A magnetic micro-robot parallel cooperative field navigation system

Abstract

The application discloses a kind of magnetic micro-robot parallel collaborative field navigation system, it is related to micro-nano operation technical field, and the micro resource conflict resolution scheduler MCDO in this system is based on the smooth single-machine path of dynamic evolution navigation planner DEN, the safety path constraint of conflict-aware evolutionary planner CAEP and the coil current with anti-interference control of double-field disturbance suppressor DFDS, to each magnetic micro-robot planning obstacle avoidance path and outputting multi-machine conflict-free instruction, to drive coil array to generate distributed gradient magnetic field, realize multi-robot independent drive;Micro-vision anchor MVAM is used to track the motion trajectory of magnetic micro-robot in real time, and provides robot pose data for MCDO and DFDS.The above-mentioned module realizes the whole process logic closed-loop management and control from path planning to physical drive, interference suppression to state monitoring, and is improved in path planning efficiency, anti-interference control accuracy, multi-machine cooperation capability, environment adaptation and the like.

Description

Technical Field

[0001] This invention relates to the field of micro-nano manipulation technology, and in particular to a parallel cooperative field navigation system for magnetic microrobots. Background Technology

[0002] With the increasing demand for minimally invasive medical procedures and precision industrial applications, magnetic field-driven microrobots have become a research hotspot in the field of intelligent control due to their advantages of being cableless and having high penetration capabilities. While traditional magnetic field-driven technologies rely on uniform magnetic fields for basic motion control and offer convenient energy transmission, they are limited by static magnetic field configurations and cannot dynamically adapt to three-dimensional trajectory planning in complex environments, resulting in insufficient autonomous obstacle avoidance and precise positioning capabilities. For autonomous navigation strategies in complex environments, traditional path planning algorithms (such as RTT) generate paths with problems of "high randomness, numerous inflection points, and long distances." For example, in complex vascular environments, paths may lead to detours or collisions due to local optima.

[0003] In recent years, intelligent navigation technology has emerged, which endows robots with environmental perception capabilities through distributed magnetic field arrays and reinforcement learning algorithms. However, when dealing with dynamic disturbances of biological fluids, it can cause real-time bottlenecks due to the high-dimensional calculation of the magnetic field-material coupling model. In particular, there are decision delay problems in multi-robot collaborative scenarios. For example, when multiple collaborating entities are performing tasks, collision risks, task blockage, or reduced efficiency may occur due to overlapping or competing target positions. Furthermore, when multiple devices that rely on magnetic fields (such as magnetic navigation robots) work together in close proximity, the magnetic fields they generate are coupled with each other, resulting in sensor reading deviations, actuator motion distortion, or data transmission errors. This can severely restrict the efficiency of cluster operations. Summary of the Invention

[0004] To address the aforementioned problems and technical requirements, the inventors have proposed a parallel cooperative field navigation system for magnetic microrobots. The objectives of this invention include:

[0005] (1) A Dynamic Evolutionary Navigation Planner (DEN) is proposed. It generates an initial tree-like path through a hybrid sampling strategy and iteratively optimizes the generated path nodes by combining an adaptive adjustment search strategy (mean vector and covariance matrix), thereby reducing sharp turns and redundant transitions in the path and making the path smoother.

[0006] (2) A Conflict-Aware Evolutionary Planner (CAEP) is proposed. For the smooth single-machine path planned by the aforementioned algorithm, selection, crossover, and mutation operations of the evolutionary strategy are used to perform gene-level optimization (such as coordinate parameters) on the generated path nodes, making the path closer to the global optimum while avoiding obstacles. Simultaneously, the optimization objective can be dynamically adjusted, i.e., by defining the path length (f... l ), angle change (f) a) and obstacle distance (f d The multi-objective function dynamically adjusts the weights of each objective according to task requirements, enabling flexible path planning.

[0007] (3) A dual-field disturbance suppressor (DFDS) is proposed. To address the nonlinear dynamic coupling of microrobots in magnetic fields (such as the coupling effect between Maxwell coils and Helmholtz coils), an extended state observer (ESO) is designed to estimate unknown dynamics and external disturbances (such as fluid resistance and friction) of the system in real time, and decoupled control of magnetic field input is achieved through feedforward compensation.

