Intelligent closed-loop control system and method for underwater harvesting robot

Abstract

This invention discloses an intelligent closed-loop control system and method for an underwater harvesting robot, belonging to the field of intelligent control technology for underwater robots. The system constructs a three-layer integrated architecture comprising a perception and navigation layer, a control system layer, and a mechanical execution layer. It achieves bidirectional full-duplex communication via a high-speed industrial communication bus and forms an intelligent closed-loop control system for the entire harvesting process based on a unified spatiotemporal reference with hardware synchronization. The method includes six core steps: perception initialization and data acquisition, multi-source data fusion and path planning, adaptive motion control, target harvesting control, operation result feedback and comparison, closed-loop correction, and operation judgment. This invention achieves closed-loop feedback and adaptive adjustment throughout the entire process, from global path planning and robot motion control to end-point harvesting operations. It solves the core pain points of existing technologies, such as poor coordination, weak anti-interference ability, incomplete closed-loop coverage, and poor environmental adaptability. It can be widely applied to scenarios such as seafood harvesting, underwater salvage, and emergency rescue.

Description

Technical Field

[0001] This invention belongs to the field of intelligent control technology for underwater robots, specifically relating to an intelligent closed-loop control system and method for an underwater harvesting robot. It can be widely applied to various underwater operation scenarios such as harvesting marine delicacies in shallow sea aquaculture areas, harvesting wild aquatic products in deep sea rocky reef areas, harvesting live fish and shrimp from underwater aquaculture cages, emergency salvage of underwater equipment, and underwater emergency rescue. Background Technology

[0002] With the rapid development of the marine economy, the demand for underwater aquaculture, marine biological resource development, and underwater engineering operations continues to grow. Underwater harvesting is a core component of these operations, and traditionally, it relies mainly on manual diving. Manual diving harvesting has several unavoidable drawbacks: the operating depth is limited by human physiological limits, typically not exceeding 40 meters; the operating environment is highly risky, with underwater currents, low temperatures, high pressure, and complex terrain easily triggering diving accidents; the operating efficiency is low, as the duration of manual diving operations is limited, and prolonged operations lead to a significant decrease in harvesting accuracy; furthermore, the complex environments of nearshore aquaculture areas and deep-sea reef areas further exacerbate the difficulty and risks of manual harvesting.

[0003] Underwater harvesting robots, as core equipment to replace manual diving operations, can perform autonomous navigation, target identification, and precise harvesting in complex underwater environments, effectively solving the pain points of manual harvesting. In recent years, they have become a research hotspot in the field of marine equipment. Currently, extensive research has been conducted both domestically and internationally on control technologies for underwater harvesting robots, resulting in several existing technical solutions. However, these existing technologies still have many technical bottlenecks and shortcomings in practical engineering applications, failing to meet the demands for high-precision, high-reliability, and highly intelligent harvesting operations in complex and dynamic underwater environments. Specifically, these shortcomings manifest in the following aspects:

[0004] First, the existing control architectures of underwater harvesting robots generally suffer from decentralization and fragmentation, failing to form an integrated closed-loop control system. In current technologies, sensing units, control units, and execution units are mostly designed independently and controlled in a decentralized manner. Each unit only achieves one-way command issuance and data upload, lacking a two-way high-speed communication mechanism and a unified spatiotemporal reference for hardware synchronization. This results in the inability to synchronize and coordinate the operational sequence of each unit. For example, motion control and harvesting control are executed by two independent controllers. When the robotic arm performs harvesting actions, the motion control unit cannot perceive the robotic arm's motion status in real time and cannot synchronously compensate for the body's posture disturbances caused by the robotic arm's movements. This leads to position drift and posture deviation of the robot body during harvesting, directly reducing the accuracy and stability of the end-effector harvesting. Furthermore, most existing control architectures only achieve semi-closed-loop control for a single link, such as building a closed loop only for the robot's motion control, failing to cover the entire process from environmental perception, path planning, motion control, harvesting operations to result feedback and correction, and thus failing to achieve intelligent closed-loop adjustment of the entire harvesting operation process.

[0005] Secondly, existing underwater harvesting robots suffer from limitations in their perception systems, including reliance on a single sensor and low integration, resulting in insufficient adaptability and reliability in complex underwater environments. Most current technologies rely on a single visual sensor for environmental perception and target recognition. However, underwater environments are generally characterized by turbidity, low visibility, and insufficient light, significantly degrading the imaging quality of visual sensors or even rendering them completely ineffective, thus preventing the robot from completing positioning, navigation, and target recognition. While some solutions incorporate both sonar and visual sensors, they fail to achieve spatiotemporal registration and tight coupling fusion of multimodal sensor data. Data from different sensors exhibit temporal and spatial discrepancies, hindering complementary advantages and leading to insufficient accuracy in positioning, navigation, and target recognition. Furthermore, prolonged operation can result in accumulated errors, failing to meet the high-precision operational requirements of complex underwater environments. Simultaneously, most existing perception systems focus only on macroscopic perception of the environment and the robot itself, neglecting the operational status perception of the end effector. For example, the absence of a gripping force feedback sensor at the gripper end prevents refined perception and closed-loop control of the grasping process, making it highly susceptible to target damage or insufficient gripping force leading to target detachment.

[0006] Third, existing control strategies for underwater harvesting robots have weak anti-disturbance capabilities and poor adaptability to complex dynamic underwater environments. The dynamic models of underwater robots are characterized by strong nonlinearity, strong coupling, and multiple disturbances. Dynamic factors such as underwater current disturbances, target displacement, and changes in robot posture all significantly affect control accuracy. Most existing technologies employ traditional fixed-parameter PID control strategies, which can only achieve basic control functions in steady-state environments and cannot adaptively fit the nonlinear disturbance characteristics of the underwater environment. When water flow velocity changes, the target shifts, or the robot's posture is disturbed, the fixed-parameter PID cannot adjust control parameters in real time, leading to slower system response, decreased control accuracy, and even control oscillations, making it impossible to maintain stable hovering and precise harvesting operations. While some technical solutions introduce intelligent control algorithms, they are only applied to a single motion control stage and are not deeply integrated with the harvesting control stage. They also fail to achieve full-dimensional adaptive compensation for the three core dynamic factors: water flow disturbance, target displacement, and robot posture disturbance, thus failing to solve the control robustness problem in complex dynamic environments.

