[0207] Q i (t) represents the newly added sensor data quality evaluation factor, which is defined as:

[0208]

[0209] where SNR i1 (t) is the sensor signal-to-noise ratio, SNR th，i1 is the threshold constant, ρ i1 and are adjustment parameters, and this factor can dynamically adjust the weight according to the sensor signal-to-noise ratio;

[0210] H i1(t) represents the newly added sensor historical performance credibility, which is updated by an exponential moving average method:

[0211] H i1 (t) = a H H i1 (t-1) + (1-a H )P i1 (t);

[0212] Where P i1 (t) is the current performance score, and a H is a smoothing coefficient;

[0213] N is the total number of multi-modal sensors.

[0214] The weight calculation method in the embodiment realizes adaptive fusion of each sensor information under different scenes and tasks, improves the robustness and adaptability of the system in a noisy environment, partial sensor failure or data loss condition; the sensor data quality evaluation and historical performance credibility factor realize more comprehensive sensor reliability evaluation, and more accurate weight allocation is realized through multi-dimensional measurement, which can improve the accuracy of fused data in a complex and variable environment, and improve the robustness of the system in the case of partial sensor failure.

[0215] In some possible embodiments of the application, the first algorithm comprises:

[0216] S11: construct a quantum representation of energy consumption-stability multi-objective optimization, and evaluate by the following composite objective function:

[0217] F(theta) = alpha E * E(theta) + alpha S * (1-S(theta)) + alpha T * (1-T(theta)) + alpha F * (1-F L (theta)) + alpha A * A D (theta);

[0218] Wherein, theta represents a robot joint control parameter vector;

[0219] E(theta) represents an energy consumption function, and the improved definition is:

[0220]

[0221] Where P i2 (t, theta) is the power consumption of the i2th joint at time t, a i2 (t, theta) is the joint acceleration, zeta i2 is the acceleration energy consumption influence coefficient, M is the total number of joints, and tf is the task completion time; the additional effect of acceleration on energy consumption is considered;

[0222] S(θ) represents the motion stability function, defined as:

[0223]

[0224] where ZMP dist (t, θ) is the minimum distance from the ZMP to the support polygon boundary, A CoM (t, θ) is the center of mass acceleration, J ang (t, θ) is the angular momentum rate of change, μ1, μ2 and μ3 are weight coefficients; the angular momentum rate of change term is added to more comprehensively evaluate dynamic stability;

[0225] T(θ) represents the task completion degree function, quantifying the degree of task goal realization;

[0226] F L (θ) represents the load adaptability function, defined as:

[0227]

[0228] where F ext (t, θ) is the external load force, F ext (t, θ) is the estimated load force, T comp (t, θ) is the load compensation torque, ψ1 and ψ2 are weight coefficients; this function evaluates the adaptability of the robot to external load changes;

[0229] A D (θ) represents the motion diversity index, defined as:

[0230]

[0231] where D KL is the K L divergence, p θ is the current motion distribution, p ref,k is the reference motion template distribution, w k is the weight coefficient, and K is the number of motion templates; this index promotes exploration of the solution space diversity;

[0232] α E , α S , α T , α F and α A are adaptive weight coefficients, satisfying α E + α S + α T + α F + α A = 1.

[0233] S12: encode each candidate solution in the solution space using qubit probability amplitude encoding, and update the solution space by a quantum annealing probability function as follows:

[0234]

[0235] wherein τ(t) = τ0·exp(-η·t / t max )·(1+δ·sin(π·t / t p )) represents a quantum annealing temperature function, τ0 is an initial temperature parameter, η is an annealing rate, t max is the maximum number of iterations, δ is a fluctuation amplitude coefficient, and t p is a fluctuation period;

[0236] is a solution space distance adjustment factor, h(x j2 , x k ) is a Hamming distance, χ and are adjustment parameters;

[0237] The quantum search network is:

[0238]

[0239] wherein U is a unitary transformation matrix of the i3 layer, representing quantum system evolution;

[0240] is a perturbation operator, and the parameter ε i3 controls the quantum exploration-exploitation balance;

[0241] is a newly added quantum measurement operator, σ i3 is a measurement strength parameter, and quantum collapse is implemented to prevent the solution space from over-diverging; and L is the number of quantum network layers.

