Multi-source fusion localization and mapping method and system for spraying robots under strong interference conditions

By integrating multi-source data and using adaptive control, the problem of low positioning accuracy of the spraying robot under strong interference conditions was solved, and a stable spraying effect was achieved in environments with high concentrations of paint mist and metallic reflection.

CN123103706APending Publication Date: 2026-11-17SHANGHAI COSCO SHIPPING HEAVY IND CO LTD +2
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611615707.2
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-19
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing mobile spraying robots struggle to effectively separate interference features in spraying environments with high concentrations of paint mist and reflective metal surfaces, leading to reduced positioning accuracy and fluctuations in the normal distance between the spray gun tip and the workpiece surface during the spraying process. Existing control systems cannot balance computational load and dynamic response.

Method used

Data is collected by synchronously triggering the LiDAR, vision module, and inertial measurement unit in hardware to construct real-time environmental geometric signals. Frequency domain decoupling is performed using the high-frequency kinematic characteristics and orthogonal subspace projection of the inertial measurement unit to purify the actual terrain topology space matrix. Reference velocity and attitude feedforward vectors are generated by combining rolling time domain optimization, and composite control commands are generated in conjunction. The collaborative compensation weights of the chassis and robotic arm are adaptively allocated to achieve closed-loop control.

Benefits of technology

The positioning accuracy of the painting robot was improved under strong interference conditions, the position fluctuation of the painting trajectory was reduced, and the stability and accuracy of the painting process were ensured.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123103706A_ABST
    Figure CN123103706A_ABST
Patent Text Reader

Abstract

This invention relates to the field of robot position control technology, specifically disclosing a multi-source fusion positioning and mapping method and system for a painting robot under strong interference conditions. The method is applied to a position adjustment system for a mobile painting robot integrating a lidar, vision module, inertial measurement unit, and including a mobile chassis, robotic arm, and spray gun. It includes synchronously triggering the lidar, vision module, and inertial measurement unit to collect data from the painting site via hardware, aligning timestamps, and constructing a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamps, and source identifiers. This invention synchronously collects multi-source data via hardware, utilizes the kinematic characteristics of the inertial measurement unit in conjunction with orthogonal subspace projection, and separates the true workpiece topology during the front-end data processing stage, reducing the impact of false features caused by paint mist and reflections on positioning and mapping.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of robot position control technology, and in particular to a multi-source fusion positioning and mapping method and system for a painting robot under strong interference conditions. Background Technology

[0002] Existing mobile painting robots typically integrate LiDAR, vision modules, and inertial measurement units to fuse multi-source data, enabling autonomous localization and mapping. A position control system guides the chassis and robotic arm along a preset trajectory. In a typical operating environment, the control system sequentially processes the perceived data to extract environmental features, uses a fixed predictive model to calculate control commands, and then issues independent motion compensation signals to the chassis and robotic arm to adjust the robot's overall pose and the trajectory of the end-effector spray gun.

[0003] In a spraying site containing high concentrations of paint mist and metal surface reflection, the scattering and reflection of paint mist will produce spatial geometric interference characteristics similar to those of the actual workpiece surface. Existing control systems have difficulty effectively separating these interference characteristics, which can easily lead to distortion of the map structure of the input control loop.

[0004] Meanwhile, when dealing with disturbed sensing data, the existing position adjustment mechanism has difficulty balancing computational load and dynamic response in a fixed control time domain. Furthermore, the independent control modes of the chassis and robotic arm cannot adaptively allocate weights based on real-time position residuals, which can easily cause fluctuations in the normal distance between the spray gun tip and the workpiece surface during the spraying process, leading to deviations in the overall position adjustment. Summary of the Invention

[0005] This invention aims to at least partially solve one of the technical problems in related technologies. Therefore, the objective of this invention is to propose a multi-source fusion localization and mapping method and system for spraying robots under strong interference conditions, in order to improve positioning accuracy.

[0006] To achieve the above objectives, a first aspect of the present invention proposes a multi-source fusion localization and mapping method for a painting robot under strong interference conditions. This method is applied to a position adjustment system for a mobile painting robot integrating a lidar, vision module, inertial measurement unit, and comprising a mobile chassis, robotic arm, and spray gun, and includes:

[0007] Step 1: The lidar, vision module and inertial measurement unit are synchronously triggered by hardware to collect spraying site data, and after aligning the timestamps, a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamp and source identifier is constructed.

[0008] Step 2: Construct an adaptive spatial geometry filtering operator using the high-frequency kinematic features of the inertial measurement unit, and combine it with orthogonal subspace projection to decouple the real-time environmental geometric signal in the frequency domain, thereby refining the actual terrain topology space matrix that represents the real workpiece boundary and robot pose.

[0009] Step 3: Inject the actual terrain topology space matrix into the trajectory control loop, extract the workpiece surface morphology and normal constraints, dynamically adjust the prediction time domain step size and control time domain step size according to the environmental signal-to-noise ratio, and generate the reference velocity feedforward vector and reference attitude feedforward vector through rolling time domain optimization.

[0010] Step 4: Combine the reference velocity feedforward vector, the reference attitude feedforward vector, and the current pose tracking residual to generate a composite control command. Adaptively allocate the collaborative compensation weights of the moving chassis and the robotic arm according to the amplitude of the current pose tracking residual, and drive each servo motor to complete the collaborative position adjustment of the spraying trajectory.

[0011] Step 5: Based on trajectory tracking error, control increment and spraying distance fluctuation, construct a comprehensive performance index, and realize closed-loop control of position adjustment by updating the decoupling filter parameters and the collaborative compensation weights online.

[0012] To achieve the above objectives, a second aspect of the present invention proposes a multi-source fusion localization and mapping system for a painting robot under strong interference conditions. This system is applied to a mobile painting robot integrating a lidar, a vision module, an inertial measurement unit, and comprising a mobile chassis, a robotic arm, and a spray gun. The system includes:

[0013] The real-time environmental geometric signal construction module is used to synchronously trigger the lidar, the vision module and the inertial measurement unit to collect spraying site data through hardware, and construct a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamp and source identifier after aligning the timestamps.

[0014] The actual terrain topology space matrix purification module is used to construct an adaptive spatial geometry filtering operator using the high-frequency kinematic features of the inertial measurement unit, and combine orthogonal subspace projection to decouple the real-time environmental geometric signal in the frequency domain, thereby purifying the actual terrain topology space matrix that represents the real workpiece boundary and robot pose.

[0015] The rolling time-domain optimization feedforward generation module is used to inject the actual terrain topology space matrix into the trajectory control loop, extract the workpiece surface morphology and normal constraints, dynamically adjust the prediction time-domain step size and control time-domain step size according to the environmental signal-to-noise ratio, and generate reference velocity feedforward vector and reference attitude feedforward vector through rolling time-domain optimization.

[0016] The collaborative position adjustment control module is used to generate a composite control command by combining the reference velocity feedforward vector, the reference attitude feedforward vector and the current pose tracking residual. Based on the amplitude of the current pose tracking residual, it adaptively allocates the collaborative compensation weights of the mobile chassis and the robotic arm, and drives each servo motor to complete the collaborative position adjustment of the spraying trajectory.

[0017] The closed-loop control adaptive iterative module is used to construct a comprehensive performance index based on trajectory tracking error, control increment and spraying distance fluctuation, and realize closed-loop control of position adjustment by updating the decoupling filtering parameters and the collaborative compensation weights online.

[0018] To achieve the above objectives, a third aspect of the present invention provides an electronic device, including a memory, a processor, and a computer program stored in the memory. When the computer program is executed by the processor, it implements the above-described method for multi-source fusion localization and mapping of a painting robot under strong interference conditions.

[0019] This invention synchronously acquires multi-source data through hardware and utilizes the kinematic characteristics of the inertial measurement unit in conjunction with orthogonal subspace projection to separate the true workpiece topology during the front-end data processing stage, reducing the impact of false features caused by paint mist and reflection on positioning and mapping. In the control stage, the system injects the purified topological space matrix into the trajectory control loop, extracts the actual normal constraints, and dynamically adjusts the prediction time step in combination with the environmental signal-to-noise ratio, improving the environmental adaptability of the control command generation. At the same time, it adaptively allocates the compensation weights of the chassis and robotic arm according to the amplitude of the real-time pose tracking residual, enabling the underlying actuators to perform coordinated position corrections. Through closed-loop iterative updates of control and decoupling parameters, the mobile painting robot can maintain the set running distance and trajectory position under disturbed working conditions. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating the multi-source fusion localization and mapping method for spraying robots under strong interference conditions provided by the present invention.