[0007] Fourth, existing underwater harvesting robots have low levels of intelligence and lack a closed-loop feedback and adaptive correction mechanism throughout the entire process. In current technologies, most underwater harvesting robots operate in a pre-planned open-loop mode, meaning the navigation path and harvesting actions are preset before operation. During operation, only preset commands are executed, without real-time feedback or comparison of the results, and they cannot automatically correct control commands based on operational deviations. When environmental changes, target displacement, or body attitude disturbances occur during operation, causing deviations in a single harvesting operation, existing systems cannot automatically identify the deviations and correct subsequent control strategies. Manual remote intervention is necessary to replan the path and harvesting actions, significantly reducing operational efficiency and preventing unmanned autonomous operation. While some technical solutions include feedback mechanisms, they only provide feedback on the execution status of a single action, failing to cover the entire process from global path planning to the final harvesting operation. This lack of closed-loop feedback and adaptive adjustment across the entire operational chain means the level of intelligence cannot meet the needs of practical engineering applications.

[0008] Fifth, existing underwater harvesting robots suffer from insufficient operational safety and reliability, lacking multi-task collaborative control and a comprehensive safety protection mechanism. In current technologies, motion control and harvesting control mostly operate independently, without a collaborative control mechanism. This prevents the synchronous coordination of the two modules' execution sequences based on a unified spatiotemporal reference, leading to a disconnect between the robotic arm's harvesting operations and the robot's posture stabilization control. This significantly increases the risk of safety hazards such as posture instability and collisions with obstacles. Furthermore, existing systems lack comprehensive operational status monitoring and emergency protection mechanisms. They cannot compare operational results with expected states in real time. When operational deviations exceed controllable limits, system malfunctions, or collision risks exist, emergency protection procedures cannot be triggered promptly. This can easily result in serious accidents such as robotic arm jamming, gripper overload damage, target loss, and robot collisions and overturning. This drastically reduces the system's operational reliability and lifespan, making it unsuitable for long-term continuous operation in complex underwater environments.

[0009] In summary, existing underwater harvesting robot control technologies have significant technical bottlenecks in terms of control architecture, perception system, control strategy, closed-loop mechanism, and safety protection. They cannot meet the requirements of high-precision, high-reliability, and highly intelligent harvesting operations in complex and dynamic underwater environments. There is an urgent need to develop a new intelligent closed-loop control system and method to solve the core pain points of existing technologies. Summary of the Invention

[0010] The purpose of this invention is to overcome the aforementioned defects in existing underwater harvesting robot control technologies and provide an intelligent closed-loop control system and method for underwater harvesting robots. This invention addresses the core technical problems of existing technologies, such as dispersed control architecture with poor coordination, single perception system with low reliability, weak anti-interference capability of control strategies, incomplete closed-loop coverage with low intelligence, and insufficient safety protection with poor reliability. It achieves intelligent closed-loop control of the entire underwater harvesting operation process, significantly improving the operational accuracy, success rate, environmental adaptability, and safety of underwater harvesting robots in complex and dynamic underwater environments.

[0011] To achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0012] An intelligent closed-loop control system for an underwater harvesting robot includes an underwater harvesting robot body, and a sensing unit, a control unit, and an execution unit mounted on the robot body. The sensing unit, control unit, and execution unit are respectively constructed as a sensing and navigation layer, a control system layer, and a mechanical execution layer. The three layers are connected in pairs through a high-speed industrial communication bus for bidirectional full-duplex communication, and together they form an intelligent closed-loop control system for the entire harvesting operation process.

[0013] The perception and navigation layer is used to collect multimodal sensing data of the underwater working environment and working status, upload the multimodal sensing data to the control system layer in real time, and after the mechanical execution layer completes a single working action, collect the working result feedback data and send it back to the control system layer.

[0014] The control system layer is mounted in an industrial-grade high-performance main controller inside the pressure-resistant sealed cabin of the robot body. It has built-in a fusion positioning and navigation module, a motion control module, a capture control module, and a closed-loop feedback correction module. Each module interacts bidirectionally with the main controller and is connected to each other through an internal high-speed bus.

[0015] The control system layer is used to receive multimodal sensing data uploaded by the perception and navigation layer, complete the real-time pose calculation and global operation path planning of the robot through the fusion positioning and navigation module, generate motion control commands for the robot body based on the global operation path and real-time posture feedback through the motion control module, generate capture control commands for the robotic arm and flexible gripper based on the target recognition and positioning results through the capture control module, and send all control commands to the mechanical execution layer through the high-speed industrial communication bus.

[0016] The control system layer is based on a unified spatiotemporal reference triggered by hardware synchronization. It synchronously coordinates the operation sequence of the motion control module, the acquisition control module and the closed-loop feedback correction module to achieve integrated collaborative control of multiple tasks.

[0017] The mechanical execution layer includes a bionic thruster unit that is communicatively connected to the motion control module, a robotic arm execution unit and a flexible gripper unit that are communicatively connected to the capture control module, and is used to receive control commands issued by the control system layer to complete the robot's navigation, hovering and target capture actions.

[0018] The closed-loop feedback correction module is used to compare the operation result feedback data returned by the perception and navigation layer with the preset expected operation state in real time, calculate the state deviation, and correct the motion control command and the capture control command in real time based on the state deviation and the real-time environmental disturbance parameters. It completes the adaptive compensation for three types of dynamic factors: water flow interference, target displacement, and body attitude disturbance, and realizes the closed-loop feedback and adaptive adjustment of the whole process from global path planning and body motion control to end capture operation.

[0019] Furthermore, the multimodal sensor group of the perception and navigation layer includes a forward-looking imaging sonar, a binocular underwater vision camera, an inertial measurement unit, a depth pressure sensor, and a clamping force feedback sensor fixedly installed on the flexible gripper's gripping end; the clamping force feedback sensor is used to collect clamping force data in real time during the target grasping process, and transmit the clamping force data back to the acquisition control module and closed-loop feedback correction module of the control system layer in real time.

[0020] Furthermore, both the motion control module and the acquisition control module of the control system layer adopt a control strategy that integrates classical control and intelligent control. The classical control uses adaptive PID closed-loop control to achieve stable execution of basic actions, while the intelligent control uses a fuzzy neural network anti-disturbance algorithm to achieve adaptive compensation for three types of complex dynamic environments: underwater ocean current disturbances, visibility changes, and target movement.

[0021] Furthermore, the closed-loop feedback correction module incorporates a deviation comparison unit, an adaptive parameter tuning unit, and a command correction unit that are sequentially bidirectionally connected. The deviation comparison unit is used to calculate the real-time state deviation between the operation result feedback data and the preset expected operation state. The adaptive parameter tuning unit is used to adjust the control tuning parameters of the motion control algorithm and the acquisition control algorithm based on the state deviation and the real-time environmental disturbance parameters collected by the perception and navigation layer. The command correction unit is used to generate corrected motion control commands and acquisition control commands based on the adjusted tuning parameters.