[0242] In the embodiment, the quantum heuristic optimization algorithm can quickly find an approximate optimal solution in a complex multi-objective optimization problem, and the calculation efficiency is improved compared with a traditional optimization method; load adaptability and action diversity evaluation are increased, and the quantum annealing strategy and search network structure are optimized, so that the algorithm has better local optimal escape ability and global search efficiency, and the calculation efficiency is improved in a complex multi-objective optimization problem, and the convergence speed and the quality of the optimization result are improved.

[0243] In some possible embodiments of the application, the method further comprises the steps of:

[0244] S21: The central controller generates a feedforward control signal using the following state space prediction equation based on the robot dynamics model and the optimized trajectory:

[0245]

[0246] wherein:

[0247] represents the system state vector predicted at time t+k1; X(t) represents the actual system state vector at time t, containing position, velocity, and acceleration information; U(t+j3|t) represents the control input planned at time t for time t+j3; Φ(t) is a time-varying state transition matrix that describes the dynamic characteristics of the system and is updated as follows: Φ(t) = Φ0+ ΔΦ(t); where Φ0is the nominal model and ΔΦ(t) is the model deviation identified online; Γ(t) is a time-varying control input matrix that takes into account the changes in actuator efficiency:

[0248] Γ(t) = Γ0· E(t); where Γ0is the nominal control input matrix and E(t) is the actuator efficiency diagonal matrix;

[0249] Ψ(t+k1|t) is the environmental disturbance prediction term, which is estimated by a high-order adaptive Kalman filter:

[0250]

[0251] wherein Ψi4is the i4th disturbance model, Ψi4j3is the disturbance interaction term, ω i4 (t) is a first-order adaptive weight, ωi4j3(t) is a second-order interaction weight, N d is the size of the disturbance model library;

[0252] Ω(t+k1|t) represents the contact dynamics prediction term, which is defined as:

[0253]

[0254] wherein N c is the number of possible contact points, p c (t+k1|t) is the contact probability prediction, J c (t+k1|t) is the contact Jacobian matrix, F c (t+k1|t) is the contact force prediction;

[0255] Λ(t, k1) represents the model uncertainty compensation term, which is defined as:

[0256] Λ(t, k1) = B(t) · tanh(W(t) · Z(t, k1));

[0257] where B(t) and W(t) are adaptive parameter matrices, Z(t, k) is the eigenvector, and tanh is the hyperbolic tangent function;

[0258] S22: Each joint second controller (local controller) adopts a sliding horizon model predictive control strategy to achieve accurate tracking by solving the following constrained optimization problem:

[0259]

[0260] Subject to constraints:

[0261] Umin(t)≤U(t+j4|t)≤Umax(t), j4=0, 1,..., H c -1;

[0262] ΔUmin(t)≤ΔU(t+j4|t)≤ΔUmax(t), j4=0, 1,..., H c -1;

[0263]

[0264] where:

[0265] Y ref (t+j4) represents the reference trajectory; ΔU(t+j4|t) represents the control increment; H p represents the prediction horizon length; H c represents the control horizon length, satisfying H c ≤H p ; Q(t, j4) and R(t, j4) are time-varying output tracking error and control increment weight matrices, respectively; Y target is the terminal target state, S is the terminal weight matrix, and Γ smooth is the smoothing penalty function, which promotes smooth transition of the control signal; C safety is the safety constraint function, which ensures that the robot avoids dangerous states;

[0266] S23: The edge computing architecture dynamically schedules the computing load through an adaptive resource allocation algorithm, defined as:

[0267]

[0268] Subject to constraints:

[0269]

[0270]

[0271] R reliability(A(t)) ≥ R min ;

[0272] wherein:

[0273] A(t) denotes a task allocation matrix, element A i5j5 (i,j) (t) denotes an indicator variable that the ith control task is allocated to the jth edge computing node; C i5j5 (i,j) (t) denotes the computing cost of the ith task at the jth node; L i5j5 (i,j) (t) denotes the communication delay cost; E i5j5 (i,j) (t) denotes the newly added energy consumption cost;

[0274] I i5j5 (i,j) (t) denotes the newly added information security risk indicator; D i5 (i,j) (t) denotes the computing resource requirement of the ith task; K j5 (i,j) (t) denotes the available computing resource of the jth node; P i5 (i,j) (t) denotes the power requirement of the ith task; P max，j5 (i,j) (t) denotes the maximum power limit of the jth node; R reliability (A(t)) denotes the system overall reliability function; R min denotes the minimum reliability requirement; M1 is the total number of control tasks; N e is the number of edge computing nodes; w1(t), w2(t), w3(t) and w4(t) are time-varying weight coefficients, satisfying w1(t)+w2(t)+w3(t)+w4(t)=1.