[0008] (4) A micro-visual anchor (MVAM) is proposed to improve the positioning accuracy of multi-robots, track the motion trajectory of micro-robots in real time, and correct deviations under target occlusion or similar targets through edge recognition algorithms, so that the tracking error is controlled within a certain range. Secondly, it can adapt to complex backgrounds and dynamic scenes, and ensure real-time feedback of robot position information in microfluidic operations.

[0009] (5) A Micro-Resource Conflict Resolution Orchestrator (MCDO) is proposed to realize multi-robot cooperative navigation in a parallel cooperative field. In the conflict resolution layer, obstacle avoidance paths are planned for multiple microrobots through a two-layer search (high-level conflict detection and low-level path adjustment), generating a cluster of collision-free paths while satisfying the safety path constraints given by CAEP. In the resource orchestration layer, the coil current sequence in the multi-robot conflict-free command is optimized, and magnetic field resources are dynamically allocated to each robot. In the parallel cooperative field output layer, the coil array is driven to generate independent gradient magnetic fields to perform multi-robot transportation / hybrid tasks.

[0010] Compared with the prior art, the beneficial technical effects of the present invention are:

[0011] 1. In terms of path planning efficiency, DEN and CAPE, through collaborative optimization, have solved the problems of strong randomness, tortuousness and long length of paths generated by traditional algorithms (such as RRT-Connect), which are difficult to meet the requirements of path smoothness and efficiency for microrobots in medical scenarios (such as vascular navigation), as well as the lack of optimization of global path features, which leads to a large number of sharp turns in the path, increasing the difficulty of motion control and energy consumption of microrobots. This has effectively improved the path planning quality of magnetic microrobots in complex environments.

[0012] 2. In terms of disturbance rejection control accuracy, DFDS integrates three levels of control—feedforward decoupling, state observer disturbance observation, and PID feedback—through composite control. This avoids problems faced by magnetically driven microrobots, such as nonlinear dynamic coupling (e.g., mutual influence of magnetic forces on the x / y axes), unknown environmental disturbances (e.g., blood flow), and the difficulty of ensuring tracking accuracy with traditional PID control. Furthermore, it solves the problem of magnetic field-biological fluid coupling, improving tracking accuracy to the sub-millimeter level and effectively enhancing the navigation accuracy and robustness of magnetically controlled microrobots in complex environments.

[0013] 3. In terms of multi-robot collaborative capability, the parallel collaborative field is equipped with a 12×12 micro-coil array system (single-layer circular design). The current direction and intensity of each coil are controlled by the multi-robot conflict-free instructions output by the MCDO, realizing discrete point-to-point drive of the micro-robot. At the same time, the MCDO solves the collision problem of multiple robots in dynamic tasks by generating collision-free trajectories based on dual-layer search. In addition, traditional global magnetic fields are difficult to distinguish the same robots. The MCDO enables each robot to be addressed independently through local magnetic field arrays.

[0014] 4. In terms of environmental adaptability, MVAM integrates the Kernel Correlation Tracking (KCTA) algorithm and the Visual Edge Estimation (VEEA) algorithm to form a robust perception closed loop. Furthermore, MVAM's KCTA+VEEA dual-mode localization mechanism solves the problem of insufficient real-time tracking accuracy. Simultaneously, as a crucial component of the entire system's closed loop, MVAM can maintain high-precision pose feedback even in dynamic occlusion environments, supporting the system's real-time response to sudden environmental changes. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of the framework of the magnetic microrobot parallel cooperative field navigation system provided in this application. Detailed Implementation

[0016] The specific embodiments of the present invention will be further described below with reference to the accompanying drawings.

[0017] One embodiment of this application discloses a parallel collaborative field navigation system for magnetic microrobots. This system is built around the precise collaboration needs of multiple robots in complex environments. It uses a "planning-scheduling-driving-perception-feedback" closed-loop logic and designs a Dynamic Evolutionary Navigator (DEN), a Conflict-Aware Evolutionary Planner (CAEP), a Micro-Resource Conflict Dissolution Orchestrator (MCDO), a Dual-Field Disturbance Suppressor (DFDS), and a Micro-Vision Anchoring Module (MVAM). It encompasses an adaptive decision-making layer (corresponding to DEN and CAEP), a robust motion control layer (corresponding to DFDS), a multi-robot collaborative scheduling layer (corresponding to MCDO), and a real-time perception and feedback layer (corresponding to MVAM). This enables end-to-end control from path planning to physical driving, interference suppression, and state monitoring, supporting precision tasks such as medical vascular navigation and industrial micro / nano assembly.