[0022] Furthermore, during the entire process of the robotic arm execution unit performing the harvesting operation, the control system layer dynamically adjusts the thrust output of the bionic thruster unit based on the real-time pose deviation of the robot body through the motion control module, compensates for the body posture disturbance caused by the robotic arm movement in real time, maintains the dynamic hovering stability of the robot body, and realizes multi-task integrated collaborative control of body motion control and end-capture operation.

[0023] This invention also provides an intelligent closed-loop control method for an underwater harvesting robot, implemented based on the intelligent closed-loop control system of the underwater harvesting robot, comprising the following steps:

[0024] S1. Perception Initialization and Data Acquisition: After the system is powered on, the perception and navigation layer completes the calibration and initialization of the multimodal sensor group. Through the synchronous trigger signal of the main controller, the multimodal sensor group is controlled to synchronously acquire real-time sensing data of the underwater working environment, including target point cloud data of forward-looking imaging sonar, environmental image data of binocular underwater vision camera, attitude data of inertial measurement unit, and pressure data of depth sensor. After performing spatiotemporal registration on all sensing data, it is uploaded to the control system layer in real time.

[0025] S2. Multi-source data fusion and path planning: The control system layer uses fusion positioning and navigation algorithms to perform tight-coupled fusion calculation on the spatiotemporally registered multimodal sensor data, construct a three-dimensional grid map of the underwater operation environment, complete the robot's six-degree-of-freedom real-time pose calculation, and generate a global collision-free operation path and target navigation plan based on the preset operation target.

[0026] S3. Adaptive motion control and attitude stabilization: The motion control algorithm of the control system layer is based on the generated global operation path and combined with real-time attitude feedback data. It calculates the multi-channel thrust drive parameters of the bionic thruster unit, generates motion control commands and sends them to the mechanical execution layer to control the robot to complete stable navigation and dynamic precise hovering.

[0027] S4. Target recognition and capture operation control: When the perception and navigation layer recognizes that the target has entered the preset operation range, the capture control algorithm of the control system layer plans the motion trajectory and grasping sequence of the end of the robotic arm based on visual recognition and real-time positioning results. Combined with the clamping force feedback mechanism, it generates capture control commands and sends them to the mechanical execution layer to control the robotic arm execution unit and flexible gripper unit to complete the target capture operation.

[0028] S5. Feedback on work results and comparison of status: After the mechanical execution layer completes a single work action, the perception and navigation layer immediately collects feedback data on the current work environment, robot pose and work results, and transmits the feedback data back to the control system layer in real time. The control system layer compares the received feedback data with the preset expected work status in real time and calculates the status deviation.

[0029] S6. Closed-loop correction and operation process determination: Based on the calculated state deviation, the control system layer determines whether the deviation is within the preset correctable threshold range. If the deviation is within the correctable threshold range, the subsequent motion control commands and harvesting control commands are corrected in real time through the closed-loop feedback correction algorithm to compensate for the uncertainties caused by water flow interference, target displacement, and body attitude disturbance. The system then returns to step S1 to re-execute the entire closed-loop control process based on the corrected control target. If the deviation exceeds the correctable threshold range, the emergency protection procedure is triggered. If there is no deviation, the harvesting operation is determined to be completed, and the control process ends.

[0030] Furthermore, in step S2, the fusion positioning and navigation algorithm adopts a tightly coupled data fusion strategy based on factor graph optimization, which fuses sonar target point cloud data, visual environment image data, inertial measurement unit attitude data and depth sensor pressure data to construct a three-dimensional grid map of the underwater operation environment, and simultaneously completes the robot's six-degree-of-freedom real-time pose calculation and global collision-free operation path planning.

[0031] Furthermore, in step S4, the clamping force feedback mechanism integrated into the capture control algorithm is as follows: during the process of the flexible gripper clamping the capture target, the clamping force data is collected in real time through the clamping force feedback sensor. When the clamping force reaches the preset target safety clamping threshold, the clamping feed action of the gripper is stopped immediately to maintain the current clamping force; when the clamping force exceeds the preset safety overload threshold, the gripper is driven to release to the safety clamping range in the reverse direction, so as to realize flexible and non-destructive grasping and overload protection of the capture target.

[0032] Furthermore, in step S5, the preset expected operating state includes the expected pose of the robot body with six degrees of freedom, the expected operating pose of the robotic arm end effector, the expected grasping state of the target and the expected clamping force range; the state deviation includes one or more combinations of position deviation, posture deviation, clamping force deviation and target displacement deviation.

[0033] Furthermore, the emergency protection procedure triggered in step S6 specifically includes: the control system layer immediately stopping the harvesting operation, controlling the flexible gripper to maintain the current gripping state, and simultaneously controlling the robot body to return to the preset safe point through the motion control module, and uploading the fault data, deviation data and operation log to the remote monitoring terminal in a synchronized manner; the emergency protection procedure is in a real-time monitoring state throughout the closed-loop control process, and the triggering conditions include operation deviation exceeding the threshold, sensor failure, thruster failure, robotic arm jamming, communication interruption, collision risk, and depth exceeding the range, and it will be executed immediately once any triggering condition is met.

[0034] Compared with the prior art, the present invention has the following outstanding technical advantages:

[0035] First, a three-layer integrated intelligent closed-loop control architecture was constructed, fundamentally solving the core pain points of existing technical architectures, such as fragmentation, poor coordination, and incomplete closed-loop coverage.

[0036] This invention reconstructs the traditionally dispersed sensing, control, and execution units into a three-layer integrated architecture comprising a sensing and navigation layer, a control system layer, and a mechanical execution layer. A high-speed industrial communication bus enables bidirectional full-duplex communication between these three layers. Simultaneously, a unified spatiotemporal reference is established through a hardware synchronization triggering mechanism, achieving precise synchronization of the runtime sequence across the three layers and breaking down the barriers of independent unit operation and timing discrepancies in existing technologies. Based on this architecture, this invention constructs a closed-loop control system covering the entire harvesting operation process, from environmental perception to path planning, motion control, harvesting operations, result feedback, and command correction. Unlike existing technologies that only address a single aspect of semi-closed-loop control, this invention achieves a complete closed-loop feedback chain from global path planning and body motion control to the final harvesting operation. This architecture achieves deep integration of perception, decision-making, and execution. The control system layer can synchronously coordinate the operation of motion control, harvesting control, and closed-loop correction modules based on a unified spatiotemporal reference, realizing multi-task integrated collaborative control. It solves the core problems of existing underwater harvesting robots, such as decentralized control, slow response, and poor coordination, from the architectural level, and greatly improves the system's overall performance, response speed, and collaborative control capabilities, laying the architectural foundation for high-precision and highly intelligent harvesting operations.