[0021] Figure 2 This is a schematic diagram illustrating the implementation of the multi-source fusion positioning and mapping system for spraying robots under strong interference conditions provided by the present invention.

[0022] Figure 3 This is a schematic diagram of the structure of the electronic device provided by the present invention. Detailed Implementation

[0023] Embodiments of the present invention are described in detail below, examples of which are illustrated in the accompanying drawings, wherein the same or similar reference numerals denote the same or similar elements or elements having the same or similar functions throughout. The embodiments described below with reference to the accompanying drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0024] The following description, with reference to the accompanying drawings, outlines a multi-source fusion positioning and mapping system, method, and electronic device for a painting robot under strong interference conditions, according to embodiments of the present invention.

[0025] Example 1:

[0026] like Figure 1 As shown, this embodiment provides a multi-source fusion localization and mapping method for a painting robot under strong interference conditions. Specifically, the method of this embodiment is applied to a position adjustment system of a mobile painting robot that integrates LiDAR, a vision module, an inertial measurement unit, and includes a mobile chassis, a robotic arm, and a spray gun.

[0027] The mobile chassis is a tracked chassis driven independently by four AC servo motors; the robotic arm is a serial industrial robotic arm with six degrees of freedom, each joint equipped with an absolute high-precision encoder and a harmonic reducer; the spray gun is fixed to the end flange of the robotic arm. The computing platform of the positioning and mapping system adopts a heterogeneous edge computing node based on the ARM Cortex-A architecture, including an eight-core central processing unit with a main frequency of two gigahertz and a neural network acceleration engine with a computing power of twenty trillion operations per second, and equipped with a low-power double data rate synchronous dynamic random access memory with a capacity of sixteen gigabytes.

[0028] To meet the response requirements of safety-critical scenarios, the underlying operating system kernel of the control system adopts a preemptive real-time Linux operating system with hard real-time characteristics. Based on the actual computing power limit of this computing platform, the control cycle of global position adjustment in this embodiment is strictly set to 10 milliseconds, that is, the system completes the acquisition, decoupling, optimization and command issuance of all data at a control frequency of 100 Hz.

[0029] Specifically, the multi-source fusion localization and mapping method for the painting robot under strong interference conditions includes the following steps:

[0030] Step 1: The lidar, vision module and inertial measurement unit are synchronously triggered by hardware to collect data on the spraying site. After aligning the timestamps, a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamp and source identifier is constructed.

[0031] Specifically, in step 1, constructing the real-time environmental geometric signal includes the following sub-steps: obtaining the original three-dimensional point cloud set output by the lidar, and extracting the three-dimensional coordinates of the spatial sampling points under the system and the corresponding laser reflection intensity values.

[0032] For example, the lidar is a 16-line industrial-grade 3D lidar, with its built-in rotary motor speed set to 10 Hz. The original 3D point cloud set output in a single scan contains approximately 30,000 spatial sampling points. The carrier system is a 3D Cartesian coordinate system established with the geometric center of the robot's mobile chassis as the origin. The 3D coordinates are stored using 32-bit single-precision floating-point numbers, and the laser reflection intensity value is stored using 16-bit unsigned integers to reflect the material reflection characteristics of the workpiece surface.

[0033] The original visual feature point set output by the vision module is obtained, and the pixel coordinates and corresponding grayscale gradient magnitudes of the feature points are extracted. For example, the vision module uses a global shutter binocular industrial camera with a frame rate set to 30 Hz. Corner points and edge points in the image are extracted using an edge detection algorithm to form the original visual feature point set. The grayscale gradient magnitude represents the degree of brightness change at that pixel coordinate.

[0034] Obtain the original measurement vector output by the inertial measurement unit, and extract the three-dimensional specific force vector and three-dimensional angular velocity vector under the load system.

[0035] For example, the inertial measurement unit employs a six-axis microelectromechanical system-level sensor, internally integrating a three-axis accelerometer and a three-axis gyroscope, with its data output frequency set to 200 Hz. The three-dimensional specific force vector represents the non-gravitational acceleration component measured by the accelerometer, and the three-dimensional angular velocity vector represents the rotational angular rate of the three axes of the load-bearing system measured by the gyroscope.

[0036] Based on a preset sampling interval, the three-dimensional coordinates and the pixel coordinates at the same sampling time are integrated into the spatial coordinates, the laser reflection intensity value and the grayscale gradient amplitude are integrated into the feature intensity, and the corresponding timestamp and source identifier are added to generate the real-time environmental geometric signal in matrix form.

[0037] It should be noted that, in order to eliminate asynchronous errors caused by inconsistent sampling frequencies of different sensors, this embodiment is configured with an independent field-programmable gate array chip as a clock synchronization generator.

[0038] The generator sends pulse trigger signals with a globally precise time protocol to all sensors at a preset sampling interval of 10 milliseconds (corresponding to a 100 Hz control cycle). During the data alignment phase, each preset sampling interval is used as a data frame window, and a linear interpolation algorithm is used to resample and project the 10 Hz laser data and 30 Hz visual data onto a 100 Hz time domain reference, so that all data have completely consistent timestamps.

[0039] The final generated real-time environmental geometric signal is presented in memory as a two-dimensional feature matrix. Each row of the matrix corresponds to a feature point, and the column vectors store the aforementioned spatial coordinates, feature intensity, globally unified timestamp, and sensor identifier indicating the data source (e.g., using zero to represent LiDAR and one to represent a vision module). This processing method ensures absolute alignment of multi-source data in the physical time dimension before entering subsequent algorithms.

[0040] Step 2: Construct an adaptive spatial geometry filtering operator using the high-frequency kinematic characteristics of the inertial measurement unit, and combine it with orthogonal subspace projection to decouple the real-time environmental geometry signal in the frequency domain, thereby refining the actual terrain topology space matrix that represents the real workpiece boundary and robot pose.

[0041] Specifically, in step 2, before constructing the adaptive spatial geometry filtering operator using the high-frequency kinematic characteristics of the inertial measurement unit, the following steps are also included:

[0042] A preset navigation coordinate system is established. The attitude rotation matrix from the previous moment is integrated with the current moment's angular velocity increment after zero-bias compensation to obtain the attitude rotation matrix at the current moment. For example, the preset navigation coordinate system uses a northeast-northeast coordinate system, i.e., the X-axis points due east, the Y-axis points due north, and the Z-axis points upwards, representing the geographic coordinate system. In practical engineering implementation, the gyroscope's measurement output includes zero-bias errors that drift with ambient temperature and operating time. These errors must be subtracted online from the pre-calibrated angular velocity zero-bias value before multiplying by the preset sampling interval to obtain the angular velocity increment. This increment is then converted into an attitude update matrix using the Lie algebra exponential mapping method to ensure the numerical orthogonality of the integration process.

[0043] The three-dimensional specific force vector after acceleration zero bias compensation is transformed to the preset navigation coordinate system using the attitude rotation matrix at the current moment, and the constant vector of gravitational acceleration is subtracted to obtain the linear acceleration in the preset navigation coordinate system at the current moment.

[0044] In this calculation process, the value of the constant vector of gravitational acceleration is initialized based on the actual gravity constant of the local area where the equipment is deployed. This is intended to eliminate the physical interference of the gravitational component on the pure kinematic acceleration.

[0045] By integrating and recursively applying the linear acceleration, a high-frequency kinematic state containing the current position and linear velocity of the painting robot is obtained.

[0046] For example, the linear acceleration is discretely integrated in the time domain using the first-order Euler numerical integration method. The linear velocity at the current moment is equal to the linear velocity at the previous moment plus the product of the linear acceleration at the current moment and the preset sampling interval.

[0047] Similarly, the current position is recursively obtained by integrating the previous position and the current linear velocity. This process has extremely low computation time, fully leveraging the advantages of the inertial measurement unit (IMU) for short-time high precision.