[0018] Please refer to Figure 1 As shown, DEN, acting as the "first-generation designer" of single-robot paths, employs a hybrid sampling strategy (integrating "random sampling + target bias") to generate initial tree-like paths from the start point to the end point for each magnetic microrobot. It iteratively optimizes the path nodes, constraining path length (reducing it by 10%-20%) and smoothness (angle change ≤15°), outputting smooth single-robot paths and providing "raw materials" for feasible single-robot trajectories in multi-robot collaboration. CAEP focuses on multi-robot collaborative safety constraints, performing gene-level optimization on each smooth single-robot path through a multi-objective evolutionary strategy. It incorporates factors such as "distance between path points and obstacles" and "robot spacing" into the optimization objectives, generating safe path constraints with obstacle avoidance rules (e.g., robot spacing ≥15mm), providing a rule-based basis for subsequent conflict resolution. Figure 1 The input from DEN is fed into the MCDO in parallel to supplement the "safety guidelines" for multi-machine collaboration.

[0019] As the core hub of the system, the MCDO compares the optimized smooth single-robot path output by DEN with the safe path constraints of CAEP, and combines the pose data fed back by MVAM to plan obstacle avoidance paths for each magnetic microrobot and output multi-robot conflict-free commands, scheduling the coil array (12×12). The multi-robot conflict-free commands include the parameters of the acting coils, the direction and intensity of the coil current. Through the parallel collaborative field connecting the scheduling layer and the control layer, the digital commands are transformed into physical magnetic field "power." This manifests as: parsing the multi-robot conflict-free commands from the MCDO, controlling the Helmholtz coils (direction control) and Maxwell coils (position control) to work collaboratively, driving the coil array to generate a distributed gradient magnetic field, enabling independent drive of multiple robots and avoiding magnetic field cross-interference. The MCDO connects path planning and physical drive, realizing the transformation from "single-robot trajectory" to "multi-robot collaborative commands."

[0020] DFDS, based on coil current information in the command and pose data fed back from MVAM, achieves decoupled control of the magnetic field driving force (such as the cross-influence of x / y axis magnetic forces) through feedforward compensation and feedback correction. It outputs a coil current with disturbance rejection control to the MCDO, guiding the correction of conflict-free multi-machine commands and ensuring path tracking accuracy. DFDS is a "secondary calibration" after magnetic field driving, counteracting the influence of environmental interference on the trajectory.

[0021] MVAM uses the Kernel Correlation Tracking Algorithm (KCTA) to track the motion trajectory of the magnetic microrobot in real time and output pose data. KCTA achieves high-speed tracking at 30Hz, adapting to the robot's rapid movements. Additionally, MVAM incorporates the Visual Edge Extraction Algorithm (VEEA), which is triggered when there is target occlusion or interference from similar targets, correcting deviations through edge extraction. MVAM is crucial for system closed-loop control, providing "real-time pose data" for MCDO collision detection and DFDS disturbance rejection control, supporting dynamic system adjustments.

[0022] The following is a detailed explanation of the structure of each of the above modules:

[0023] (1) Dynamic Evolutionary Navigation Planner (DEN)

[0024] The core objective of DEN (Device Execution Environment) is to generate a single-machine initial path through rolling optimization, balancing path length and smoothness. First, the search space is initialized, with input parameters including the robot's starting coordinates S(x). s ,y s ,z s ), endpoint coordinates G(x) g ,y g ,z gThe obstacle 3D model (discrete into 144×144×144 voxels using a mesh method) is used, with constraints set: minimum path node spacing of 1mm and obstacle safety distance ≥1.5mm. Next, a hybrid sampling strategy is employed to generate the initial tree-like path, including: randomly sampling environmental space points with a certain probability (e.g., 60%) to avoid local optima, and biasing sampling towards the target point with a certain probability (e.g., 40%) to accelerate convergence; tree structure expansion: starting from the starting point, after each sampling of a new node, the feasibility of connection with the nearest tree node is checked (no collision), gradually expanding to the endpoint to form the initial tree-like path. Then, the iterative optimization process for the path nodes includes:

[0025] Generate a new population: Sample 100 individuals based on the mean vector μ and covariance matrix Σ, where each individual is a sequence of path nodes. Evaluate individual fitness: Calculate the path length f for each individual. l and angle change f a The formula is as follows: r k Let f be the k-th path node among the individuals. Select the top 20% of individuals with higher fitness values ​​(individual fitness sorted from high to low) as elite individuals, and update μ and Σ based on the elite individuals to adjust the search direction. When f has been continuous for five generations... l and f a When the variation range is less than the preset value (e.g., 5%), the individual corresponding to the highest fitness value in the current generation is output as the smooth single-machine path. Finally, rolling window optimization is performed, dividing the long path into windows such as 5mm. Each time, only the nodes of the current window and the next two windows are optimized, which can reduce the amount of computation and adapt to dynamic environments.

[0026] (2) Conflict-Aware Evolutionary Planner (CAEP)

[0027] The core objective of CAPE is to resolve multi-objective conflicts and generate Pareto optimal paths that balance obstacle avoidance and efficiency. The process of optimizing each smooth single-machine path at the gene level through a multi-objective evolutionary strategy includes:

[0028] Define a multi-objective function: The input consists of a smooth single-machine path generated by DEN, the obstacle distribution, the task type, and the path length f. l With angle change f a The definition is the same as DEN, the obstacle distance f d It is the reciprocal of the distance between the path node and the nearest obstacle, and the formula is: (∈=0.1mm, to avoid denominators of 0). Weights are automatically assigned based on task type; for example, in a medical scenario, ω... l =0.4, ω a =0.3, ω d =0.3 allows for priority obstacle avoidance. In industrial scenarios, ω l=0.6, ω a =0.2, ω d =0.2 can prioritize efficiency, and the final multi-objective function is the path length f. l , Angle change f a Obstacle distance f d The weighted sum is expressed as F = ω l f l +ω a f a +ω d f d .

[0029] Encode the node coordinates on the smooth single-machine path as real-valued genes, such as r. k (x k ,y k ,z k The system first generates 50 mutant individuals based on a smooth single-machine path. Then, it randomly selects two parent individuals and generates offspring individuals by crossover gene fragments with a first probability (e.g., 0.8). Gaussian noise (standard deviation 0.5mm) is added to each gene with a second probability (e.g., 0.1). Conflict detection is then performed, checking each offspring individual to ensure it meets the obstacle safety distance constraint (≥1.5mm). The fitness values ​​of all individuals are calculated according to the defined multi-objective function F, and the top 10 individuals with the lowest fitness values ​​(sorted from lowest to highest fitness) are selected as the next generation parents. After a predetermined number of iterations (e.g., 10), the individuals corresponding to the lowest fitness values ​​under different weights during the iteration process are collected as the optimal path, forming a Pareto solution set. This set is used by the system to select the optimal path based on the real-time environment, such as automatically switching to a higher ω when encountering sudden obstacles. d Weighted path.

[0030] In this embodiment, DEN and CAEP have a collaborative mechanism. The smooth single-machine path output by DEN serves as the initial population for CAEP's gene-level optimization. The optimized smooth single-machine path output by CAEP is fed back to DEN for rolling window optimization and then input to MCDO. When the rate of environmental change (e.g., obstacle movement distance > 2mm) or the probability of path conflict (collision risk > 30% in multi-robot scenarios) exceeds a set threshold, CAEP is automatically restarted to replan the smooth single-machine path. Each CAEP optimization only adjusts the curvature of the path by more than 1.5mm. -1 The inflection point node, because f a The smoothness of the path is reflected, which indirectly affects the selection and adjustment of curvature inflection points. For paths with inflection points of large curvature (i.e., large angle changes), its f... aThe value will be affected, which in turn will change the function F value. When selecting offspring individuals, individuals with smaller F values ​​will be given priority, that is, individuals with smoother paths (more reasonable curvature inflection points), thereby reducing the amount of computation (saving 40% of the time compared to global optimization).