[0037] Second, a navigation and positioning system integrating multimodal perception and tight coupling was established, which significantly improved the perception reliability and positioning and navigation accuracy in complex underwater environments.

[0038] The perception and navigation layer of this invention constructs a multimodal sensor array, integrating a forward-looking imaging sonar, a binocular underwater vision camera, a high-precision inertial measurement unit (IMU), a depth and pressure sensor, and a gripping force feedback sensor. This forms a comprehensive perception system encompassing far-field, near-field, macroscopic, and microscopic dimensions, distinguishing it from the single-sensor perception modes of existing technologies and enabling complementary advantages of different sensors. Specifically, the forward-looking imaging sonar enables long-range environmental perception and target recognition in turbid water and low-visibility environments; the binocular vision camera provides high-precision target imaging and detail recognition in the near field; the IMU and depth sensor provide high-frequency body attitude and position data; and the gripping force feedback sensor enables refined state perception during the end-effector grasping process. Furthermore, this invention employs a tightly coupled data fusion strategy based on factor graph optimization. After spatiotemporal registration of all multimodal sensor data, tightly coupled fusion calculation is performed under a unified spatiotemporal reference, eliminating temporal and spatial biases between different sensor data and avoiding the cumulative errors of a single sensor. This allows for the construction of a high-precision 3D grid map in complex underwater environments, enabling real-time six-degree-of-freedom pose calculation for the robot. This sensing system can adapt to complex underwater environments with varying visibility and terrain, completely solving the problems of single visual sensors failing in turbid water, insufficient positioning accuracy, and large cumulative errors in existing technologies. At the same time, it incorporates end-point operation status sensing into the overall sensing system, providing comprehensive, accurate, and reliable sensing data support for closed-loop control throughout the entire process, and significantly improving the system's environmental adaptability and sensing reliability.

[0039] Third, an adaptive disturbance rejection control strategy that integrates classical and intelligent control is adopted, which significantly improves the system's robustness and operational accuracy in complex and dynamic underwater environments.

[0040] The motion control module and acquisition control module of this invention both adopt a control strategy that integrates adaptive PID closed-loop control and fuzzy neural network anti-disturbance algorithm. Unlike the fixed-parameter PID control of existing technologies, this approach achieves full-dimensional adaptive compensation for complex dynamic underwater environments. The adaptive PID closed-loop control, as the basic control layer, enables rapid response and steady-state precision control of the robot's basic navigation, hovering, and robotic arm movements, ensuring the stability and reliability of the system's basic actions. The fuzzy neural network anti-disturbance algorithm, as the intelligent compensation layer, takes the robot's attitude deviation, deviation rate of change, and real-time environmental disturbance parameters as inputs, and the correction amount of the control parameters as outputs. Without establishing a precise underwater robot nonlinear dynamic model, it can fit the nonlinear characteristics of three complex dynamic environments—underwater current disturbance, visibility changes, and target movement—in real time, adaptively adjusting the tuning parameters of the PID control to achieve full-dimensional adaptive compensation for three core dynamic factors: water flow interference, target displacement, and body attitude disturbance. This integrated control strategy retains the advantages of classic PID control, such as fast response speed, high steady-state accuracy, and good stability, while solving the problems of poor adaptability and weak anti-disturbance capability of traditional PID control to nonlinear, strongly coupled, and strongly disturbed underwater dynamic systems through intelligent algorithms. It can maintain the robot's stable navigation, precise hovering, and high-precision end-capture control in complex dynamic underwater environments, greatly improving the system's control robustness and environmental adaptability.

[0041] Fourth, a closed-loop feedback and adaptive correction mechanism was established throughout the entire process, enabling unmanned autonomous control of harvesting operations and significantly improving the intelligence level and success rate of the operation.

[0042] This invention incorporates a closed-loop feedback correction module at the control system layer, constructing a full-process closed-loop feedback and adaptive correction mechanism of "deviation comparison - parameter tuning - command correction." This differs from the open-loop operation mode and single-stage feedback of existing technologies, achieving closed-loop feedback and adaptive adjustment throughout the entire acquisition process. After a single operation is completed at the mechanical execution layer, the perception and navigation layer immediately collects comprehensive operation result feedback data, including the current working environment, robot pose, robotic arm pose, gripping force state, and target grasping state, and transmits this data back to the control system layer. The closed-loop feedback correction module compares the feedback data with the preset expected working state in real time, accurately calculating state deviations such as position deviation, posture deviation, gripping force deviation, and target displacement deviation. Subsequently, based on the deviations and real-time environmental disturbance parameters, the adaptive parameter tuning unit adjusts the tuning parameters of the motion control and acquisition control algorithms, and the command correction unit generates corrected control commands to compensate for various uncertainties during the operation process in real time. When the operational deviation is within the correctable threshold range, the system automatically re-executes the closed-loop control process based on the corrected control target, completing the correction and secondary operation without manual intervention. Only when the deviation exceeds the correctable threshold will the emergency protection procedure be triggered. This full-process closed-loop feedback mechanism completely solves the problems of existing technologies, such as reliance on manual intervention after a single operational failure, low intelligence, and low operational efficiency. It achieves fully unmanned autonomous control of the entire harvesting operation, adaptively responding to various dynamic changes during the operation, and significantly improving the intelligence and success rate of the operation.

[0043] Fifth, a multi-task integrated collaborative control and full-process safety protection mechanism was constructed, which significantly improved the system's operational reliability and security.

[0044] The control system layer of this invention is based on a unified spatiotemporal reference with hardware synchronization, realizing deep collaboration between the motion control module and the harvesting control module. During the entire process of the robotic arm execution unit performing the harvesting operation, the motion control module will collect the pose deviation of the robot body in real time, dynamically adjust the thrust output of the bionic thruster unit, and compensate for the body posture disturbance caused by the robotic arm movement in real time, so as to maintain the dynamic hovering stability of the robot body. This realizes the multi-task integrated collaborative control of body motion control and end-effector harvesting operation, completely solving the problem of body posture instability and harvesting accuracy reduction caused by robotic arm movement in the prior art, and fundamentally avoiding the collision and overturning risks caused by body posture disturbance. Meanwhile, this invention establishes a full-process status monitoring and emergency protection mechanism. The emergency protection program is in real-time monitoring throughout the closed-loop control process, capable of monitoring operational deviations, system operating status, and environmental risks in real time. It covers seven trigger scenarios: operational deviation exceeding thresholds, sensor failure, thruster malfunction, robotic arm jamming, communication interruption, collision risk, and depth exceeding range. Once any trigger condition is met, the emergency protection program is immediately activated, stopping the harvesting operation, controlling the flexible gripper to maintain the current gripping state to prevent target loss, and simultaneously controlling the robot body to automatically return to a preset safe point. Fault data, deviation data, and operation logs are synchronously uploaded to the remote monitoring terminal. In addition, the gripping force feedback mechanism integrated into the harvesting control algorithm enables flexible and non-destructive grasping of the harvested target and overload protection, avoiding damage to the harvested target and protecting the end effector from overload damage. This collaborative control and safety protection mechanism realizes risk prevention and safety protection throughout the entire harvesting operation process, significantly improving the system's operational reliability, equipment lifespan, and operational safety, and enabling long-term continuous operation in complex underwater environments. Attached Figure Description

[0045] Figure 1 This is a block diagram of the overall architecture of the intelligent closed-loop control system for the underwater harvesting robot of the present invention.