[0048] Specifically, in step 2, frequency domain decoupling of the real-time environmental geometric signal is performed by combining orthogonal subspace projection, which specifically includes:

[0049] By using the current attitude rotation matrix and the current position, the spatial coordinates in the real-time environmental geometric signal are transformed into the preset navigation coordinate system, resulting in a navigation system spatial coordinate matrix. This spatial transformation unifies the dynamically changing coordinates mounted on the robot body into an absolutely stationary reference system, laying a physical foundation for eliminating free-floating interference features such as paint mist particles.

[0050] Two strictly orthogonal feature subspaces are obtained by pre-training with typical working condition samples. The feature subspaces include a real terrain feature subspace representing a fixed structure and a noise feature subspace representing interference features.

[0051] To comply with the requirement of sufficient disclosure regarding the construction and training of algorithmic models, the training and acquisition methods for these two sets of feature subspaces are described in detail below:

[0052] First, training data was collected, and the painting robot was driven to operate in a cleanroom environment without paint mist spraying. The set of multi-source spatial coordinates for 1000 control cycles was recorded as noise-free samples. Then, it was operated under strong interference conditions with high-pressure spraying and paint mist concentration reaching 80 mg / m³. The set of coordinates for 1000 control cycles was recorded as noisy samples.

[0053] The two sets of samples are aligned and recombined to construct a spatial covariance matrix. On the offline server, the covariance matrix is ​​decomposed using the singular value decomposition algorithm to extract its eigenvectors and singular values. According to the principle of energy conservation, the singular values ​​are arranged in descending order.

[0054] Because real workpieces such as ship bulkheads and large steel structures have high continuity and structural rigidity in spatial distribution, their corresponding spatial variance dominates the overall energy distribution; while paint mist particles and electromagnetic noise points exhibit an isotropic random distribution in space, and their corresponding variance energy is small and discrete.

[0055] Therefore, the eigenvectors corresponding to the first three largest singular values ​​are extracted to form a basis vector set, and the space spanned by this basis vector set is defined as the true terrain feature subspace. The basis vector set formed by the eigenvectors corresponding to the remaining singular values ​​is defined as the noise feature subspace. Due to the mathematical properties of singular value decomposition, these two subspaces are strictly orthogonal in geometric properties.

[0056] Through the training methods described above, the system can solidify mathematical models that characterize the structural invariance of the real physical world in the offline stage.

[0057] The navigation system spatial coordinate matrix is ​​orthogonally projected onto the real terrain feature subspace, and the components corresponding to the noise feature subspace are removed to obtain the purified real terrain coordinate matrix. The calculation formula for this orthogonal projection is as follows:

[0058] ;

[0059] in, This represents the purified real terrain coordinate matrix; The basis matrix representing the real terrain feature subspace; The transpose of the basis matrix of the real terrain feature subspace; This represents the spatial coordinate matrix of the navigation system.

[0060] The algorithm involves only basic matrix multiplication calculations, and a single execution can take less than 0.2 milliseconds in the heterogeneous edge computing nodes, which is strictly matched to the underlying hardware processing capabilities.

[0061] The purified real terrain coordinate matrix is ​​converted back to the carrier system, and combined with the corresponding feature intensity, timestamp and source identifier, the actual terrain topology space matrix at the current moment is constructed.

[0062] Optionally, the adaptive spatial geometry filtering operator is a spatial transformation matrix constructed based on a third-order identity matrix, the linear velocity in the preset navigation coordinate system, and adaptive constraint coefficients, wherein the adaptive constraint coefficients are positively correlated with the environmental signal-to-noise ratio.

[0063] After obtaining the actual terrain topology space matrix, the motion consistency is verified using the adaptive spatial geometry filtering operator, specifically as follows:

[0064] Calculate the deviation distance between the coordinate vectors of each feature point in the actual terrain topology space matrix and the adaptive spatial geometry filtering operator under spatial mapping. The formula for calculating this deviation distance is as follows:

[0065] ;

[0066] in, This represents the deviation distance of each feature point under this mapping; Indicates the adaptive constraint coefficient; This represents the linear velocity vector in the preset navigation coordinate system; This represents the transpose of the linear velocity vector; The squared norm of the linear velocity vector; This represents the coordinate vector of any extracted feature point.

[0067] If the deviation distance is less than the preset distance verification threshold, the corresponding feature point is retained; otherwise, it is determined to be an outlier that does not conform to the rigid body motion law and is discarded. The preset distance verification threshold is set by the median of the ranging error of the LiDAR at the factory calibration, avoiding the subjectivity of manual setting.

[0068] The algorithm's behavior design under extreme conditions reflects strict physical laws: in a real environment, the workpiece is stationary, and the motion trend of its coordinates relative to the mobile robot should be absolutely consistent with the robot's reverse linear velocity. If the absolute coordinates of a point in space (such as paint mist particles drifting in the wind) change, that point will have a huge deviation in the above transformation formula.

[0069] By verifying this formula, the system effectively eliminates dynamic false features from interfering with map data based on physical motion attributes.

[0070] Step 3: Inject the actual terrain topology space matrix into the trajectory control loop, extract the workpiece surface morphology and normal constraints, dynamically adjust the prediction time domain step size and control time domain step size according to the environmental signal-to-noise ratio, and generate the reference velocity feedforward vector and reference attitude feedforward vector through rolling time domain optimization.

[0071] Specifically, in step 3, before extracting the workpiece surface morphology and normal constraints, a unified kinematic state-space model of the painting robot is also established, specifically as follows:

[0072] The three-dimensional position vector and three-dimensional velocity vector of the center point of the spray gun under the load system, as well as the angle vectors of each joint of the robotic arm, are extracted and combined to form the system state vector. The three-dimensional velocity command vector of the mobile chassis and the angular velocity command vectors of each joint of the robotic arm are extracted and combined to form the system control input vector. The kinematic models of the chassis and the robotic arm are combined to establish a discrete state equation relating the system state vector and the system control input vector.

[0073] For example, considering the sliding steering kinematics of a tracked chassis and the improved standard DiNavier-Hatenberg parameter matrix of a robotic arm, the angular velocity in joint space is mapped to the linear velocity in the end-effector space by calculating the Jacobian matrix, thus obtaining the global discrete state transition equations of the system. The rigor of this state-space model construction ensures that the displacement mapping of the control actions to the physical entity during subsequent optimization is realistic and reliable.

[0074] Specifically, in step 3, generating the reference velocity feedforward vector and the reference attitude feedforward vector includes:

[0075] The point cloud on the workpiece surface in the actual terrain topology space matrix is ​​fitted with a surface to extract the position vector and normal vector of the point to be sprayed.

[0076] In practical engineering, the least squares method can be used to perform optimal plane fitting on a set of three-dimensional coordinate points in a local space. The number of normals in the plane equation represents the normal vector at that point on the workpiece surface.

[0077] The normal distance from the center point of the spray gun to the point to be sprayed is constrained to be less than a preset standard spraying distance, with the difference between this distance and a minimum allowable distance deviation threshold being used as the normal constraint. The spraying robot is required to always spray along this normal direction. This distance deviation threshold is determined based on the fluid atomization cone angle performance parameters of the specific paint, thus preventing sagging or missed spraying caused by process violations.

[0078] Construct a rolling optimization objective function that relates the current prediction system state to the reference target state. The discretized expression of this rolling optimization objective function is as follows:

[0079] ;

[0080] in, This represents the scalar evaluation value of the rolling optimization objective function; This represents the prediction time-domain step size; This represents the control time-domain step size; The model represents the future... The system state vector predicted in the step-by-step output; This represents the reference target state vector given based on the user-sprayed trajectory. This represents a positive semidefinite diagonal matrix that defines the weights for the state tracking error penalty. This represents the increment of the control input vector; This represents a positive definite diagonal matrix that defines the weights controlling the incremental penalty.

[0081] Under the conditions of satisfying the normal constraint and the physical constraints of robot operation, the minimum value of the rolling optimization objective function is obtained by solving the quadratic programming problem, and the optimal control input sequence in the control time domain is obtained.