[0031] (3) Dual-Field Disturbance Suppressor (DFDS)

[0032] The core of DFDS is disturbance rejection control. Firstly, it involves decoupling and feedforward control of the magnetic field drive. The input is the magnetic field drive command transmitted in parallel (including the Helmholtz coil current I). helm Maxwell coil current I maxw The decoupling calculation is based on the magnetic field-force model, analyzing the magnetic field-force coupling relationship along the x, y, and z axes (e.g., the x-axis current of a Helmholtz coil affects the y-axis magnetic field distribution), as shown in the following formula:

[0033]

[0034] Where k1 and k2 are the coil force coefficients, ΔF cross The cross-coupling force acting on the coil refers specifically to the undesirable force generated by the control current of a certain axis (such as the x-axis coil current) on other non-target axes (such as the y-axis). It can be obtained analytically through the multi-axis current-force coupling matrix, which will not be elaborated here.

[0035] The process of generating coil current with disturbance rejection control in this module includes: firstly, obtaining the pre-calculated coupling force compensation amount ΔI for feedforward compensation. comp =K -1 ΔF cross To correct the initial coil current of the MCDO output, the corrected initial coil current is I. new =I cmd +ΔI comp ;

[0036] Among them, I cmd The initial coil current acting on the Helmholtz coil or Maxwell coil is the "required value" of the coil current initially calculated by the system based on the target motion (such as the planned path or desired pose), and it serves as the basic input for subsequent current compensation and correction.

[0037]

[0038] Where, k xx k yy The coaxiality factor (the influence of x-axis current on x-axis force, and the same applies to y-axis) determines the desired force; k xy k yx The cross-coupling coefficient (the effect of x-axis current on y-axis force, and the same applies to y-axis) determines the cross-coupling force.

[0039] Secondly, define the robot's motion state as x = [p, v, a]. T The position, velocity, and acceleration of the magnetic microrobot, along with environmental disturbances d (fluid resistance, sensor noise), are used as extended states. An extended state observer is constructed, with the core formula being: Where A and B are state matrices, L and M are observation gains, and y is the pose data fed back by MVAM. The pose is predicted by the observer, and u is the equivalent control force for achieving magnetic field actuation. The compensation current is generated based on the real-time estimated environmental disturbance force output by the extended state observer. Specifically, this involves: using the observer to simultaneously estimate the "disturbance" and "state" from the pose measurement value y fed back by the MVAM, and then converting it into an environmental disturbance force. Observer output If the disturbance is equivalent to acceleration, then the real-time estimated environmental disturbance force can be obtained using Newton's second law as follows: m is the mass of the magnetic microrobot, which in turn generates a compensating current. The coil drive current then changes further to I. final =I new +ΔI d .

[0040] Then, feedback control and error correction are performed: the pose data p fed back by MVAM is calculated. real The pose p of the nearest node on the obstacle avoidance path planned by MCDO ref The error between them is e = p ref -p real The current is then fed into the PID controller for feedback adjustment, and the PID correction current is output. Further compensation for errors.

[0041] Finally, by integrating the corrected initial coil current, disturbance compensation current, and PID correction current obtained based on feedforward compensation, a coil current I with disturbance rejection control is obtained. drive =I final +ΔI pid .

[0042] (4) Micro-visual anchor (MVAM)

[0043] The core objective of MVAM is to provide "real-time pose data" to support dynamic system adjustments. The real-time tracking of the magnetic microrobot's trajectory using a kernel-correlation tracking algorithm includes: first, acquiring real-time images of the robot's motion using a high-resolution camera (1920×1080) and performing image preprocessing operations, such as adaptive threshold binarization to denoise the images, enhance contrast, and highlight the robot's contours. Second, extracting image features to construct an appearance model f(z), and calculating the target response in each frame of the real-time tracked image through cyclic cross-correlation. To quickly locate the center coordinates (x) of the magnetic microrobot kcf ,y kcf The output pose data includes the center coordinates and attitude angles of the magnetic microrobot.