[0275] In the embodiment, the distributed predictive control system realizes the organic combination of global planning and local fine control through the central-local double-layer control architecture, the edge computing architecture reduces the system delay, and improves the response speed of the humanoid robot in a high dynamic environment; by introducing a time-varying system model, contact dynamics prediction, model uncertainty compensation, terminal constraint and a more comprehensive edge computing resource allocation strategy, the robustness and adaptability of the control system are significantly improved, the prediction accuracy is improved in a dynamic and variable environment, the system response delay is reduced, the control stability is improved, and the energy efficiency is improved.

[0276] Referring to Fig. 2 , another embodiment of the present application provides a motion control system for an intelligent robot, for executing a motion control method for an intelligent robot, comprising: a humanoid robot; a central controller; an intelligent sensor group and a second controller arranged on the humanoid robot; and a server;

[0277] The intelligent sensor group is configured to: collect environment data and robot state data;

[0278] The server is configured to:

[0279] input the environment data and the robot state data into a pre-trained first neural network model to obtain motion prediction data of the robot and environment analysis results;

[0280] construct a motion planning model according to the motion prediction data and the environment analysis results, the model comprising:

[0281] perform energy consumption-stability multi-objective optimization on joint motion trajectories in the motion planning model by using a preset first algorithm;

[0282] The central controller is configured to generate reference trajectories of joint positions, velocities and torques based on the optimization results;

[0283] The second controller configured locally at each joint is configured to locally adjust the reference trajectories within a prediction time domain according to real-time feedback data.

[0284] It should be understood that, Fig. 2 The block diagram of the motion control system for the intelligent robot shown is only schematic, and the number of each module shown does not limit the protection scope of the present application. The motion control system for the intelligent robot provided in the embodiment can be used to execute each embodiment scheme of the corresponding motion control method for the intelligent robot. For the specific implementation process, please refer to the description of each method embodiment, which will not be repeated here.

[0285] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited by the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily necessary for the present application.

[0286] In the above embodiments, the description of each embodiment has its own emphasis, and the parts not described in detail in a certain embodiment can be referred to the related description of other embodiments.

[0287] In several embodiments provided in the present application, it should be understood that the disclosed apparatus can be implemented in other manners. For example, the division of the apparatus embodiments described above is merely illustrative, and the division of the units can be changed according to actual needs. For example, the units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the displayed or discussed mutual couplings or direct couplings or communication connections can be indirect couplings or communication connections through some interfaces, devices or units, and can be in electrical, mechanical or other forms.

[0288] The units described as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, i.e., can be located in one place, or can be distributed on multiple network units. Some or all of the units can be selected according to actual needs to achieve the purpose of the embodiments.

[0289] In addition, the functional units in each embodiment of the present application can be integrated in one processing unit, or each unit can be physically present separately, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0290] The integrated unit, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable memory. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that contributes to the technical solutions or all or part of the technical solutions can be embodied in the form of a software product, which is stored in a memory and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the above-mentioned method of each embodiment of the present application. The aforementioned memory includes: a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0291] Those of ordinary skill in the art can understand that all or part of the steps of the various methods in the above embodiments can be completed by a program instructing relevant hardware, and the program can be stored in a computer readable memory, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0292] The above has carried out the detailed introduction to the embodiment of the application, the principle and implementation mode of the application have been set forth in this article by applying specific examples, the above embodiment explanation is only for helping to understand the method of the application and its core idea; at the same time, for the general technical personnel in the art, according to the idea of the application, the specific implementation mode and application range will have changes, and the above is not understood as the limitation of the application.

[0293] Although the present application is disclosed as above, the present application is not limited to this. Any person skilled in the art can easily think of changes or substitutions without departing from the spirit and scope of the present application, and various changes and modifications can be made, including the combination of different functions and implementation steps, including software and hardware implementation, which are all within the protection scope of the present application.