[0082] Here, physical constraints include the maximum speed limit and acceleration limit corresponding to the maximum output torque of each servo motor, preventing the algorithm from outputting erroneous commands that would overload the motors. To solve this quadratic programming problem under the limited computing power of the underlying controller, the algorithm uses an interior-point solver, limiting the maximum number of iterations to 10.

[0083] When the environmental signal-to-noise ratio is low, resulting in sparse feature points, the system uses a pre-defined linear mapping relationship to predict the time-domain step size. Automatic reduction sacrifices long-term smoothness in exchange for a reduction in the solution dimension, thereby ensuring that the quadratic programming solution can be completed within 2 milliseconds, and guaranteeing that the hard real-time characteristics of the system are not compromised.

[0084] The first element of the optimal control input sequence is extracted as the optimal control quantity at the current moment, and the spray gun reference linear velocity and spray gun reference angular velocity are separated from it and combined to form the reference velocity feedforward vector and the reference attitude feedforward vector.

[0085] Step 4: Combine the reference velocity feedforward vector, the reference attitude feedforward vector, and the current pose tracking residual to generate a composite control command. Adaptively allocate the collaborative compensation weights of the moving chassis and the robotic arm according to the amplitude of the current pose tracking residual, and drive each servo motor to complete the collaborative position adjustment of the spraying trajectory.

[0086] Specifically, in step 4, the reference velocity feedforward vector, the reference attitude feedforward vector, and the current pose tracking residual are combined to generate a composite control command, which specifically includes:

[0087] The deviation between the reference state vector generated based on the preset spraying trajectory and the current system state vector is calculated to obtain the current pose tracking residual, which includes the position residual sub-vector.

[0088] The reference velocity feedforward vector and the reference attitude feedforward vector are used as the optimal feedforward control quantities, and the feedback correction quantity is calculated based on the current pose tracking residual using a proportional-integral-derivative algorithm. The optimal feedforward control quantity and the feedback correction quantity are superimposed to generate the composite control command, which includes the chassis speed command and the robotic arm joint angular velocity command.

[0089] This composite control architecture, which combines feedforward and feedback, retains the immediate forward response of the feedforward link to environmental changes, while eliminating the steady-state error generated by the system under the action of transmission clearance and friction through the feedback integral term.

[0090] Specifically, in step 4, the collaborative compensation weights of the mobile chassis and the robotic arm are adaptively allocated based on the amplitude of the current pose tracking residual, including:

[0091] A first deviation threshold and a second deviation threshold greater than the first deviation threshold are preset. It is important to note that the physical meaning and setting basis of these two thresholds are subject to strict objective constraints. The preset first deviation threshold is calibrated by the maximum interference-free adjustment space of the robotic arm end flange without causing any singular joint positions, and is exemplarily set to 5 mm; the second deviation threshold is defined by the maximum braking friction slip distance of the tracked chassis servo motor, and is exemplarily set to 20 mm.

[0092] The magnitude of the position residual subvector is then calculated. When the magnitude is less than or equal to the first deviation threshold, the chassis speed command maintains feedforward speed tracking, and the position deviation is fully compensated by the joint angle feedback of the robotic arm.

[0093] The physical basis for this strategy is that the chassis has an extremely large mass momentum angular momentum, and requiring it to respond to high-frequency small deviations of 5 millimeters would cause the tracks to slip frequently on the ground; while the robotic arm is lightweight and has a very high bandwidth, making it the most reasonable choice to deal with this level of deviation.

[0094] When the amplitude is greater than the first deviation threshold and less than or equal to the second deviation threshold, the compensation weight assigned to the mobile chassis increases linearly with the amplitude, and the position deviation is compensated by the mobile chassis and the robotic arm in a coordinated manner.

[0095] The formula for calculating the linear compensation weight coefficient at this stage is as follows:

[0096] ;

[0097] in, This represents the dimensionless linear compensation weighting coefficient assigned to the mobile chassis, with a numerical range constrained between 0 and 1. This represents the current actual magnitude of the position residual subvector; This represents the preset first deviation threshold; This represents the preset second deviation threshold. The compensation weight assigned to the robotic arm is 1 minus... .

[0098] When the amplitude exceeds the second deviation threshold, it indicates that the robot is in a dangerous state of significant yaw. In this case, the mobile chassis takes the lead in correcting the large deviation, while the robotic arm only maintains relative posture stability to avoid overextension of the robotic arm beyond its mechanical limits, which could lead to a shutdown and power failure.

[0099] Step 5: Based on trajectory tracking error, control increment and spraying distance fluctuation, construct a comprehensive performance index, and realize closed-loop control of position adjustment by updating the decoupling filter parameters and the collaborative compensation weights online.

[0100] Specifically, in step 5, closed-loop control of position adjustment is implemented, which includes:

[0101] The standard deviation of the spraying distance fluctuation is statistically analyzed in real time based on the actual terrain topology space matrix; performance weight coefficients are assigned to the trajectory tracking error, the control increment, and the standard deviation, respectively, and the comprehensive performance index is constructed by linear weighted summation; the adaptive constraint coefficient in the decoupling filtering parameters and the collaborative compensation weight are used as the parameter vector to be tuned, and the gradient descent method is used to calculate the gradient vector of the comprehensive performance index relative to the parameter vector to be tuned; the parameter vector to be tuned is updated online based on the gradient vector and the preset learning rate to reduce the comprehensive performance index.

[0102] In practical engineering operations, to avoid excessively large learning rates that could cause parameter jumps and lead to motor vibration and loss of control, the amplitude of the preset learning rate is clamped and limited to a very small fixed value range. This online iterative process is allocated to a low-priority thread in the computing node kernel and executed in an asynchronous non-blocking mode to avoid occupying the critical time slices of the aforementioned closed-loop control calculations, thus ensuring timing safety during continuous device operation.

[0103] In summary, considering the existing serial processing architecture and passive response mechanism, current painting robots often face problems such as dynamic lag and oscillations in the painting trajectory caused by feature interference.

[0104] The implementation scheme disclosed in this embodiment fully realizes a closed loop of mutual support between technical features and algorithmic features. By fusing algorithms such as inertial prior and orthogonal feature projection with specific physical control variables and environmental perception boundaries, a parallel decoupling purification-temporal feedforward-weight allocation joint architecture is constructed. This scheme explicitly eliminates the interference of random disturbances such as paint mist and reflections on the laser point cloud, thus eliminating the uncertainty of trajectory planning. Through adaptive cooperative control of the robotic arm and chassis based on physical threshold characteristics, it not only achieves millisecond-level execution response capability within the safe range allowed by hardware constraints, but also significantly narrows the position fluctuation error of the center point of the spraying trajectory as it moves.

[0105] This embodiment is not only mathematically solvable in a closed loop, but its hardware settings are also fully compatible with current industrial standard manufacturing processes and bus control protocol designs, demonstrating significant and objective feasibility for engineering implementation.

[0106] Example 2:

[0107] Based on Example 1, this example further introduces an emergency safety feedforward compensation step for sudden high-concentration paint mist blinding conditions.

[0108] The technical solution of this embodiment serves as a high-level safety interruption service routine in the overall control process. It is specifically embedded between steps 3 and 4 of Embodiment 1 to solve the problem of trajectory divergence and hard collision of equipment caused by the lack of physical prior support when the existing technology faces severe loss of transient sensing signals.

[0109] Specifically, this embodiment relies on the underlying hardware processing platform defined in Embodiment 1. To meet the requirements for deterministic response latency under the extreme condition of perception deprivation, all artificial intelligence inference tasks and kinematic calculation modules involved in this embodiment are executed within the on-chip static random access memory of the edge computing node to avoid bus latency caused by external dynamic random access memory refresh. The execution logic and algorithm parameter settings of this emergency safety feedforward compensation step are entirely limited by the 10-millisecond system global control cycle limit.

[0110] Phase 1: Determination and triggering of transient perception-induced blindness.

[0111] Specifically, when spraying operations are carried out inside complex, enclosed ship sections or large storage tanks, sudden changes in the ventilation airflow or the instantaneous full opening of the spray gun can easily cause a large number of micron-sized paint mist particles to form a high-density aerosol reflective interference layer in front of the sensor's field of view within a very short time. This physical phenomenon can block the pulse-time-of-flight ranging optical path of the lidar and cause severe overexposure of the vision module, making it impossible for the orthogonal subspace projection algorithm in step 2 of Example 1 to obtain sufficient effective geometric information to characterize the workpiece contour after purification.

