Event-based vision and uncertain nmpc-based control method, program, device and storage medium for underwater robot
By combining event vision and uncertainty NMPC control methods with inertial navigation and event cameras, the problem of unstable hovering of underwater robots in turbulent and low-light environments was solved, achieving high-precision underwater operations.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- HARBIN ENGINEERING UNIVERSITY SANYA NANHAI INNOVATION & DEVELOPMENT BASE
- Filing Date
- 2026-05-15
- Publication Date
- 2026-06-12
AI Technical Summary
Underwater robots face the challenge of instability during hovering in conditions of strong turbulence, low light, and high-speed movement due to motion blur caused by traditional vision and lag in feedback control.
A control method based on event vision and uncertainty NMPC is adopted. Effective visual feature points are obtained through an inertial navigation system and an event camera. A reprojection residual vector is constructed. Combined with the hydrodynamic model of the underwater robot and historical sliding data, rolling temporal optimization is performed to achieve high-bandwidth flow field learning and stable control.
It enables underwater robots to perform centimeter-level high-precision fixed-point operations in complex sea conditions, solving the systemic collapse problem caused by traditional visual ambiguity and control lag, and ensuring the stability and precision operation capability of the robot under strong disturbance conditions.
Smart Images
Figure CN122195050A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of underwater robot control technology, specifically relating to underwater robot control methods, programs, devices, and storage media based on event vision and uncertainty NMPC. Background Technology
[0002] Deep-sea precision operations (such as subsea pipeline welding and wet-plug interface docking) require ROVs to have centimeter-level hovering capabilities. However, the complex underwater environment poses severe challenges to traditional sensing and control technologies. Mainstream solutions typically employ "traditional CMOS camera + IMU VIO navigation" combined with "PID or ADRC disturbance rejection control." Traditional frame cameras capture images at a fixed frame rate (e.g., 30fps). In underwater lighting conditions, longer exposure times are required. If the robot or water currents move violently, severe motion blur occurs, leading to loss of feature point tracking and divergence in the positioning system. Furthermore, in areas alternating between strong light (direct reflection from searchlights) and darkness, the dynamic range of ordinary cameras is insufficient, resulting in overexposed or completely black images. Existing disturbance rejection control is mostly based on instantaneous error feedback. When turbulence occurs, the robot is first veered off course before the controller begins to correct it. This "lag" causes the robot to oscillate repeatedly near its fixed point, failing to meet the requirements of precision operations.
[0003] Although current operational ROVs are generally equipped with high-definition vision and feedback control systems, they still face a vicious cycle of "visual perception failure" and "control response lag" when dealing with complex deep-sea currents. On the one hand, limited by the exposure-motion paradox of traditional frame-type CMOS sensors, the low-light underwater environment forces the camera to extend the exposure time. Once the robot is impacted by turbulence and experiences high-frequency jitter, the image will produce severe motion blur. In addition, the limited dynamic range makes it difficult to take into account both the reflection of strong searchlights and the deep background, resulting in the loss of visual features and divergence in VIO positioning. On the other hand, traditional control strategies based on error feedback (such as PID / SMC) are inherently passive and have phase lag. They can only compensate after the position deviation occurs, making it difficult to overcome the huge inertia of the water and lacking the ability to remember and predict the spatial structure of the flow field, causing the robot to oscillate repeatedly near the fixed point. This intertwining of perception layer failure and control layer lag makes it highly susceptible to systemic collapse under strong disturbance conditions, leading to a cycle of "jitter causing ambiguity—ambiguity causing positioning drift—drift causing erroneous thrust—thrust exacerbating jitter," which fails to meet the stability requirements of precision contact operations. Summary of the Invention
[0004] The purpose of this invention is to solve the problem of instability in stationary hovering caused by traditional visual motion blur and feedback control lag in underwater robots under conditions of strong turbulence, low light, and high-speed movement. It provides an underwater robot control method, program, device, and storage medium based on event vision and uncertainty NMPC.