[0046] Figure 2 This is a block diagram showing the internal module composition and communication connections of the control system layer of the present invention;

[0047] Figure 3 This is a block diagram showing the internal unit composition and data flow of the closed-loop feedback correction module of the present invention;

[0048] Figure 4 This is an overall flowchart of the intelligent closed-loop control method for underwater harvesting robots of the present invention;

[0049] Figure 5 This is a schematic diagram illustrating the principle of the closed-loop feedback control throughout the entire process of this invention.

[0050] Figure 6 This is a flowchart illustrating the execution process of the emergency protection procedure of this invention.

[0051] In the diagram: 1-Perception and navigation layer; 2-Control system layer; 3-Mechanical execution layer; 4-High-speed industrial communication bus; 5-Underwater harvesting robot body; 101-Forward-looking imaging sonar; 102-Binocular underwater vision camera; 103-Inertial measurement unit; 104-Depth pressure sensor; 105-Grip force feedback sensor; 201-Industrial-grade high-performance main controller; 202-Fusion positioning and navigation module; 203-Motion control module; 204-Harvesting control module; 205-Closed-loop feedback correction module; 2051- Deviation comparison unit; 2052-Adaptive parameter tuning unit; 2053-Command correction unit; 301-Bionic thruster unit; 302-Robotic arm execution unit; 303-Flexible gripper unit. Detailed Implementation

[0052] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. The described embodiments are merely preferred embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0053] All components used in this invention are commercially available general-purpose industrial-grade products, requiring no custom development and possessing the feasibility of mass production; all software algorithms of this invention are implemented based on existing industrial-grade real-time operating systems and development platforms, and those skilled in the art can repeat this invention without creative effort based on the content disclosed in this specification.

[0054] In an embodiment of the present invention, an intelligent closed-loop control system for an underwater harvesting robot has the following overall architecture: Figure 1 As shown, the system includes an underwater harvesting robot body 5, and a perception and navigation layer 1, a control system layer 2, and a mechanical execution layer 3 mounted on the robot body 5. These three layers are connected in pairs via a high-speed industrial communication bus 4, operating synchronously based on a unified spatiotemporal reference triggered by hardware synchronization. Together, they form an intelligent closed-loop control system for the entire harvesting operation. The output of the perception and navigation layer 1 is connected to the input of the control system layer 2 via the high-speed industrial communication bus 4. The output of the control system layer 2 is connected to the input of the mechanical execution layer 3 via the high-speed industrial communication bus 4. The execution feedback of the mechanical execution layer 3 and the operation result feedback of the perception and navigation layer 1 are both connected to the feedback input of the control system layer 2 via the high-speed industrial communication bus 4, forming a complete closed-loop feedback circuit.

[0055] The multimodal sensor group of the perception and navigation layer 1 includes a forward-looking imaging sonar 101 and a binocular underwater vision camera 102 fixedly installed at the front centerline of the robot body 5; an inertial measurement unit 103 and a depth and pressure sensor 104 fixedly installed at the center of gravity of the robot body 5; and a clamping force feedback sensor 105 fixedly installed on the inner side of the gripping end of the flexible gripper unit 303. The forward-looking imaging sonar 101 is a multi-beam imaging sonar with a detection range of 0.5-100 meters, a horizontal beam angle of 120°, and a vertical beam angle of 20°, used for long-range underwater environment 3D point cloud data acquisition and target recognition; the binocular underwater vision camera 102 is an IP68 waterproof sealed binocular camera with a resolution of 1920×1080 and a frame rate of 30fps, equipped with an underwater LED supplementary lighting unit with adjustable supplementary lighting power from 0-30W, used for near-field environment image acquisition and target detail recognition; the inertial measurement unit 103 uses a high-precision fiber optic IMU or an industrial-grade MEMS. The IMU (Induction Unit), with a sampling frequency of 1000Hz, an angular velocity measurement range of ±400° / s, and an acceleration measurement range of ±16g, is used to collect high-frequency attitude data of the robot's three-axis acceleration and three-axis angular velocity. The depth pressure sensor 104 is a high-precision piezoresistive pressure sensor with a measurement range of 0-200 meters, an accuracy of ±0.05%FS, and a sampling frequency of 100Hz, used to collect underwater depth data. The clamping force feedback sensor 105 is a multi-point array thin-film pressure sensor with a sampling frequency of 200Hz and a range of 0-200N, used to collect clamping force distribution data in real time during the grasping process.

[0056] The internal modules of control system layer 2 are composed as follows: Figure 2As shown, the industrial-grade high-performance main controller 201 is installed in the pressure-resistant sealed chamber of the robot body 5. The main controller 201 adopts an NXP i.MX 8QuadMax industrial-grade multi-core processor and runs on the Linux PREEMPT_RT real-time operating system. The kernel real-time performance is better than 100μs, and the operating temperature range is -40℃~85℃. It has the ability to resist vibration, resist electromagnetic interference, and operate in a wide temperature range, and can meet the requirements of real-time processing of multi-source data and parallel operation of multi-tasks. The main controller 201 integrates a fusion positioning and navigation module 202, a motion control module 203, a acquisition control module 204, and a closed-loop feedback correction module 205. All four modules interact bidirectionally with the main controller 201 and are connected to each other through the AXI internal high-speed bus, with a communication bandwidth of not less than 10Gbps. Among them, the fusion positioning and navigation module 202 is responsible for the fusion and calculation of multimodal sensor data, the construction of a 3D grid map, the robot pose calculation, and global path planning; the motion control module 203 is responsible for the navigation and hovering control of the robot body and the calculation of the thrust driving parameters of the bionic thruster; the grasping control module 204 is responsible for the grasping operation control of the robotic arm and the flexible gripper, and the planning of the robotic arm's motion trajectory and grasping sequence; the closed-loop feedback correction module 205 is responsible for the feedback comparison of operation results, deviation calculation, and control command correction. Its internal unit composition is as follows: Figure 3 As shown, it includes a deviation comparison unit 2051, an adaptive parameter tuning unit 2052, and an instruction correction unit 2053 that are connected in a bidirectional communication manner.