[0112] To this end, the system calculates the density of effective feature points in the actual terrain topology space matrix in real time. The calculation method for the density of effective feature points is as follows: count the number of effective feature points that have been verified and retained by the adaptive spatial geometry filtering operator within the current control cycle, and divide it by the average total number of feature points extracted by the system under interference-free benchmark calibration conditions to obtain a normalized percentage value.

[0113] When the density of effective feature points is lower than the preset perception blindness safety threshold for multiple consecutive sampling periods, the system is determined to enter a transient perception blindness state.

[0114] In actual engineering calibration, the perception blindness safety threshold is set to 5%. The physical meaning of this value is that the effective geometric features at this time are insufficient to support the full-rank solution of the six-degree-of-freedom pose matrix in three-dimensional space.

[0115] Meanwhile, the density index is required to be below the threshold for multiple consecutive sampling periods (for example, five consecutive control periods, i.e., 50 milliseconds) before a judgment is triggered. The purpose is to use the low-pass filtering characteristics of time series to filter out false positive triggers caused by transient shot noise from a single sensor or instantaneous packet loss on the data bus, and to prevent the system from switching between normal and emergency states with high-frequency oscillations.

[0116] It is important to note that the moment the system is determined to enter a transient perception blinding state, the underlying real-time operating system will immediately trigger a high-priority hardware interrupt, suspending the original rolling time-domain optimization calculation thread based on environmental feedback, thereby cutting off the erroneous guidance of the failed external perception data to the subsequent motion actuators.

[0117] Phase 2: Architecture and forward inference of the one-dimensional temporal convolutional network prediction model.

[0118] Specifically, in the transient perception blinding state, after suspending the rolling temporal optimization, the system immediately activates the one-dimensional temporal convolutional network prediction model pre-deployed in the edge controller. Since external optical and laser sensing devices fail at this time, the system can only maintain local navigation by relying on the internal inertial measurement unit, which is unaffected by aerosol physical obstruction.

[0119] The three-dimensional force sequence and three-dimensional angular velocity sequence of the inertial measurement unit within a preset time sliding window are extracted as input feature sequences. Forward inference is performed through the one-dimensional temporal convolutional network prediction model to output the pose increment prediction value at the current moment.

[0120] For example, the preset time window is set to include a data queue of the past 200 sampling points, corresponding to the high-frequency kinematic history of the past second at a sampling rate of 200 Hz. Since the time window is only 1 second, the integral drift error of the microelectromechanical system-level inertial sensor is still limited to the sub-millimeter physical range, which makes short-term high-precision trajectory estimation objectively feasible.

[0121] To complete model inference within a 10-millisecond control period and extract high-frequency dynamic features from time-series data, the one-dimensional temporal convolutional network prediction model in this embodiment employs a specific lightweight network structure design. The one-dimensional temporal convolutional network prediction model uses a three-layer causal convolutional structure with progressively increasing hole rates in each layer, no pooling layers, and an output layer that is a linear fully connected layer.

[0122] The physical significance of employing a causal convolution structure lies in ensuring that when the model calculates the pose increment at the current moment, it does not prematurely call upon or utilize data from future time steps that have not yet occurred. This strictly conforms to the real-time control logic of causality in the physical world. The design of a pooling-free layer is to avoid losing high-frequency angular velocity abrupt changes during temporal downsampling. These abrupt changes often correspond to minute vibrations at the end effector of the robotic arm or transient slippage of the chassis.

[0123] To disclose the core computational logic of this network model in detail, the temporal discrete convolution operation of the m-th hidden layer in the above causal convolution structure is precisely defined by the following formula:

[0124] ;

[0125] In the formula, Indicates the first Layered networks in discrete time steps The output hidden layer feature vector; This indicates the size of the one-dimensional convolution kernel; Indicates the first Layer offset index in one-dimensional space The convolution kernel weight matrix at the point; Indicates input up to the number The layer at the corresponding discrete time step Temporal feature vector at the location; Indicates the first The void ratio parameter set for the layer; Indicates the first Layer bias correction vector.

[0126] Explain the effect of the above formula in practical applications: by introducing a void ratio parameter that increases exponentially. (For example, the three-layer network is set to 1, 2, and 4 respectively), this formula can achieve this without increasing the kernel size. and weight matrix Given the limited number of parameters, ensure the output of the last layer of the network is optimal. The receptive field can expand exponentially, thus completely covering a preset time window of up to one second.

[0127] This computational mechanism compresses the model parameters to less than 500 bytes, and the edge processor's single-instruction multiple-data register can load all weight data into the L1 cache within one clock cycle. Real-world testing shows that the forward inference process defined by this formula executes stably within 0.4 milliseconds on the central processing unit, effectively avoiding resource conflicts between theoretical parameters and actual data throughput.

[0128] Phase 3: Offline training and data binding of the prediction model.

[0129] Optionally, to ensure the reliability of the output of the one-dimensional temporal convolutional network prediction model, its training parameters are obtained offline, and the mean squared error loss function and adaptive moment estimation optimization algorithm are used to iteratively converge based on multi-source synchronous data under historical normal operation.

[0130] In detail, the construction of the training data source strictly relies on the normal spraying operation logs of the robot in a clean and structured experimental workshop. When collecting data in the offline environment, the LiDAR and vision module operate at the optimal signal-to-noise ratio. At this time, the multi-source fusion high-precision pose increment generated by the system is recorded as the absolute ground truth label for model training; the historical data sequence of the inertial measurement unit recorded simultaneously serves as the input for training.

[0131] In terms of specific training methods, the training dataset is divided into a training set and a validation set in an 8:2 ratio. The mean squared error between the predicted values ​​generated by the linear fully connected layer connected after the output of the above formula and the ground truth labels is used as the loss function.

[0132] To overcome the common gradient explosion or vanishing problems in time series network training, an adaptive moment estimation optimization algorithm is used to update the weight parameters through backpropagation. The initial learning rate is set to 0.01%, and a dynamic learning rate adjustment strategy based on validation set loss decay is configured.

[0133] After 500 iterations, training stops and the model weights are fixed when the mean square error of the pose increment on the validation set converges to within 0.1 mm. This training mechanism ensures that the extracted algorithm model is an effective high-dimensional nonlinear fit to the kinematic differential equations of rigid bodies under real physical conditions.

[0134] Phase 4: Implementation of emergency safety control strategies and physical rollback constraints.

[0135] Specifically, after the model outputs, a pure inertial trajectory is calculated based on the predicted pose increment, generating an emergency pose feedforward to replace the reference velocity feedforward vector and the reference attitude feedforward vector. In this step, the system numerically integrates the six-degree-of-freedom differential increment output by the neural network based on the current known precise pose, maintaining the system's continuous awareness of its own spatial state.

[0136] Furthermore, and this is the core of the safety control logic in this embodiment, it is to simultaneously apply a physical backoff constraint to the composite control command, drive the robotic arm to back off by a preset safety margin distance in the opposite direction of the last effectively extracted normal constraint, and control the moving chassis to perform braking at a constant deceleration.

[0137] When blindness occurs, there may be complex unstructured obstacles around the robot (such as unfinished ship ribs). If the system is allowed to blindly continue to execute the original painting operation trajectory by relying on inertia, the high cumulative error can easily lead to physical collision between the end of the robotic arm and the workpiece.

[0138] Physical laws show that in the last time slice before the blind zone occurs, because the robotic arm strictly executes the spraying process perpendicular to the workpiece surface, the normal direction of the workpiece surface at that moment is usually a collision-free path pointing to an open and safe space.

[0139] Therefore, the robotic arm is driven to retract in the opposite direction to the last effectively extracted normal constraint. To ensure that this retraction process conforms to the dynamic boundaries of the servo drive and to guarantee the smoothness of the deceleration process to avoid damage to the harmonic reducer, the emergency retraction speed command of the robotic arm end effector is generated by the following formula:

[0140] ;

[0141] In the formula, This represents the emergency retraction speed control vector generated by the robotic arm's end effector in the three-dimensional operating space; This represents the maximum permissible escape scalar velocity at the end of the robotic arm, calculated in advance based on the rated power of the motors at each joint of the robotic arm. The base of the natural logarithm; This indicates the remaining backtracking distance from the target stopping point at the current moment; This indicates the preset safety margin distance; This represents the last validly extracted spatial normal unit vector.