[0005] The underwater robot control method based on event vision and uncertainty NMPC includes the following steps:
[0006] The underwater robot is equipped with an inertial navigation system and an event camera. For the image acquired by the event camera, if the logarithmic change in light intensity of a pixel in the image exceeds the event polarity contrast threshold, then the pixel is taken as a valid visual feature point.
[0007] The position vector, linear velocity vector, and angular velocity vector of the underwater robot are obtained through the inertial navigation system. Based on the attitude vector of the underwater robot at the previous moment, the prior attitude vector of the underwater robot at the current moment is calculated and projected onto the pixel plane. The difference between the prior attitude vector and the pixel coordinates of each effective visual feature point is calculated to construct the reprojection residual vector.
[0008] The position error components and attitude error components of the event camera are taken as the objectives to be solved. Based on the reprojection residual vector, a nonlinear optimization objective function is constructed. The optimal position error components and attitude error components of the event camera that minimize the value of the nonlinear optimization objective function are solved. The prior attitude vector is corrected to obtain the attitude vector of the underwater robot at the current moment.
[0009] The position vector, linear velocity vector, attitude vector of the underwater robot, and the zero bias vectors of the accelerometer and gyroscope in the inertial navigation system are used to construct the state vector. Based on the hydrodynamic model of the underwater robot, the flow field disturbance vector experienced by the underwater robot at the current moment is calculated and stored in the historical sliding dataset. Based on the state vector of the underwater robot at the current moment, the posterior distribution is calculated using the kernel function vector to obtain the predicted disturbance mean vector and the predicted disturbance variance vector.
[0010] The predicted disturbance mean vector is substituted into the state transition equation as a feedforward disturbance term, obstacle state information is introduced as a constraint, and the predicted disturbance variance vector is substituted into the safe collision avoidance distance constraint equation. The tracking control problem is transformed into a rolling time domain optimization problem, and a cost function is constructed. The optimal predictive control sequence is obtained by solving the cost function, and the optimal control torque vector of the underwater robot at the current moment is determined.
[0011] Furthermore, for the image acquired by the event camera, if the logarithmic change in the light intensity of a pixel in the image exceeds the event polarity contrast threshold, specifically:
[0012]
[0013] in, for pixel coordinates in the image at any time The light intensity at that location; The contrast threshold; pixel coordinates The polarity of events at that location .
[0014] Furthermore, the underwater robot's attitude vector from the previous moment... Calculate the prior attitude vector of the underwater robot at the current moment. :
[0015]
[0016] in, for The angular velocity vector of the underwater robot is constantly obtained through the inertial navigation system; This is the zero bias vector of the gyroscope in the inertial navigation system;
[0017] The prior attitude vector of the underwater robot at the current moment. Project the vector onto the pixel plane and subtract it from the pixel coordinates of each effective visual feature point to construct the reprojection residual vector. :
[0018]
[0019] in, for Time of the first Pixel coordinates of effective visual feature points , for The number of valid visual feature points at any given time; This is the intrinsic parameter matrix of the event camera; The difference in the installation rotation angle of the event camera; The installation translation distance of the event camera relative to the inertial navigation system; According to the first Pixel coordinates of effective visual feature points Position error components of the event camera With attitude error components The calculated first The position vector of effective visual feature points; for The position vector of the underwater robot is obtained in real time through the inertial navigation system; This is the homogeneous coordinate normalized projection function.