[0057] The mechanical execution layer 3 consists of a bionic thruster unit 301, a robotic arm execution unit 302, and a flexible gripper unit 303. The bionic thruster unit 301 includes four horizontal vector thrusters and two vertical vector thrusters. The horizontal thrusters are symmetrically installed on the left and right sides of the robot body 5, and the vertical thrusters are symmetrically installed on the top and bottom ends of the robot body 5. All thrusters are driven by waterproof brushless DC motors, with a maximum thrust of 500N per unit, enabling six-degree-of-freedom motion control of the robot, completing forward, backward, heave, lateral, turning, and pitching movements. The robotic arm execution unit 302 adopts a 6-7 degree-of-freedom underwater robotic arm, which is fixedly installed at the front centerline of the robot body 5. It has an IP68 protection rating, a pressure resistance rating of 200 meters, an operating radius of 0.8-1.5 meters, and an end-effector repeatability of ±0.5mm, enabling flexible multi-degree-of-freedom movements in an underwater environment. The flexible gripper unit 303 adopts a pneumatically driven or servo motor driven flexible gripper, which is fixedly installed on the end flange of the robotic arm execution unit 302. The gripping range is 0-200mm, which can adapt to targets of different shapes and materials, achieving flexible and non-destructive grasping.

[0058] The high-speed industrial communication bus 4 adopts EtherCAT or CANopen industrial fieldbus, which features high real-time performance, high reliability, and strong anti-interference capability. The communication rate is not less than 100Mbps, which can realize bidirectional high-speed data transmission between the three layers. The delay of issuing control commands is no more than 1ms, and the delay of uploading sensor data is no more than 1ms, ensuring the real-time performance of closed-loop control.

[0059] Based on the above system, the intelligent closed-loop control method for underwater harvesting robots of the present invention has the following overall process: Figure 4 As shown, the specific steps include:

[0060] S1. Sensing Initialization and Data Acquisition: After the system is powered on, the sensing and navigation layer 1 completes the calibration initialization of the multimodal sensor group, specifically including: room temperature zero bias calibration and temperature compensation calibration of the inertial measurement unit 103, intrinsic and extrinsic parameter calibration of the binocular underwater vision camera 102, sound velocity calibration and parameter calibration of the forward-looking imaging sonar 101, surface zero-point calibration of the depth pressure sensor 104, and range calibration and zero-point calibration of the clamping force feedback sensor 105. After calibration, the main controller 201 outputs a 1kHz hardware synchronous trigger PWM signal to control the multimodal sensor group to synchronously acquire real-time sensing data of the underwater operating environment, including target point cloud data of the forward-looking imaging sonar 101, environmental image data of the binocular underwater vision camera 102, attitude data of the inertial measurement unit 103, and pressure data of the depth pressure sensor 104. The main controller 201 adds a unified timestamp to all sensing data, and after completing spatiotemporal registration, it uploads the data to the control system layer 2 in real time through the high-speed industrial communication bus 4.

[0061] S2. Multi-source data fusion and path planning: The control system layer 2 receives multimodal sensor data after spatiotemporal registration. Through the fusion positioning and navigation module 202's fusion positioning and navigation algorithm, a tightly coupled data fusion strategy based on factor graph optimization is adopted to fuse sonar target point cloud data, visual environment image data, inertial measurement unit attitude data, and depth sensor pressure data to construct a 3D grid map of the underwater working environment. Simultaneously, the robot's six-degree-of-freedom real-time pose calculation is completed. Subsequently, based on the preset working target and working area, combined with the constructed 3D grid map, an improved... The algorithm generates a global collision-free operation path and target navigation plan, while setting the expected pose and control accuracy requirements for path tracking.

[0062] S3. Adaptive Motion Control and Attitude Stabilization: The motion control module 203 of the control system layer 2, based on the generated global operation path and combined with the robot's real-time pose feedback data, calculates the robot's pose and velocity deviations. It employs a motion control algorithm that integrates adaptive PID and fuzzy neural networks to calculate the multi-channel thrust drive parameters of the bionic thruster unit 301, generating motion control commands, which are then sent to the mechanical execution layer 3 via the high-speed industrial communication bus 4. The bionic thruster unit 301 of the mechanical execution layer 3 receives the control commands, adjusts the thrust output of each thruster, and controls the robot to complete stable navigation along the planned path. Simultaneously, it compensates for attitude deviations caused by water flow disturbances in real time. Upon reaching the target point, the robot achieves dynamic and precise hovering, providing a stable operating platform for subsequent harvesting operations.

[0063] S4. Target Recognition and Acquisition Operation Control: When the perception and navigation layer 1 recognizes that the acquisition target has entered the preset operation range, it uses forward-looking imaging sonar 101 and binocular underwater vision camera 102 to fuse positioning, obtain the three-dimensional spatial coordinates and shape feature data of the acquisition target, and uploads it to the control system layer 2 in real time. The acquisition control module 204 of the control system layer 2, based on the visual recognition and real-time positioning results of the target, uses a fifth-order polynomial interpolation algorithm to plan the motion trajectory and grasping sequence of the robotic arm end effector, and generates acquisition control commands in combination with the gripping force feedback mechanism, which are sent to the mechanical execution layer 3 through the high-speed industrial communication bus 4. The robotic arm execution unit 302 and the flexible gripper unit 303 of the mechanical execution layer 3 receive the control commands. The robotic arm execution unit 302 drives the flexible gripper unit 303 to move along the planned trajectory to the acquisition target position, and the flexible gripper unit 303 performs a closing grasping action. During the grasping process, the clamping force feedback sensor 105 collects clamping force data in real time and transmits it back to the capture control module 204. When the clamping force reaches the preset target safety clamping threshold, the clamping and feeding action of the gripper is stopped immediately to maintain the current clamping force. When the clamping force exceeds the preset safety overload threshold, the gripper is driven to release to the safety clamping range in reverse, so as to realize flexible and non-destructive grasping and overload protection of the target.

[0064] S5. Work Result Feedback and State Comparison: After the mechanical execution layer 3 completes a single capture operation, the perception and navigation layer 1 immediately collects the current working environment data, the robot's six-degree-of-freedom pose data, the working pose data of the robotic arm's end effector, the gripping force data of the flexible gripper, and the grasping state data of the target being captured, as work result feedback data. This data is transmitted back to the control system layer 2 in real time via the high-speed industrial communication bus 4. The closed-loop feedback correction module 205 of the control system layer 2 compares the received feedback data with the preset expected work state in real time. The preset expected work state includes the expected pose of the robot's six degrees of freedom, the expected working pose of the robotic arm's end effector, the expected grasping state of the target being captured, and the expected gripping force range. The state deviation is calculated through comparison, including one or more combinations of position deviation, posture deviation, gripping force deviation, and target displacement deviation.