[0142] To illustrate the function of the above formula in practical applications: when a blinding event has just occurred, the remaining backtracking distance... equal to safety margin distance As the exponential term approaches the initial decay value, the robotic arm rapidly retreats from the dangerous, close-to-work area at a high reverse retreat speed; as the robotic arm continues to retreat along the collision-free corridor, the remaining retreat distance... As the exponent gradually approaches 0, the exponent term rapidly approaches 1, causing the velocity decay factor formed by the expression within the parentheses to approach 0.

[0143] This, through both mathematical mechanisms and physical execution, ensures that the robotic arm can smoothly and undisturbedly reduce its speed to zero after reaching the safe zone, achieving a soft landing and hovering. Subsequently, the angle feedforward commands for each joint are mapped through the pseudo-inverse operation of the Jacobian matrix and directly fed into the underlying servo drive bus.

[0144] At the same time, for tracked mobile chassis with huge mass and inertia, directly locking the brake will cause the static friction between the track and the ground to suddenly change into sliding friction, causing the chassis to produce uncontrollable sideslip and yaw.

[0145] Therefore, the mobile chassis is controlled to brake at a constant deceleration. This constant deceleration is calculated precisely based on the maximum static friction coefficient between the track rubber material and the steel workshop floor, ensuring the chassis stops at the shortest possible distance within the pure rolling physical limit. This strategy ensures that the chassis and robotic arm retreat collaboratively to a safe envelope, avoiding kinematic interference.

[0146] Phase 5: Smooth handover of perception recovery and control weights.

[0147] It is important to note that the blinding effect is not permanent. As the workshop's ventilation system continues to operate, the aerosol concentration will gradually decrease. While the system performs physical retraction and braking shutdown, the backend sensor channels remain operational.

[0148] The system applies braking until the density of the effective feature points recovers to above the perception blindness safety threshold, then exits the transient perception blindness state. When the exit condition is met, the system does not directly perform a state transition. Instead, it re-injects the final pose obtained based on inertial calculations as the initial state into the rolling temporal optimization framework. In the subsequent 10 control cycles, linear interpolation is used to gradually transfer the control weights of the composite control commands back to the conventional control loop based on multi-source fusion and actual terrain topology matrix derivation, ensuring that the robot can smoothly resume the normal autonomous localization, mapping, and painting operation cycle.

[0149] In summary, considering the limitations of existing technologies, traditional graph-optimized positioning frameworks or loosely coupled systems based on Kalman filtering often suffer from complete failure of feature matching when encountering blinding conditions caused by high concentrations of paint fog. This leads to divergence of the control matrix, abnormal movement of the actuator, and ultimately mechanical damage.

[0150] The overall effectiveness of this embodiment lies in the following: by combining a lightweight temporal convolutional network model with strict physical and mechanical braking constraints, a purely inertial-driven safe retreat path is constructed with extremely low computational overhead without adding any expensive fog-penetrating radar or other additional hypothetical redundant hardware. Even in extreme conditions where the external optical physical channel is instantly severed, this solution not only maintains the short-term high-precision continuous operation of the core positioning system, but more importantly, it forcibly regulates the optimal physical retreat path to avoid workpiece collisions at the execution layer, demonstrating extremely high system robustness and objectively real-world engineering manufacturing value.

[0151] Example 3:

[0152] like Figure 2 As shown, this embodiment provides a multi-source fusion positioning and mapping system for a painting robot under strong interference conditions, which is used to execute and implement the position adjustment and mapping methods described in Embodiment 1 and Embodiment 2.

[0153] In complex industrial automated production, existing painting robots often suffer from problems such as map topology distortion, computing power overflow, and end-point trajectory oscillation when encountering high-concentration paint mist or strong reflective metal workpieces due to the lack of a real-time separation mechanism for dynamic disturbance characteristics and a closed-loop weight allocation mechanism for collaboration among multiple actuators.

[0154] To address the aforementioned technical issues, this embodiment is deployed on a real heterogeneous multi-core edge computing hardware platform (such as an industrial control motherboard integrating an ARM central processing unit, a digital signal processing unit (DSP), and a field-programmable gate array (FPGA). The overall architecture covers the following core functional modules.

[0155] Specifically, a multi-source fusion positioning and mapping system for a painting robot under strong interference conditions is applied to a mobile painting robot integrating lidar, vision module, inertial measurement unit, and including a mobile chassis, robotic arm, and spray gun. The system includes:

[0156] Real-time environmental geometric signal construction module: This module is used to synchronously trigger the lidar, vision module and inertial measurement unit to collect spraying site data through hardware, and construct a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamp and source identifier after aligning the timestamp.

[0157] Specifically, in complex painting workshops, the heterogeneity of different sensors can easily lead to timing misalignments in data sampling. This module relies on a hard real-time synchronous clock generator provided by an FPGA chip at the physical level. By sending a consistent Precision Time Protocol (PTP) pulse sequence to the LiDAR's rotary motor controller, the industrial camera's global shutter trigger pin, and the MEMS inertial measurement unit, it enforces underlying hardware alignment. After the data stream merges into the system bus, the module packages and stitches together the discrete point cloud 3D coordinates, the grayscale feature intensity of image pixels, the unified global system timestamp, and the sensor device identifier bits that distinguish data attributes into a multi-dimensional state matrix. This module establishes the spatiotemporal uniformity of the perceived information from the data source, preventing minute time phase differences from being amplified into pose drift errors in subsequent high-speed motion calculations.

[0158] Actual terrain topology space matrix purification module: This module is used to construct an adaptive spatial geometry filtering operator using the high-frequency kinematic features of the inertial measurement unit, and combine orthogonal subspace projection to decouple the real-time environmental geometric signal in the frequency domain, thereby purifying the actual terrain topology space matrix that represents the real workpiece boundary and robot pose.

[0159] Specifically, in practical engineering applications, paint mist particles and metallic reflections on the workpiece surface create random, high-density illusory obstacle points in three-dimensional space. The computational tasks of this module are primarily distributed within the DSP unit of the edge controller.

[0160] First, a kinematic state observer in three-dimensional space is constructed using high-frequency linear and angular velocities obtained through real-time integration by the inertial measurement unit, forming an adaptive spatial geometric filtering operator to verify the rigid body motion laws. Then, the system performs eigenvalue decomposition to obtain the orthogonal subspace basis matrix representing the static stable environment.

[0161] To ensure the robustness of this module under extreme conditions, during the preliminary research and setting of the basis matrix parameters, a massive amount of typical interference data is typically extracted using an offline engineering simulation platform for signal de-noising testing and verification. By comparing the power spectral density curves before and after environmental feature projection and the state tracking waveforms of the controller, the convergence of the filter parameters in a noisy environment and the accuracy of reconstructing the underlying sensor data are quantitatively verified.

[0162] After parameter verification and solidification, the algorithm model can stably project real-time environmental geometric signals containing high-frequency noise into an orthogonal subspace, eliminate random scattering and high-frequency interference components, and purify the actual terrain topology space matrix that represents the rigid structure of the real physical environment.

[0163] Rolling temporal optimization feedforward generation module: This module is used to inject the actual terrain topology space matrix into the trajectory control loop, extract the workpiece surface morphology and normal constraints, dynamically adjust the prediction time domain step size and control time domain step size according to the environmental signal-to-noise ratio, and generate reference velocity feedforward vector and reference attitude feedforward vector through rolling temporal optimization.

[0164] Specifically, in the spraying process, the spray gun must maintain a specific normal distance and angle with the workpiece surface. After receiving the actual terrain topology space matrix, this module uses the least squares method to fit and extract the local normal vector and geometric shape surface of the workpiece surface. The module has a built-in model predictive control (MPC) algorithm kernel based on discrete state-space equations. Its input is the current dynamic state and the above-mentioned normal constraint equations, and its output is the optimal control sequence in the future finite time domain.