[0020] Furthermore, the position error component of the event camera... With attitude error components As the objective to be solved, based on the reprojected residual vector Construct a nonlinear optimization objective function:
[0021]
[0022] in, For a priori residuals; for Time of the first Weight scalars of effective visual feature points; For the first The measurement noise covariance matrix of effective visual feature points;
[0023] Find the optimal position error components of the event camera that minimize the nonlinear optimization objective function. With attitude error components ,Right now ;
[0024] For the prior attitude vector After correction, the attitude vector of the underwater robot at the current moment is obtained. ;
[0025]
[0026] Furthermore, the aforementioned Time of the first Weight scalar of effective visual feature points The method for obtaining it is as follows:
[0027] In the Calculate the covariance matrix of the event distribution within the spatiotemporal neighborhood of effective visual feature points. And calculate the covariance matrix of the event distribution. eigenvalues and , Then calculate the coherence index. :
[0028]
[0029] The coherence index is activated by the Sigmoid activation function. Mapped to weight scalar ;
[0030]
[0031] in, This is the gain coefficient; An empirical threshold for distinguishing between bubbles and static objects.
[0032] Furthermore, the position vector of the underwater robot Linear velocity vector Attitude vector and the zero bias vector of the accelerometer in the inertial navigation system With the zero bias vector of the gyroscope Construct as a state vector ;
[0033] Based on the hydrodynamic model of the underwater robot, calculate the flow field disturbance vector experienced by the underwater robot at the current moment. ;
[0034]
[0035] in, Here is the inertial matrix of the underwater robot; Add a Coriolis force matrix for the underwater robot; Here is the damping matrix for the underwater robot; This represents the static restoring force and torque matrix of an underwater robot. for The control torque vector of the underwater robot at any given time;
[0036] Will Store in the historical sliding dataset, i.e. Based on the current The state vector of the underwater robot at any time The posterior distribution is calculated using the kernel function vector to obtain the predicted perturbation mean vector. With the predicted perturbation variance vector ;
[0037]
[0038]
[0039] in, For the present The state vector of the underwater robot at any time The prior autocovariance scalar; For the front The state vector of the underwater robot at any time Compared with historical sliding datasets The cross-covariance vector between them; For historical sliding datasets The prior autocovariance vector; It is a unit array.
[0040] Furthermore, the predicted perturbation mean vector Substituting the feedforward perturbation term into the state transition equation, and introducing obstacle state information as a constraint, the predicted perturbation variance vector is used. Substituting the safety collision avoidance distance constraint equation, the tracking control problem is transformed into a rolling time-domain optimization problem, and the cost function is constructed as follows:
[0041]
[0042] in, and This is the weight matrix;
[0043] Set the prediction range and control range Set predictive control sequence and predicted output sequence :
[0044]
[0045]
[0046] The constraints are:
[0047] (1) , , ; and The minimum and maximum allowable control torques; , ;
[0048] (2) , , This is the state transition function. The perturbation input mapping matrix;
[0049] (3) Introduce obstacle state information as a constraint condition, and predict the perturbation variance vector. Substituting into the safety collision distance constraint equation, , The state vector of the underwater robot is Distance to the nearest obstacle To maintain a safe distance for underwater robots, For risk coefficient, This represents the computation of the trace of a matrix;
[0050] The optimal predictive control sequence is obtained by solving the cost function. ;
[0051]
[0052] Find the optimal predictive control sequence The first item As the optimal control torque vector of the underwater robot at the current moment. ,Right now .
[0053] A computer device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of the above-described underwater robot control method based on event vision and uncertainty NMPC.
[0054] A computer-readable storage medium storing computer instructions that, when executed by a processor, implement the steps of the above-described underwater robot control method based on event vision and uncertainty NMPC.
[0055] A computer program product includes a computer program that, when executed by a processor, implements the steps of the above-described underwater robot control method based on event vision and uncertainty NMPC.
[0056] The beneficial effects of this invention are as follows:
[0057] This invention solves the motion ambiguity problem in the perception layer through neuromorphic vision, addresses the singularity problem in state estimation through Lie group manifold filtering, and resolves the hysteresis problem in the control layer through GP-NMPC. These three components are tightly coupled: the high bandwidth of the perception layer provides high signal-to-noise ratio data for flow field learning; accurate flow field learning provides reliable feedforward information for NMPC; and the stable control of NMPC, in turn, ensures the smooth operation of the perception layer. This enables the underwater robot to perform centimeter-level high-precision stationary operations in complex sea conditions. Attached Figure Description
[0058] Figure 1 This is the overall flowchart of the present invention. Detailed Implementation
[0059] The present invention will now be further described with reference to the accompanying drawings.