[0065] S6. Closed-Loop Correction and Operation Process Judgment: Based on the calculated state deviation, the control system layer 2 determines whether the deviation is within the preset correctable threshold range. If the deviation is within the correctable threshold range, the closed-loop feedback correction module 205, through the closed-loop feedback correction algorithm, adjusts the control tuning parameters of the motion control algorithm and the harvesting control algorithm in real time based on the state deviation and real-time environmental disturbance parameters, generating corrected motion control commands and harvesting control commands to compensate for uncertainties caused by water flow interference, target displacement, and body attitude disturbances. Then, it returns to step S1, re-executes the entire closed-loop control process based on the corrected control target, and completes the subsequent harvesting operation. If the deviation exceeds the correctable threshold range, an emergency protection procedure is immediately triggered. The execution flow of the emergency protection procedure is as follows: Figure 6 As shown, control system layer 2 immediately stops the harvesting operation, controls the flexible gripper unit 303 to maintain the current gripping state to prevent the target from being lost, and simultaneously controls the robot body to automatically return to the preset safe point through the motion control module 203. Fault data, deviation data, and operation logs are simultaneously uploaded to the remote monitoring terminal. The emergency protection program is in real-time monitoring throughout the closed-loop control process and will be executed immediately once any trigger condition is met. If there is no deviation, meaning the operation result fully meets the expected operation state, the harvesting operation is considered complete, and the control process ends.

[0066] In the aforementioned full-process control process, the principle of full-process closed-loop feedback control is as follows: Figure 5As shown, the control system layer 2 is based on a unified spatiotemporal reference for hardware synchronization, and synchronously coordinates the operation sequence of the motion control module 203, the harvesting control module 204 and the closed-loop feedback correction module 205. During the entire process of the robotic arm execution unit 302 performing the harvesting operation, the motion control module 203 dynamically adjusts the thrust output of the bionic thruster unit 301 based on the real-time pose deviation of the robot body, compensates for the body posture disturbance caused by the robotic arm movement in real time, maintains the dynamic hovering stability of the robot body, and realizes multi-task integrated collaborative control of body motion control and end-capture operation.

[0067] Technical principle of the invention:

[0068] The core technical principle of this invention is based on the closed-loop control theory of "perception-decision-execution-feedback-correction," constructing a three-layer integrated intelligent closed-loop control system. Through multimodal perception fusion under a unified spatiotemporal reference with hardware synchronization, adaptive disturbance rejection control integrating classical and intelligent technologies, full-process closed-loop feedback correction, and multi-task integrated collaborative control, it achieves intelligent autonomous control of the entire underwater harvesting operation process. The specific technical principles are as follows:

[0069] 1. Operating principle of the three-layer integrated closed-loop architecture

[0070] The perception and navigation layer, control system layer, and mechanical execution layer establish bidirectional full-duplex communication between each other via a high-speed industrial communication bus. Through hardware-synchronized PWM trigger signals output by the main controller, a unified clock reference is provided for all sensors, actuators, and control modules, ensuring that all sensor data, control commands, and execution feedback operate within the same time and space coordinate system. Time synchronization accuracy is better than 1ms, completely eliminating timing and spatial deviations. The perception and navigation layer, acting as the system's "sensors," is responsible for collecting comprehensive environmental and operational status data, providing decision-making support for the control system layer. The control system layer, acting as the system's "brain," is responsible for all data processing, decision-making, and command generation, forming the core of the entire closed-loop system. The mechanical execution layer, acting as the system's "limbs," is responsible for precisely executing control commands and completing actual operational actions. The closed-loop feedback correction module, acting as the system's "adjustment center," is responsible for comparing operational results and correcting control commands, forming a complete closed-loop circuit. This three-layer architecture works in tandem and interacts bidirectionally, achieving closed-loop control of the entire harvesting operation process.

[0071] 2. Hardware Implementation Principles of a Unified Spatiotemporal Reference

[0072] The unified spatiotemporal reference of this invention is achieved through the hardware synchronization circuit of the main controller. Specifically, the timer module of the industrial-grade high-performance main controller outputs a 1kHz synchronous trigger PWM signal. This signal is connected via shielded twisted-pair cables to the synchronous trigger input terminals of the forward-looking imaging sonar, binocular underwater vision camera, inertial measurement unit, depth pressure sensor, and gripping force feedback sensor, as well as the synchronization terminals of the control drive boards of the bionic thruster unit, robotic arm execution unit, and flexible gripper unit. This ensures that all sensors acquire data at the same time, and all actuators respond to control commands at the same time. Simultaneously, the main controller adds a unified timestamp to all sensor data. Through hand-eye calibration and extrinsic parameter calibration, the data from all sensors are unified into the robot's body coordinate system, achieving dual synchronization in time and space. This provides a foundation for multi-source data fusion and multi-module collaborative control.

[0073] 3. The principle of positioning and navigation based on tightly coupled fusion of multimodal data

[0074] This invention employs a tightly coupled data fusion strategy based on factor graph optimization, and the specific implementation steps are as follows:

[0075] Step 1: Construct the basic framework of the factor graph, using the robot's six-degree-of-freedom pose, velocity, and IMU zero bias as state nodes. The state nodes have 15 dimensions, including 3D position, 3D attitude quaternion, 3D velocity, 3D gyroscope zero bias, and 3D accelerometer zero bias.

[0076] Step 2: Construct various constraint factors, including:

[0077] (1) IMU pre-integration factor: Based on the high-frequency sampling data of the inertial measurement unit, the IMU data between two adjacent key frames is pre-integrated to obtain the relative pose and relative velocity constraints between adjacent key frames, and the IMU pre-integration factor is constructed and added to the factor graph;

[0078] (2) Visual odometry factor: Based on adjacent frame images acquired by binocular underwater visual camera, the relative pose transformation between adjacent frames is calculated through ORB feature extraction and matching. Combined with the camera's intrinsic and extrinsic parameters, the visual odometry factor is constructed and added to the factor graph.

[0079] (3) Sonar point cloud matching factor: Based on the point cloud data of adjacent frames acquired by forward-looking imaging sonar, point cloud registration is performed by ICP iterative nearest point algorithm, the relative pose transformation between adjacent frames is calculated, sonar point cloud matching factor is constructed and added to the factor graph.