[0165] To balance the computational power consumption of the computing platform with the sensitivity of the robot system's response, this module calculates the spatial density of effective feature points in real time within the current sampling period, using this as an objective indicator to evaluate the signal-to-noise ratio (SNR) of the current environment. When the SNR decreases (e.g., encountering a sudden dense fog), the module adaptively shortens the prediction time-domain step size through a preset mapping logic. This parameter adjustment effectively reduces the dimension of the Hessian matrix in the quadratic programming solver, reducing the search space for algorithm optimization. Therefore, even in edge computing nodes with strictly limited computing power, it ensures that the single-operation time of the multi-objective optimization control loop never exceeds the hard control cycle limit of ten milliseconds. The optimal feedforward obtained through the solution serves as the reference velocity feedforward vector and reference attitude feedforward vector sent to the execution layer.

[0166] Cooperative position adjustment control module: This module is used to generate composite control commands by combining the reference velocity feedforward vector, the reference attitude feedforward vector and the current pose tracking residual. Based on the amplitude of the current pose tracking residual, it adaptively allocates the cooperative compensation weights of the moving chassis and the robotic arm, and drives each servo motor to complete the cooperative position adjustment of the spraying trajectory.

[0167] Specifically, this module is the hub for scheduling and distributing physical motion of the equipment via the bus. Its output signal is converted into pulse width modulation (PWM) commands to directly control the chassis traction motor and the servo drive of the robotic arm joint. The system performs parallel closed-loop calculations with the ideal reference feedforward vector and the current true pose tracking residual obtained through feedback calculation by the servo encoder.

[0168] Due to the significant differences in the physical and dynamic properties of the multiple actuators in the system, the module incorporates a frequency-band mechanical threshold allocation mechanism: the tracked mobile chassis has a large moment of inertia and static friction with the ground, making it suitable for handling large-scale low-frequency displacement following; while the six-axis robotic arm has a small moment of inertia and extremely high control bandwidth, making it suitable for real-time response to small high-frequency errors. By calculating the scalar amplitude of the position residual in real time, the module dynamically adjusts the collaborative compensation weight between the two: when the residual amplitude is within the safety tolerance, all high-frequency compensation commands are allocated to the robotic arm; when the residual amplitude exceeds the robotic arm's compensation limit threshold due to external impact, the following weight of the chassis is gradually increased. This trade-off strategy based on physical laws effectively avoids mechanical transmission oscillations and trajectory loss of control caused by drive saturation of a single actuator.

[0169] Closed-loop control adaptive iterative module: This module is used to construct a comprehensive performance index based on trajectory tracking error, control increment and spraying distance fluctuation, and realize closed-loop control of position adjustment by updating the decoupling filtering parameters and the collaborative compensation weights online.

[0170] Specifically, considering that equipment in actual industrial manufacturing environments has long-term time-varying physical factors such as mechanical gear wear and changes in lubrication damping, it is difficult to maintain the initial motion accuracy by operating the control weights fixed for a long time.

[0171] This module is configured to run in a low-priority background thread of the operating system, and collects multi-dimensional operational status data in real time in non-blocking mode, including absolute position error, servo control command increment, and fluctuation of physical spraying normal distance.

[0172] The algorithm employs a linear weighting method to synthesize the aforementioned multi-source evaluation indicators into a single cost function. During the optimization iteration process, the system combines the distribution characteristics of Pareto Front Scatters constructed in the offline testing phase to objectively evaluate the physical trade-off boundaries between multiple objectives such as energy consumption, error, and control smoothness, thereby determining the optimal gradient descent direction that does not violate kinematic physical limits. Using this gradient direction, the filtering threshold of the orthogonal decoupling matrix and the aforementioned collaborative compensation weights are adjusted online asynchronously, enabling the control system to achieve self-driven parameter closed-loop evolution during production operation.

[0173] The system described in this embodiment effectively filters out geometric signal distortion and sensor contamination in strong interference environments containing dense paint fog and reflections by extending precise clock synchronization from the hardware level to the frequency domain orthogonal subspace for data purification, without the need to add expensive redundant detection hardware.

[0174] Furthermore, the system dynamically adjusts the time step of time-domain optimization by utilizing the environmental signal-to-noise ratio. This ensures the real-time performance of the trajectory feedforward inference algorithm under extreme conditions without exceeding the computing power limits of existing general-purpose industrial control platforms, thus avoiding control phase shift. It also performs adaptive physical weighting and online parameter iteration on the chassis and robotic arm through residual amplitude. The system significantly suppresses high-frequency jitter of the robot's end-effector trajectory under large-scale displacement, maintains the absolute stability of the normal working distance required for spraying operations, and possesses a practical and reliable engineering implementation capability.

[0175] Example 4:

[0176] Corresponding to the above embodiments, the present invention also proposes an electronic device.

[0177] like Figure 3 The diagram shows a structural schematic of an electronic device according to the present invention. The electronic device 100 includes a processor 101 and a memory 103. The processor 101 and the memory 103 are connected, for example, via a bus 102. Optionally, the electronic device 100 may further include a transceiver 104. It should be noted that in practical applications, the transceiver 104 is not limited to one unit, and the structure of this electronic device 100 does not constitute a limitation on the embodiments of the present invention.

[0178] Processor 101 may be a CPU, a general-purpose processor, a DSP, an ASIC, an FPGA, or other programmable logic device, transistor logic device, hardware component, or any combination thereof. It may implement or execute the various exemplary logic blocks, modules, and circuits described in connection with this disclosure. Processor 101 may also be a combination that implements computational functions, such as including one or more microprocessor combinations, a combination of a DSP and a microprocessor, etc.

[0179] Bus 102 may include a pathway for transmitting information between the aforementioned components. Bus 102 may be a PCI bus or an EISA bus, etc. Bus 102 may be divided into an address bus, a data bus, a control bus, etc. For ease of representation, Figure 3 The bus is represented by a single thick line, but this does not mean that there is only one bus or one type of bus.

[0180] The memory 103 stores a computer program corresponding to the multi-source fusion localization and mapping method for a painting robot under strong interference conditions according to the above embodiments of the present invention. This computer program is controlled and executed by the processor 101. The processor 101 executes the computer program stored in the memory 103 to implement the content shown in the aforementioned method embodiments.

[0181] Among them, electronic devices 100 include, but are not limited to: mobile terminals such as laptops and PADs (tablet computers) and fixed terminals such as desktop computers. Figure 3 The electronic device 100 shown is merely an example and should not be construed as limiting the functionality and scope of the embodiments of the present invention.

[0182] Although embodiments of the present invention have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting the present invention. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of the present invention.

Claims

1. A multi-source fusion localization and mapping method for a painting robot under strong interference conditions, applied to a position adjustment system of a mobile painting robot integrating lidar, vision module, inertial measurement unit, and including a mobile chassis, robotic arm, and spray gun, characterized in that... include: Step 1: The lidar, vision module and inertial measurement unit are synchronously triggered by hardware to collect spraying site data, and after aligning the timestamps, a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamp and source identifier is constructed. Step 2: Construct an adaptive spatial geometry filtering operator using the high-frequency kinematic characteristics of the inertial measurement unit, and combine it with orthogonal subspace projection to decouple the real-time environmental geometry signal in the frequency domain, thereby refining the actual terrain topology space matrix that represents the real workpiece boundary and robot pose. Step 3: Inject the actual terrain topology space matrix into the trajectory control loop, extract the workpiece surface morphology and normal constraints, dynamically adjust the prediction time domain step size and control time domain step size according to the environmental signal-to-noise ratio, and generate the reference velocity feedforward vector and reference attitude feedforward vector through rolling time domain optimization. Step 4: Combine the reference velocity feedforward vector, the reference attitude feedforward vector, and the current pose tracking residual to generate a composite control command. Adaptively allocate the collaborative compensation weights of the moving chassis and the robotic arm according to the amplitude of the current pose tracking residual, and drive each servo motor to complete the collaborative position adjustment of the spraying trajectory. Step 5: Based on trajectory tracking error, control increment and spraying distance fluctuation, construct a comprehensive performance index, and realize closed-loop control of position adjustment by updating the decoupling filter parameters and the collaborative compensation weights online.