[0060] The underwater robot control method based on event vision and uncertainty NMPC provided by this invention includes the following steps:
[0061] Step 1: The underwater robot is equipped with an inertial navigation system and an event camera. Currently... At any given moment, images are acquired via an event camera, and the three-dimensional position vector of the underwater robot is obtained via an inertial navigation system. 3D linear velocity vector and three-dimensional angular velocity vector ;
[0062] Step 2: Current At any given moment, an image is acquired via the event camera. If the logarithmic change in the light intensity of a pixel in the image exceeds the event polarity contrast threshold, then that pixel is taken as the current [event / location]. Effective visual feature points at any given moment;
[0063] The logarithmic change in light intensity of pixels in the image exceeds the event polarity contrast threshold:
[0064]
[0065] in, for Time of the first Two-dimensional pixel coordinates of effective visual feature points , for The number of valid visual feature points at any given time; For the present Pixel coordinates in the image at time [time] The light intensity at that location; For the previous moment, that is Pixel coordinates in the image at time [time] The light intensity at that location; The contrast threshold; pixel coordinates Event polarity at the location;
[0066]
[0067] Step 3: For the current The first moment Effective visual feature points, in the first Calculate the covariance matrix of the event distribution within the spatiotemporal neighborhood of effective visual feature points. And calculate the covariance matrix of the event distribution. eigenvalues and , Then calculate the coherence index. ;
[0068]
[0069] The coherence index is activated by the Sigmoid activation function. Mapped to weight scalar ;
[0070]
[0071] in, This is the gain coefficient; An empirical threshold for distinguishing between bubbles and static objects;
[0072] Step 4: Based on the previous moment, i.e. Attention vector of underwater robot at any time Based on the current Angular velocity vector of the underwater robot at any moment Calculate the prior attitude vector ;
[0073]
[0074] in, Let be the zero bias vector of the gyroscope in the inertial navigation system; superscript As an antisymmetric matrix operator, it transforms a three-dimensional vector into... Matrix, and then through The rotation matrix is obtained by performing matrix exponential mapping;
[0075] Step 5: Calculate the position error components of the event camera. With attitude error components As the objective to be solved; the prior attitude vector is used as the target. Projected onto the pixel plane, the two-dimensional pixel coordinates of each effective visual feature point Divide the values to construct the reprojected residual vector. ;
[0076] Weight scalar based on each effective visual feature point With reprojection residual vector Construct a nonlinear optimization objective function Find the optimal position error components of the event camera that minimize the nonlinear optimization objective function. With attitude error components ;
[0077]
[0078] Nonlinear optimization objective function :
[0079]
[0080]
[0081] in, For a priori residuals; For the first The measurement noise covariance matrix of effective visual feature points; For homogeneous coordinate normalized projection function; This is the intrinsic parameter matrix of the event camera; The difference in the installation rotation angle of the event camera; The installation translation distance of the event camera relative to the inertial navigation system; According to the first Two-dimensional pixel coordinates of effective visual feature points Position error components of the event camera With attitude error components The calculated first Three-dimensional position vectors of effective visual feature points;
[0082] Step 6: Based on the prior attitude vector Optimal position error components With attitude error components Calculate the current Attention vector of underwater robot at any time ;
[0083]
[0084] Step 7: Based on the current The three-dimensional position vector of the underwater robot at any time 3D linear velocity vector Attitude vector and the zero bias vector of the accelerometer in the inertial navigation system With the zero bias vector of the gyroscope Construct state vector ;
[0085] Step 8: Set the current Attention vector of underwater robot at any time With the three-dimensional linear velocity vector as well as Control torque vector at any moment The data is input into the hydrodynamic model of the underwater robot to calculate the current state of the underwater robot. The three-dimensional flow field disturbance vector that is constantly subjected to ;
[0086]
[0087] in, for The derivative with respect to time, i.e., the current... The three-dimensional linear acceleration vector of the underwater robot at any given moment; Here is the inertial matrix of the underwater robot; Add a Coriolis force matrix for the underwater robot; Here is the damping matrix for the underwater robot; This represents the static restoring force and torque matrix of an underwater robot.