Claims

1. A motion control method for a smart robot, characterized by, The application relates to a human-shaped robot motion planning method based on multi-sensor fusion and quantum annealing mechanism. Collecting environment data and robot state data through a group of intelligent sensors arranged on the human-shaped robot; Inputting the environment data and the robot state data into a pre-trained first neural network model to obtain robot motion prediction data and environment analysis results; According to the motion prediction data and the environment analysis results, a motion planning model is constructed, which comprises a first sub-model for generating a global path and a behavior sequence according to a task target and environment analysis results; A preset first algorithm is used to perform energy consumption-stability multi-objective optimization on joint motion trajectories in the motion planning model, specifically including: constructing a composite objective function containing an energy consumption term, a motion stability term and a task completion degree term; using a quantum annealing mechanism to quickly search in the solution space to achieve optimal motion control parameters in complex terrain and dynamic environment; and dynamically adjusting joint stiffness and damping parameters according to the optimization results to minimize energy consumption while maintaining stability; A central controller generates reference trajectories of joint positions, velocities and torques based on the optimization results; A second controller configured locally at each joint adjusts the reference trajectories locally within a predicted time domain according to real-time feedback data. 2.The motion control method for a smart robot according to claim 1, wherein The step of inputting the environment data and the robot state data into a pre-trained first neural network model to obtain robot motion prediction data and environment analysis results comprises: Preprocessing the environment data and the robot state data; Inputting the preprocessed environment data and the robot state data into a spatio-temporal attention sub-model of the first neural network model; The spatio-temporal attention sub-model establishes an attention weight matrix for the time dimension; The time attention mechanism is used to calculate the correlation between the current time and the historical time data to identify the time sequence dependence relationship of the motion mode; Based on the spatial attention mechanism, the spatial correlation between different sensor data is analyzed to establish a mutual influence relationship diagram among the sensors; The time and space attention weights are fused to generate a comprehensive spatio-temporal attention distribution matrix; According to the spatio-temporal attention distribution matrix, a sensor fusion adaptive weight sub-model calculates the fusion weight coefficients of each sensor data; Based on the current environment complexity and task demand, the importance weights of different sensor data are dynamically adjusted; The multi-modal data after weight adjustment are weighted and fused to form a unified feature representation vector; The information flow is controlled through a gating mechanism to filter out the most critical sensor information for current motion decision; The fused feature representation vector is input into a motion prediction sub-model to extract the current motion trend and target orientation of the robot; The feature representation vector is input into an environment analysis sub-model to analyze the geometric structure, obstacle distribution and dynamic changes of the surrounding environment; Combining the motion intention and the environment understanding result, motion prediction data and environment analysis results are generated. 3.The motion control method for a smart robot according to claim 2, wherein The step of constructing a motion planning model according to the motion prediction data and the environment analysis results comprises: The first sub-model is constructed, including: according to the motion prediction data and the environment analysis result, combining the preset task target parameter, a global task decomposition tree is established; based on the obstacle distribution information and the terrain characteristics in the environment analysis result, a global path from the starting point to the target point is generated; according to the task complexity and the environment dynamics, the global path is divided into a plurality of sub-path segments, and each sub-path segment corresponds to a behavior primitive; an execution priority and a time window are assigned to each behavior primitive, and a time sequence execution sequence containing basic behaviors such as walking, turning, grabbing and obstacle avoidance is generated; an alternative scheme library of the global path is established, which can quickly switch to an alternative path when the main path is blocked; The second sub-model is constructed, including: receiving the behavior sequence output by the first sub-model, calling the corresponding motion mode library for each behavior primitive; based on the kinematics and dynamics constraints of the robot, the behavior primitive is decomposed into a set of joint motion trajectories; considering the coupling relationship and coordination constraints between joints, the inverse kinematics algorithm is used to calculate the position, velocity and acceleration target values of each joint; according to the environment constraint information in the environment analysis result, the joint trajectory is subjected to feasibility test and collision detection; the trajectory that does not meet the constraint condition is locally adjusted and optimized to ensure the smoothness and continuity of the trajectory; a detailed joint motion trajectory table containing a time stamp is generated, and a checkpoint mechanism for trajectory execution is established; The third sub-model