[0044] in, For the kernel correlation matrix, The label is the Fourier domain, ⊙ represents pointwise multiplication, and F... -1 {} represents the inverse Fast Fourier Transform operator. In one possible implementation, the target response is calculated. The process includes: (1) Obtaining the kernel correlation matrix: Extracting HOG (Histogram of Oriented Gradients) features from the target region (such as a micro-robot) in the initial frame, denoted as feature vector x, and extracting candidate features of the same dimension from the candidate region (sliding window around the predicted position of the target) in subsequent frames, denoted as feature vector y. The feature dimension is usually d (e.g., the HOG feature dimension of 64×64 pixels is about 2000). Then, a kernel function is used to map the HOG features and candidate features to a high-dimensional space. In this example, the Gaussian kernel function is selected: Where σ is the kernel width parameter (controlling the sensitivity of the similarity measure). Finally, the HOG feature x is compared with all candidate features y. i (i = 1, 2, ..., N, where N is the number of candidate regions) Calculate the kernel function values ​​to construct an N×1 kernel correlation matrix. If a global similarity matrix is ​​required, then calculate the kernel function pairwise for all candidate features to obtain an N×N kernel correlation matrix.

[0045] (2) Obtain Fourier domain labels: In the initial frame, define a two-dimensional Gaussian function as the ideal response label with the target center as the origin:

[0046]

[0047] Where (u,v) are the coordinates relative to the target center, σ y The Gaussian width (controlling peak sharpness, typically set to 1 / 2 of the target radius) is used. The physical meaning of this label is: ideally, the matching response is highest at the target center, and attenuates when deviating from the center. Finally, a two-dimensional Fourier transform is performed on the ideal response label y(u,v) to obtain the Fourier domain label. F{} is the Fast Fourier Transform operator.

[0048] In this embodiment, when the target responds When the peak value is below a set threshold, the Visual Edge Extraction Algorithm (VEEA) is triggered to correct for target occlusion or deviations under similar targets. Specifically, this includes the following: First, the real-time acquired robot motion images are sequentially processed using image grayscale conversion, Gaussian filtering for noise reduction, and adaptive histogram equalization (CLAHE) to enhance image contrast. Then, edge detection is performed to extract the "skeleton" of the target contour, including: using the Canny operator to extract image edges; calculating the gradients in the x and y directions, and the magnitude G of the gradients. mag With direction; Non-maximum suppression: Preserving local maxima G along the gradient direction mag,max Remove non-edge pixels around the edge to refine the edge from a "wide strip" to a "single pixel line"; dual threshold filtering: set a high threshold T. h and low threshold T l (e.g. T) h =0.2G mag,max T l =0.1T h This is used to delineate the extracted edges, i.e., those higher than T. h The edges are "strong edges" (preserved directly), while those below T... l The edges are designated as "weak edges" (directly removed), while edges in between are designated as "candidate edges." If an edge is connected to a strong edge, it is retained to ensure edge continuity and prevent breakage of the microrobot's contour.

[0049] (5) Micro-resource conflict resolution orchestrator (MCDO)

[0050] The core objective of MCDO is to dynamically allocate magnetic field resources and conflict-free paths. This module includes a conflict resolution layer, a resource orchestration layer, and a parallel cooperative field output layer. In the conflict resolution layer, a two-layer search algorithm is employed. The higher layer identifies temporal-spatial conflict points (a) between multiple robots based on the input DEN rolling-optimized smooth single-machine path, CAEP output safety path constraints (such as robot spacing ≥ 15mm), and MVAM feedback pose data. i ,a j (p,t)(represents robot a) i With a j At time t, position p is occupied, and a conflict resolution strategy (such as "wait, detour") is generated. The lower layer alleviates the conflict by adjusting the path nodes or inserting the conflict resolution strategy, generates a collision-free path cluster that meets the safety path constraints, and converts it into a multi-machine conflict-free instruction to drive the coil array.

[0051] At the resource orchestration layer, the coil current sequence in the multi-robot conflict-free command is optimized to reduce system power consumption. Based on the optimized coil current sequence, the magnetic field resources of the 12×12 coil array are dynamically allocated to each robot (e.g., prioritizing activation of coils within 3mm of the robot's center). This drives the coil array to generate a distributed gradient magnetic field at the parallel cooperative field output layer, enabling the execution of multi-robot transport / hybrid tasks. Simultaneously, the resource orchestration layer also receives the correction current I from the DFDS feedback. drive This ensures successful resolution of path conflicts in multi-machine collaborative tasks and maximizes the efficiency of magnetic field drive.