[0088] Step 9: Position the underwater robot as currently... The three-dimensional flow field disturbance vector that is constantly subjected to Store historical sliding dataset Based on the current The state vector of the underwater robot at any time The posterior distribution is calculated using the kernel function vector to obtain the predicted perturbation mean vector. With the predicted perturbation variance vector ;
[0089]
[0090]
[0091] in, For the present The state vector of the underwater robot at any time The prior autocovariance scalar; For the front The state vector of the underwater robot at any time Compared with historical sliding datasets The cross-covariance vector between them; For historical sliding datasets The prior autocovariance vector;
[0092] Step 10: Set the prediction interval and control range Set predictive control sequence and predicted output sequence :
[0093]
[0094]
[0095] Predict the mean vector of the disturbance Substituting the feedforward perturbation term into the state transition equation, and introducing obstacle state information as a constraint, the predicted perturbation variance vector is used. Substituting the equations for safe collision avoidance distances, the tracking control problem is transformed into a rolling time-domain optimization problem, and a cost function is constructed. :
[0096]
[0097] in, and This is the weight matrix. , ;
[0098] The constraints are:
[0099] (1) , , ; and The minimum and maximum allowable control torques; , ;
[0100] (2) , , This is the state transition function. The perturbation input mapping matrix;
[0101] (3) Introduce obstacle state information as a constraint condition, and predict the perturbation variance vector. Substituting into the safety collision distance constraint equation, , The state vector of the underwater robot is Distance to the nearest obstacle To maintain a safe distance for underwater robots, For risk coefficient, This represents the computation of the trace of a matrix;
[0102] The optimal predictive control sequence is obtained by solving the optimization problem. ;
[0103]
[0104] Pick The first item in the list, namely, let This yields the optimal control torque vector for the underwater robot. .
[0105] 1. A robust state estimation method based on spatiotemporal attention mechanism and Lie group manifold:
[0106] This invention proposes a coupled architecture of "front-end attention denoising + back-end manifold filtering". It abandons the traditional global spatiotemporal surface algorithm, which is sensitive to underwater bubbles and suspended impurities, and innovatively introduces a spatiotemporal attention mechanism. By calculating the consistency score of local optical flow, a dynamic attention mask is automatically generated, assigning high weights to structured features and low weights to discrete bubble noise. Furthermore, these attention weights are explicitly injected into the Lie group error state Kalman filter through the observation noise covariance matrix. This not only utilizes the tangent space update of the Lie algebra space to avoid gimbal deadlock, but also fundamentally eliminates the interference of unstructured water noise on pose calculation through the attention weighting mechanism, ensuring the long-term drift stability of the underwater robot in turbid, multi-bubble environments.
[0107] 2. Uncertainty-aware sparse GP-NMPC opportunity-constrained control architecture:
[0108] This invention constructs a disturbance-resistant control closed loop. Addressing the computational limitations of embedded platforms, a sparse incremental Gaussian process (GP) is employed. Only a limited number of "inducible points" are maintained to approximate the flow field distribution, reducing computational complexity and enabling high-frequency real-time online learning. Secondly, uncertainty perception and opportunity constraints are introduced. The controller not only utilizes the mean of the GP prediction for feedforward compensation but also dynamically adjusts the safety constraint boundary using the variance of the GP prediction (cognitive uncertainty). When facing unknown strong turbulence (large variance), the controller automatically shrinks the feasible region and adopts a conservative strategy; when in a stable flow field (small variance), an aggressive strategy is adopted. This mechanism effectively solves the system oscillation or instability problem caused by model mismatch in traditional NMPC.