is constructed, including: establishing a multi-thread monitoring mechanism to detect environmental changes and sudden conditions in real time, including dynamic obstacle appearance, ground condition change and external interference; when a sudden condition is detected, an emergency response program is immediately started, the current trajectory execution is suspended and a reaction control mode is entered; based on the current sensor data and the robot state, the threat level and the influence range of the sudden condition are quickly evaluated; according to the threat level, the corresponding reaction strategy is selected: for slight interference, the trajectory is fine-tuned; for medium threat, an avoidance action is started; for serious threat, an emergency stop is executed; in the reaction control mode, a rule-based fast decision algorithm is used to generate a local avoidance trajectory or a protection action; the execution effect of the reaction action is evaluated in real time, and when the threat is eliminated, a connection trajectory from the current state to the original target is re-planned; A coordination and information transmission mechanism between the sub-models is constructed; The planning result is output and verified. 4.The motion control method for a smart robot according to claim 3, wherein The step of generating the reference trajectory of joint position, velocity and torque by the central controller based on the optimization result includes: Performing optimization result analysis and parameter extraction; Generating a joint position reference trajectory according to the data obtained by optimization result analysis and parameter extraction; Based on the joint position reference trajectory, a joint velocity reference trajectory is generated; Based on the robot dynamics model, the joint position reference trajectory and the joint velocity reference trajectory, a joint torque reference trajectory is generated; Performing multi-joint coordination and trajectory synchronization; Integrate and distribute the joint position reference trajectory, the joint velocity reference trajectory and the joint torque reference trajectory. 5.The motion control method for a smart robot according to claim 4, wherein The step of locally adjusting the reference trajectory in the prediction time domain by the second controller configured locally at each joint based on real-time feedback data includes: The current joint state data of the joint is collected in real time by a position encoder, a speed sensor and a torque sensor built in the joint; The current joint state data is time-synchronized and aligned with the reference trajectory, and the trajectory deviation between the current joint state and the reference trajectory is calculated; Based on the current joint state data, a prediction time domain is established and state prediction is performed; The trajectory deviation is analyzed, and a corresponding current adjustment strategy is formulated; Based on the current state data and the current adjustment strategy, the corresponding local trajectory in the parameter trajectory is re-planned and optimized; Adjacent joint coordination and conflict resolution are performed; The adjustment effect is evaluated and feedback is provided. 6.The motion control method for a smart robot according to claim 5, wherein The sensor fusion adaptive weight sub-model of the first neural network model calculates the dynamic weight of each sensor data in the following manner: Wherein, represents a weight coefficient of the ith sensor at the time point t; represents a reliability score of the ith sensor data, calculated by the following formula: ， wherein is a sensor data variance, is a data packet loss rate, is a historical data consistency bias, , and are balancing coefficients and satisfy , and are sensor characteristic parameters; represents the relevance score of the ith sensor to the current task, generated by the task context encoder, increasing environmental condition adaptability; denotes the time factor of the ith sensor, defined as: ； wherein is the time stamp of the last valid data of the ith sensor, is a time decay parameter, is the sensor data sampling frequency, and is a tuning parameter; represents the newly added sensor data quality assessment factor, defined as: wherein is a sensor signal-to-noise ratio, is a threshold constant, and is an adjustment parameter, which factor can dynamically adjust the weight according to the sensor signal-to-noise ratio; represents the newly added sensor historical performance credibility, which is updated by exponential moving average method: ； wherein is the current performance score, is a smoothing coefficient; N is the total number of multi-modal sensors. 7.The motion control method for a smart robot according to claim 6, wherein The first algorithm includes: S11: Construct a quantum representation of energy consumption-stability multi-objective optimization, and evaluate by the following compound objective function: ； Wherein, θ represents the robot joint control parameter vector; represents the energy consumption function, the improvement definition is: ； wherein is the power consumption of the ith joint at time t, is the joint acceleration, is the acceleration energy consumption impact coefficient, M is the total number of joints, is the task completion time; the additional impact of acceleration on energy consumption is considered; denotes the motion stability function, defined as: wherein is the minimum distance from the zero moment point to the support polygon boundary, is the center of mass acceleration, is the angular momentum rate of change, , and are weight coefficients; the angular momentum rate of change term is added to more comprehensively evaluate dynamic stability; represents a task completion degree function, quantifying the degree of achievement of the task goal; denotes the load adaptivity function, defined as: wherein is the external load force, is the estimated load force, is