2. The method according to claim 1, characterized in that, Step 1, in which the real-time environmental geometric signal is constructed, specifically includes: Obtain the original three-dimensional point cloud set output by the lidar, and extract the three-dimensional coordinates of the spatial sampling points in the carrier system and the corresponding laser reflection intensity values; Obtain the original visual feature point set output by the visual module, and extract the pixel coordinates and corresponding grayscale gradient magnitudes of the feature points; Obtain the original measurement vector output by the inertial measurement unit, and extract the three-dimensional specific force vector and three-dimensional angular velocity vector under the load system; Based on a preset sampling interval, the three-dimensional coordinates and the pixel coordinates at the same sampling time are integrated into the spatial coordinates, the laser reflection intensity value and the grayscale gradient amplitude are integrated into the feature intensity, and the corresponding timestamp and source identifier are added to generate the real-time environmental geometric signal in matrix form.

3. The method according to claim 1, characterized in that, In step 2, before constructing the adaptive spatial geometry filtering operator using the high-frequency kinematic characteristics of the inertial measurement unit, the following steps are also included: Set a preset navigation coordinate system, integrate the attitude rotation matrix of the previous moment with the current moment's angular velocity increment after zero angular velocity offset compensation, and obtain the attitude rotation matrix of the current moment. The three-dimensional specific force vector after acceleration zero bias compensation is transformed to the preset navigation coordinate system using the attitude rotation matrix at the current moment, and the constant vector of gravitational acceleration is subtracted to obtain the linear acceleration in the preset navigation coordinate system at the current moment. By integrating and recursively applying the linear acceleration, a high-frequency kinematic state containing the current position and linear velocity of the painting robot is obtained.

4. The method according to claim 3, characterized in that, In step 2, frequency domain decoupling of the real-time environmental geometric signal is performed by combining orthogonal subspace projection, specifically including: Using the attitude rotation matrix at the current moment and the position at the current moment, the spatial coordinates in the real-time environmental geometric signal are transformed to the preset navigation coordinate system to obtain the navigation system spatial coordinate matrix; Two sets of strictly orthogonal feature subspaces are obtained in advance through training with typical working condition samples. The feature subspaces include a real terrain feature subspace representing a fixed structure and a noise feature subspace representing interference features. The navigation system spatial coordinate matrix is ​​orthogonally projected onto the real terrain feature subspace, and the components corresponding to the noise feature subspace are removed to obtain the purified real terrain coordinate matrix. The purified real terrain coordinate matrix is ​​converted back into the loading system, and combined with the corresponding feature intensity, timestamp and source identifier, the actual terrain topology space matrix at the current moment is constructed.

5. The method according to claim 4, characterized in that, The adaptive spatial geometry filtering operator is a spatial transformation matrix constructed based on a third-order identity matrix, the linear velocity in the preset navigation coordinate system, and adaptive constraint coefficients. The adaptive constraint coefficients are positively correlated with the environmental signal-to-noise ratio. After obtaining the actual terrain topology space matrix, the motion consistency is verified using the adaptive spatial geometry filtering operator, specifically as follows: Calculate the deviation distance between the coordinate vectors of each feature point in the actual terrain topology space matrix and the adaptive spatial geometry filtering operator under spatial mapping; If the deviation distance is less than the preset distance verification threshold, the corresponding feature point is retained; otherwise, it is determined to be an outlier that does not conform to the rigid body motion law and is removed.

6. The method according to claim 1, characterized in that, In step 3, before extracting the workpiece surface morphology and normal constraints, a unified kinematic state-space model of the painting robot is also established, specifically as follows: Extract the three-dimensional position vector and three-dimensional velocity vector of the center point of the spray gun under the load system, as well as the angle vectors of each joint of the robotic arm, and combine them to form the system state vector; The three-dimensional velocity command vector of the mobile chassis and the angular velocity command vector of each joint of the robotic arm are extracted and combined to form the system control input vector; By combining the kinematic models of the chassis and the robotic arm, a discrete state equation is established that relates the system state vector to the system control input vector.

7. The method according to claim 6, characterized in that, Step 3, which generates the reference velocity feedforward vector and the reference attitude feedforward vector, specifically includes: Surface fitting is performed on the point cloud of the workpiece surface in the actual terrain topology space matrix to extract the position vector and normal vector of the point to be sprayed. The difference between the normal distance from the center point of the spray gun to the point to be sprayed and the preset standard spraying distance is less than the allowable distance deviation threshold, which is used as the normal constraint. Construct a rolling optimization objective function that associates the current prediction system state with the reference target state; Under the conditions of satisfying the normal constraint and the physical constraints of robot operation, the minimum value of the rolling optimization objective function is obtained by solving the quadratic programming problem, and the optimal control input sequence in the control time domain is obtained. The first element of the optimal control input sequence is extracted as the optimal control quantity at the current moment, and the spray gun reference linear velocity and spray gun reference angular velocity are separated from it and combined to form the reference velocity feedforward vector and the reference attitude feedforward vector.

8. The method according to claim 7, characterized in that, In step 4, the reference velocity feedforward vector, the reference attitude feedforward vector, and the current pose tracking residual are combined to generate a composite control command, specifically including: The deviation between the reference state vector generated based on the preset spraying trajectory and the current system state vector is calculated to obtain the current pose tracking residual, which includes the position residual sub-vector. The reference velocity feedforward vector and the reference attitude feedforward vector are used as the optimal feedforward control quantities, and the feedback correction quantity is calculated based on the current pose tracking residual using a proportional-integral-derivative algorithm. The optimal feedforward control quantity and the feedback correction quantity are superimposed to generate the composite control command, which includes chassis speed command and robotic arm joint angular velocity command.

9. The method according to claim 8, characterized in that, In step 4, the collaborative compensation weights of the mobile chassis and the robotic arm are adaptively allocated based on the amplitude of the current pose tracking residual, specifically including: A first deviation threshold and a second deviation threshold greater than the first deviation threshold are preset; Calculate the magnitude of the position residual subvector; When the amplitude is less than or equal to the first deviation threshold, the chassis speed command maintains feedforward speed tracking, and the position deviation is fully compensated by the joint angle feedback of the robotic arm. When the amplitude is greater than the first deviation threshold and less than or equal to the second deviation threshold, the compensation weight assigned to the mobile chassis increases linearly with the amplitude, and the position deviation is compensated by the mobile chassis and the robotic arm in collaboration. When the amplitude is greater than the second deviation threshold, the large deviation correction is mainly driven by the mobile chassis, and the robotic arm only maintains relative posture stability.

10. A multi-source fusion positioning and mapping system for a spraying robot under strong interference conditions, characterized in that, The system, which integrates LiDAR, vision modules, inertial measurement units, and a mobile painting robot comprising a mobile chassis, robotic arm, and spray gun, includes: The real-time environmental geometric signal construction module is used to synchronously trigger the lidar, the vision module and the inertial measurement unit to collect spraying site data through hardware, and construct a real-time environmental geometric signal containing spatial coordinates, feature intensity, timestamp and source identifier after aligning the timestamps. The actual terrain topology space matrix purification module is used to construct an adaptive spatial geometry filtering operator using the high-frequency kinematic features of the inertial measurement unit, and combine orthogonal subspace projection to decouple the real-time environmental geometric signal in the frequency domain, thereby purifying the actual terrain topology space matrix that represents the real workpiece boundary and robot pose. The rolling time-domain optimization feedforward generation module is used to inject the actual terrain topology space matrix into the trajectory control loop, extract the workpiece surface morphology and normal constraints, dynamically adjust the prediction time-domain step size and control time-domain step size according to the environmental signal-to-noise ratio, and generate reference velocity feedforward vector and reference attitude feedforward vector through rolling time-domain optimization. The collaborative position adjustment control module is used to generate a composite control command by combining the reference velocity feedforward vector, the reference attitude feedforward vector and the current pose tracking residual. Based on the amplitude of the current pose tracking residual, it adaptively allocates the collaborative compensation weights of the mobile chassis and the robotic arm, and drives each servo motor to complete the collaborative position adjustment of the spraying trajectory. The closed-loop control adaptive iterative module is used to construct a comprehensive performance index based on trajectory tracking error, control increment and spraying distance fluctuation, and realize closed-loop control of position adjustment by updating the decoupling filtering parameters and the collaborative compensation weights online.