[0052] In one possible implementation, the least squares method is used to optimize the coil current sequence in a multi-machine conflict-free command. The goal of the least squares method is to find an optimal current sequence I such that the calculated magnetic field H·I is as close as possible to the target magnetic field B, while minimizing the sum of squares of the errors. Specifically, this method includes: first, constructing a least squares problem: transforming the magnetic field equation into an optimization problem, with the residual vector (BH·I), and the objective being to minimize its sum of squares (BH·I). T (BH·I). Next, solve the normal equation: by taking the derivative and setting it to zero, we obtain the normal equation: H T H·I=H T The normal equations transform the original problem into a linear system. Finally, the optimal current is calculated: the optimal current sequence is I = (H... T H) -1 H T Here, H is the coil-magnetic field mapping matrix, which can be obtained in advance through calibration or modeling, such as obtaining matrix H through finite element simulation in Ansys Maxwell.

[0053] The above descriptions are merely preferred embodiments of this application, and the present invention is not limited to the above embodiments. It is understood that other improvements and variations directly derived or conceived by those skilled in the art without departing from the spirit and concept of the present invention should be considered to be included within the protection scope of the present invention.

Claims

1. A parallel cooperative field navigation system for magnetic microrobots, characterized in that, The system includes: The Dynamic Evolutionary Navigation Planner (DEN) generates an initial tree-like path from the starting point to the end point for each magnetic microrobot using a hybrid sampling strategy, and iteratively optimizes the path nodes to output a smooth single-robot path. The Conflict-Aware Evolutionary Planner (CAEP) performs gene-level optimization on each smooth single-machine path through a multi-objective evolutionary strategy and generates safe path constraints with obstacle avoidance rules, providing a rule basis for subsequent conflict resolution. The Micro-Resource Conflict Resolution Orchestrator (MCDO) compares the optimized smooth single-robot path with the safety path constraints, and combines the pose data fed back by the Micro-Vision Anchor (MVAM) to plan obstacle avoidance paths for each magnetic microrobot and output multi-robot conflict-free instructions. The multi-robot conflict-free instructions include the parameters of the operating coil, the direction and intensity of the coil current, which are used to drive the coil array to generate a distributed gradient magnetic field, enabling independent driving of multiple robots. The dual-field disturbance suppressor (DFDS) uses the coil current information in the command and the pose data fed back by the MVAM to achieve decoupling control of the magnetic field driving force through feedforward compensation and feedback correction. It also outputs a coil current with disturbance rejection control to the MCDO to guide the correction of the multi-machine conflict-free command. The micro-visual anchor MVAM tracks the motion trajectory of the magnetic microrobot in real time using a kernel-function correlation tracking algorithm and outputs the pose data, including: Real-time acquisition of robot motion images and image preprocessing operations; Extract features from the image to construct an appearance model In each frame of the real-time tracked image, the target response is calculated through cyclic cross-correlation. To locate the center coordinates of the magnetic microrobot, output pose data, including the center coordinates and attitude angles of the magnetic microrobot; in, , For the kernel correlation matrix, For Fourier domain labels, To multiply point by point, For the inverse fast Fourier transform operator; In this system, the DEN and the CAEP have a collaborative mechanism. The smooth single-machine path output by the DEN serves as the initial population for the CAEP when performing gene-level optimization. The optimized smooth single-machine path output by the CAEP is fed back to the DEN for rolling window optimization and then input to the MCDO. When the rate of environmental change or the probability of path conflict exceeds a set threshold, the CAEP is automatically restarted to replan the smooth single-machine path.

2. The parallel cooperative field navigation system for magnetic microrobots according to claim 1, characterized in that, The iterative optimization of path nodes includes: A new population is generated, and n1 individuals are sampled based on the mean vector and covariance matrix, where each individual is a sequence of path nodes; Calculate the path length for each individual and angle change The mean vector and covariance matrix are updated based on the fitness values ​​of individuals, and the top n2 individuals with higher fitness values ​​are selected as elite individuals. The search direction is adjusted by updating the mean vector and covariance matrix based on the elite individuals. When m consecutive generations and When the variation range is less than the preset value, the individual corresponding to the highest fitness value in the current generation is output as the smooth single-machine path.