the load compensation torque, and is a weight factor; this function evaluates the robot's ability to adapt to external load changes; Action diversity index, defined as: ； wherein divergence, is the current action distribution, is the reference action template distribution, is the weight coefficient, K is the number of action templates; this index promotes the exploration of solution space diversity; , , , and are adaptive weight coefficients satisfying ; S12: Encode each candidate solution in the solution space by quantum bit probability amplitude, and update the solution space by the following quantum annealing probability function: wherein, represents a quantum annealing temperature function, with a periodic fluctuation term added, is an initial temperature parameter, and η is an annealing rate, is a maximum number of iterations, and δ is a fluctuation amplitude coefficient, is a fluctuation period; is the solution space distance adjustment factor, is the Hamming distance, χ and is the adjustment parameter; The quantum search network is: ； wherein, is the unitary transformation matrix of the i3rd layer, representing the quantum system evolution; for the perturbation operator, the parameter controlling the quantum exploration-exploitation balance; for the newly added quantum measurement operator, for the measurement intensity parameter, implementing quantum collapse to prevent the solution space from over-diverging; L is the number of quantum network layers. 8.The motion control method for a smart robot according to claim 7, wherein Further comprising the steps of: S21: The central controller generates a feedforward control signal based on the robot dynamics model and the optimized trajectory using the following state space prediction equation: Wherein: represents the system state vector predicted at time t, t+k1; represents the actual system state vector at time t, containing position, velocity and acceleration information; represents the control input planned at time t, t+j3; is a time-varying state transition matrix, describing the system dynamics, which is updated by: where is the nominal model, is the model bias identified online; is a time-varying control input matrix, considering actuator efficiency variation: where is the nominal control input matrix, is the actuator efficiency diagonal matrix; For the environmental disturbance prediction term, the high-order adaptive Kalman filter is used to estimate: wherein is a 4th order disturbance model, is a disturbance interaction term, is a first order adaptive weight, is a second order interaction weight, is a disturbance model library size; represents the contact kinetics prediction term, defined as: wherein is the number of possible contact points, is the contact probability prediction, is the contact Jacobian, is the contact force prediction; represents a model uncertainty compensation term, defined as: wherein and is an adaptive parameter matrix, is an eigenvector, is a hyperbolic tangent function; S22: The second controller of each joint adopts a sliding horizon model predictive control strategy to achieve accurate tracking by solving the following constrained optimization problem: Satisfy the constraint condition: Wherein: represents the system output at time t + t4predicted at time t; represents the reference trajectory; represents the control increment; represents the prediction horizon length; represents the control horizon length, satisfying ; and are the time-varying output tracking error and control increment weight matrices, respectively; is the terminal target state, is the terminal weight matrix, enhancing control stability; is the smoothness penalty function, promoting smooth transition of control signals; is the safety constraint function, ensuring the robot avoids dangerous states; S23: The edge computing architecture dynamically schedules the computing load by an adaptive resource allocation algorithm, defined as: Satisfy the constraint condition: Wherein: denotes the task allocation matrix, element denotes the indicator variable that the ith control task is allocated to the jth edge computing node; denotes the computation cost of the ith task on the jth node; denotes the communication delay cost; denotes the added energy consumption cost; denotes the added information security risk indicator; denotes the computation resource requirement of the ith task; denotes the available computation resource of the jth node; denotes the power requirement of the ith task; denotes the maximum power limit of the jth node; denotes the system overall reliability function; denotes the minimum reliability requirement; M1 is the total number of control tasks; is the number of edge computing nodes; , , and are time-varying weight coefficients satisfying . 9.A motion control system for a smart robot, for executing the motion control method for a smart robot according to any one of claims 1 to 8, characterized in that, It includes: Humanoid robot; Central controller; intelligent sensor group and second controller arranged on the humanoid robot; And server; The intelligent sensor group is configured to collect environmental data and robot state data; The server is configured to: Input the environmental data and the robot state data into a pre-trained first neural network model to obtain motion prediction data and environmental analysis results of the robot; According to the motion prediction data and the environmental analysis results, a motion planning model is constructed, which includes; A preset first algorithm is used to perform energy consumption-stability multi-objective optimization on the joint motion trajectory in the motion planning model; The central controller is configured to generate a reference trajectory of joint position, velocity and torque based on the optimization result; The second controller arranged locally at each joint is configured to make local adjustments to the reference trajectory within a prediction time domain according to real-time feedback data.

Images