[0109] 3. The microsecond-level weighted tight coupling mechanism between neuromorphic vision and inertial measurement:
[0110] This invention uses the microsecond-level timestamp of the event camera as the master clock to perform hard-triggered synchronization of IMU data, eliminating time offset between sensors. Secondly, within the tightly coupled optimization window, when constructing the joint objective function, attention scores are used to adaptively weight the visual reprojection error term. This mechanism allows the system to automatically "ignore" pixels obscured by bubbles at the optimization level, using only high-confidence feature points to constrain the zero-bias drift of the IMU. This achieves strong robustness of the sensing system, maintaining tracking even under conditions of high-speed jitter caused by strong underwater turbulence and bubble obstruction.
[0111] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.
Claims
1. An underwater robot control method based on event vision and uncertainty NMPC, characterized by: The underwater robot is equipped with an inertial navigation system and an event camera. For the image acquired by the event camera, if the logarithmic change in light intensity of a pixel in the image exceeds the event polarity contrast threshold, then the pixel is taken as a valid visual feature point. The position vector, linear velocity vector, and angular velocity vector of the underwater robot are obtained through the inertial navigation system. Based on the attitude vector of the underwater robot at the previous moment, the prior attitude vector of the underwater robot at the current moment is calculated and projected onto the pixel plane. The difference between the prior attitude vector and the pixel coordinates of each effective visual feature point is calculated to construct the reprojection residual vector. The position error components and attitude error components of the event camera are taken as the objectives to be solved. Based on the reprojection residual vector, a nonlinear optimization objective function is constructed. The optimal position error components and attitude error components of the event camera that minimize the value of the nonlinear optimization objective function are solved. The prior attitude vector is corrected to obtain the attitude vector of the underwater robot at the current moment. The position vector, linear velocity vector, attitude vector of the underwater robot, and the zero bias vectors of the accelerometer and gyroscope in the inertial navigation system are used to construct the state vector. Based on the hydrodynamic model of the underwater robot, the flow field disturbance vector experienced by the underwater robot at the current moment is calculated and stored in the historical sliding dataset. Based on the state vector of the underwater robot at the current moment, the posterior distribution is calculated using the kernel function vector to obtain the predicted disturbance mean vector and the predicted disturbance variance vector. The predicted disturbance mean vector is substituted into the state transition equation as a feedforward disturbance term, obstacle state information is introduced as a constraint, and the predicted disturbance variance vector is substituted into the safe collision avoidance distance constraint equation. The tracking control problem is transformed into a rolling time domain optimization problem, and a cost function is constructed. The optimal predictive control sequence is obtained by solving the cost function, and the optimal control torque vector of the underwater robot at the current moment is determined.
2. The underwater robot control method based on event vision and uncertainty NMPC according to claim 1, characterized in that: For images acquired by the event camera, if the logarithmic change in the light intensity of pixels in the image exceeds the event polarity contrast threshold, specifically: in, for pixel coordinates in the image at any time The light intensity at that location; The contrast threshold; pixel coordinates The polarity of events at that location .
3. The underwater robot control method based on event vision and uncertainty NMPC according to claim 1, characterized in that: The underwater robot's attitude vector from the previous moment is used as a reference. Calculate the prior attitude vector of the underwater robot at the current moment. : in, for The angular velocity vector of the underwater robot is constantly obtained through the inertial navigation system; This is the zero bias vector of the gyroscope in the inertial navigation system; The prior attitude vector of the underwater robot at the current moment. Project the vector onto the pixel plane and subtract it from the pixel coordinates of each effective visual feature point to construct the reprojection residual vector. : in, for Time of the first Pixel coordinates of effective visual feature points , for The number of valid visual feature points at any given time; This is the intrinsic parameter matrix of the event camera; The difference in the installation rotation angle of the event camera; The installation translation distance of the event camera relative to the inertial navigation system; According to the first Pixel coordinates of effective visual feature points Position error components of the event camera With attitude error components The calculated first The position vector of effective visual feature points; for The position vector of the underwater robot is obtained in real time through the inertial navigation system; This is the homogeneous coordinate normalized projection function.