3. The parallel cooperative field navigation system for magnetic microrobots according to claim 1, characterized in that, The step of performing gene-level optimization on each of the smooth single-machine paths using a multi-objective evolutionary strategy includes: The node coordinates on the smooth single-machine path are encoded as real-number genes, and n3 mutant individuals are randomly generated based on the smooth single-machine path, where each individual is a path node sequence; Two parent individuals are randomly selected, and offspring individuals are generated by cross-crossing gene fragments with the first probability. Gaussian noise is added to each gene with the second probability. The fitness values ​​of all individuals are calculated according to the defined multi-objective function, and the top n3 individuals with the lowest fitness values ​​are selected as the next generation parents; the multi-objective function is the path length. Angle changes Obstacle distance The weighted sum is automatically assigned based on the task type. After reaching the predetermined number of iterations, the individuals corresponding to the lowest fitness values ​​under different weights during the iteration process are collected as the optimal paths, forming a Pareto solution set, which is then used by the system to select based on the real-time environment.

4. The parallel cooperative field navigation system for magnetic microrobots according to claim 1, characterized in that, The process of generating a coil current with disturbance rejection control includes: Obtain the pre-calculated coupling force compensation amount for feedforward compensation This is to correct the initial coil current output by the MCDO; robot motion state An extended state observer is constructed with the environmental disturbance d as an extended state, and a compensation current is generated based on the real-time estimated environmental disturbance force output by the extended state observer. The error between the pose data fed back by the MVAM and the pose of the nearest node on the obstacle avoidance path planned by the MCDO is calculated and input to the PID controller for feedback adjustment, and the PID correction current is output. By integrating the corrected initial coil current, the compensation current, and the PID correction current, the coil current with disturbance rejection control is obtained. in, These are the position, velocity, and acceleration of the magnetic microrobot, respectively, and the environmental disturbance d is fluid resistance and sensor noise. This refers to the cross-coupling force acting on the coil; in, For coaxiality, This represents the cross-coupling coefficient.

5. The parallel cooperative field navigation system for magnetic microrobots according to claim 4, characterized in that, The generation of compensation current based on the real-time estimated environmental disturbance force output by the extended state observer includes: The extended state observer is defined as follows: Where A and B are state matrices, L and M are observation gains, and y is the pose data fed back by the MVAM. It is the pose predicted by the observer. It is the equivalent control force that enables magnetic field-driven operation; The real-time estimated environmental disturbance force output by the extended state observer is: , The mass of the magnetic microrobot is ; the generated compensation current is . .

6. The parallel cooperative field navigation system for magnetic microrobots according to claim 1, characterized in that, The MCDO comprises a conflict resolution layer, a resource orchestration layer, and a parallel cooperative field output layer, wherein: A two-layer search algorithm is used in the conflict resolution layer. The upper layer identifies the time-space conflict points between multiple machines and generates a resolution strategy based on the input optimized smooth single-machine path, the safe path constraint and the pose data. The lower layer alleviates the conflict by adjusting the path nodes or inserting the resolution strategy, generates a cluster of collision-free paths that meet the safe path constraint, and converts them into multi-machine conflict-free instructions for driving the coil array. In the resource orchestration layer, the coil current sequence in the multi-robot conflict-free instruction is optimized, and the coil array magnetic field resources are dynamically allocated to each robot according to the optimized coil current sequence. In this way, in the parallel cooperative field output layer, the coil array is driven to generate a distributed gradient magnetic field to perform multi-robot transportation / hybrid tasks.

7. The parallel cooperative field navigation system for magnetic microrobots according to claim 1, characterized in that, The calculation target response include: HOG features are extracted from the target region in the initial frame, and candidate features of the same dimension are extracted from the candidate regions in subsequent frames. A kernel function is used to map the HOG features and the candidate features to a high-dimensional space, and kernel function values ​​are calculated for the HOG features and all candidate features to construct a kernel correlation matrix. ; In the initial frame, a two-dimensional Gaussian function is defined as the ideal response label with the target center as the origin: For the ideal response label Perform a two-dimensional Fourier transform to obtain the Fourier domain labels. ; in,( u , v () represents the coordinates relative to the center of the target. σ y The width is Gaussian. This is the Fast Fourier Transform operator.

8. The parallel cooperative field navigation system for magnetic microrobots according to claim 1, characterized in that, In the MVAM, when the target response When the peak value is lower than the set threshold, the visual edge extraction algorithm is triggered to correct the deviation under target occlusion or similar targets.

Images