4. The underwater robot control method based on event vision and uncertainty NMPC according to claim 3, characterized in that: The position error component of the event camera With attitude error components As the objective to be solved, based on the reprojected residual vector Construct a nonlinear optimization objective function: in, For a priori residuals; for Time of the first Weight scalars of effective visual feature points; For the first The measurement noise covariance matrix of effective visual feature points; Find the optimal position error components of the event camera that minimize the nonlinear optimization objective function. With attitude error components ,Right now ; For the prior attitude vector After correction, the attitude vector of the underwater robot at the current moment is obtained. ; 。 5. The underwater robot control method based on event vision and uncertainty NMPC according to claim 4, characterized in that: The Time of the first Weight scalar of effective visual feature points The method for obtaining it is as follows: In the Calculate the covariance matrix of the event distribution within the spatiotemporal neighborhood of effective visual feature points. And calculate the covariance matrix of the event distribution. eigenvalues and , Then calculate the coherence index. : The coherence index is activated by the Sigmoid activation function. Mapped to weight scalar ; in, This is the gain coefficient; An empirical threshold for distinguishing between bubbles and static objects.
6. The underwater robot control method based on event vision and uncertainty NMPC according to claim 4, characterized in that: The position vector of the underwater robot Linear velocity vector Attitude vector and the zero bias vector of the accelerometer in the inertial navigation system With the zero bias vector of the gyroscope Construct as a state vector ; Based on the hydrodynamic model of the underwater robot, calculate the flow field disturbance vector experienced by the underwater robot at the current moment. ; in, Here is the inertial matrix of the underwater robot; Add a Coriolis force matrix for the underwater robot; Here is the damping matrix for the underwater robot; This represents the static restoring force and torque matrix of an underwater robot. for The control torque vector of the underwater robot at any given time; Will Store in the historical sliding dataset, i.e. Based on the current The state vector of the underwater robot at any time The posterior distribution is calculated using the kernel function vector to obtain the predicted perturbation mean vector. With the predicted perturbation variance vector ; in, For the present The state vector of the underwater robot at any time The prior autocovariance scalar; For the front The state vector of the underwater robot at any time Compared with historical sliding datasets The cross-covariance vector between them; For historical sliding datasets The prior autocovariance vector; It is a unit array.
7. The underwater robot control method based on event vision and uncertainty NMPC according to claim 6, characterized in that: The predicted perturbation mean vector Substituting the feedforward perturbation term into the state transition equation, and introducing obstacle state information as a constraint, the predicted perturbation variance vector is used. Substituting the safety collision avoidance distance constraint equation, the tracking control problem is transformed into a rolling time-domain optimization problem, and the cost function is constructed as follows: in, and This is the weight matrix; Set the prediction range and control range Set predictive control sequence and predicted output sequence : The constraints are: (1) , , ; and The minimum and maximum allowable control torques; , ; (2) , , This is the state transition function. The perturbation input mapping matrix; (3) Introduce obstacle state information as a constraint condition, and predict the perturbation variance vector. Substituting into the safety collision distance constraint equation, , The state vector of the underwater robot is Distance to the nearest obstacle To maintain a safe distance for underwater robots, For risk coefficient, This represents the computation of the trace of a matrix; The optimal predictive control sequence is obtained by solving the cost function. ; Find the optimal predictive control sequence The first item As the optimal control torque vector of the underwater robot at the current moment. ,Right now .
8. A computer device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that: The processor executes the computer program to implement the steps of the method according to any one of claims 1 to 7.
9. A computer-readable storage medium storing computer instructions thereon, characterized in that: When executed by a processor, the computer instructions implement the steps of the method according to any one of claims 1 to 7.
10. A computer program product, comprising a computer program, characterized in that: When executed by a processor, the computer program implements the steps of the method according to any one of claims 1 to 7.