[0080] (4) Depth constraint factor: Based on the depth data collected by the depth pressure sensor, the Z-axis position of the robot is constrained, the depth constraint factor is constructed, and added to the factor graph;

[0081] Step 3: Construct a nonlinear least squares optimization model with the goal of minimizing the sum of squared residuals of all constraint factors. The objective function is:

[0082]

[0083] in, The objective function to be minimized. : No. Each residual term (observed value) Compared with model predictions (deviation) : No. The weight matrix of the residuals (usually the inverse of the covariance matrix, used to balance the confidence levels of different observations). : The expansion of the square of the weighted norm;

[0084] Step 4: The Gauss-Newton iterative method is used to solve the objective function. The iteration terminates when the sum of squared residuals is less than 1e-6 or the number of iterations reaches 50. The optimal solution of the state node is updated iteratively to obtain the optimal estimate of the robot's six-degree-of-freedom pose. At the same time, based on the pose calculation results and environmental perception data, a three-dimensional grid map of the underwater working environment is constructed to realize simultaneous localization and mapping (SLAM).

[0085] This tightly coupled fusion strategy deeply integrates data from all sensors into the same optimization framework, rather than the loosely coupled combined navigation mode of existing technologies. It can make full use of the advantages of different sensors, eliminate the cumulative error of a single sensor, and even if a sensor fails for a short time, it can maintain positioning accuracy through the data of other sensors, which greatly improves the robustness and accuracy of positioning and navigation.

[0086] 4. Adaptive Disturbance Rejection Control Principle that Integrates Classic and Intelligent Approaches

[0087] The motion control and acquisition control of this invention both adopt a dual-loop fusion control structure of "adaptive PID + fuzzy neural network", and the specific implementation steps are as follows:

[0088] Step 1: Implementation of inner-loop adaptive PID control. The discrete expression for PID control is:

[0089] Where u(k) is the control output at time k, e(k) is the control deviation at time k, T is the control period (1ms), and K... p (k), K i Kd(k) and Kd(k) are the proportional, integral, and differential coefficients for adaptive adjustment at time k, respectively;

[0090] The adaptive adjustment rule is as follows: based on the magnitude and trend of the deviation, a piecewise adaptive strategy is adopted. When the absolute value of the deviation is greater than 50% of the set value, K is increased. p (1.2-1.5 times the baseline value), reduce K i (0.3-0.5 times the reference value) to improve system response speed; when the absolute value of the deviation is between 10% and 50% of the set value, adjust K. p K i K d To ensure system stability and response speed, move the system to near the reference value; when the absolute value of the deviation is less than 10% of the set value, increase K. i (1.2-1.5 times the baseline value), reduce K d (0.5-0.8 times the reference value) to eliminate system steady-state error;

[0091] Step 2: Implementation of the outer loop fuzzy neural network anti-disturbance compensation loop, using a 5-layer fuzzy neural network structure, including:

[0092] (1) Input layer: There are 5 input nodes, namely control deviation e, deviation change rate ec, water flow velocity v, target displacement Δp, and body attitude disturbance Δθ;

[0093] (2) Fuzzification layer: Each input variable is fuzzified and divided into 7 fuzzy subsets, namely {negative large, negative medium, negative small, zero, positive small, positive medium, positive large}. The Gaussian membership function is used to calculate the membership value corresponding to each input variable. The center value and width of the membership function are preset according to the range of the input variable.

[0094] (3) Rule layer: Implements fuzzy rule matching. Each node corresponds to one fuzzy control rule, and the number of fuzzy rules is 7. 5 =16807 rules, and the applicability of each rule is calculated through rule matching;

[0095] (4) Normalization layer: Normalizes the output of the rule layer to eliminate the magnitude difference in the applicability of different rules;

[0096] (5) Output layer: There are 3 output nodes, which are the correction values ​​ΔK of the PID parameters. p ΔK i ΔK d The weights of the output layer are updated in real time through an online adaptive learning algorithm, which uses gradient descent and sets the learning rate to 0.01. The learning objective is to minimize the sum of squared control deviations and adjust the network weights in real time.

[0097] Step 3: Fusion control output. The control output of the inner loop adaptive PID is superimposed with the disturbance feedforward compensation of the outer loop fuzzy neural network to obtain the final control output, which is then sent to the execution unit to achieve adaptive compensation for underwater dynamic disturbances.

[0098] This integrated control principle combines the stability of classical control with the adaptive capability of intelligent control. It can cope with various disturbances in complex dynamic environments and achieve high-precision and high-robustness control without the need to establish an accurate underwater robot dynamic model.

[0099] 5. Principle of closed-loop feedback and adaptive correction throughout the entire process

[0100] The closed-loop feedback correction module of this invention constructs a three-level correction mechanism of "deviation comparison - parameter tuning - command correction", and the specific implementation steps are as follows:

[0101] Step 1: Calculate the state deviation of the deviation comparison unit, specifically as follows:

[0102] (1) Position deviation calculation: The deviation between the actual position and the expected position of the robot body is calculated using Euclidean distance. The formula is:

[0103]

[0104] Among them, (x real ,y real ,z real (x) represents the actual position coordinates, (x) exp ,y exp ,z exp () represents the expected location coordinates;

[0105] (2) Attitude deviation calculation: Quaternion operations are used to calculate the deviation between the actual attitude and the expected attitude. The formula is as follows:

[0106] Among them, Q real Q is the actual attitude quaternion. exp The expected attitude quaternion is ⊗, which represents the quaternion multiplication operation. The attitude deviation quaternion is converted into Euler angles to obtain the attitude deviations in the roll, pitch, and yaw directions.

[0107] (3) Clamping force deviation calculation: The deviation between the actual clamping force and the expected clamping force is calculated using the difference. The formula is as follows:

[0108]

[0109] Among them, F real F represents the actual clamping force. exp The expected clamping force;

[0110] (4) Target displacement deviation calculation: The deviation between the actual position of the target and the expected grasping position is calculated using Euclidean distance;

[0111] Step 2: Parameter adjustment of the adaptive parameter tuning unit. Based on the state deviation calculated in Step 1 and the real-time environmental disturbance parameters (water flow velocity, turbulence intensity, target motion velocity) collected by the perception and navigation layer, fuzzy control rules are used to adjust the control tuning parameters of the motion control algorithm and the acquisition control algorithm in real time, including PID parameters, trajectory planning constraints, and thruster thrust distribution coefficients.

[0112] Step 3: Control command generation of the command correction unit. Based on the tuning parameters adjusted in Step 2, combined with the current robot's real-time pose, environmental state, and target state, the thrust drive parameters of the thruster, the motion trajectory of the robotic arm, and the gripper's clamping control parameters are recalculated to generate corrected motion control commands and capture control commands, which are then sent to the mechanical execution layer to achieve adaptive compensation for operational deviations.

[0113] This principle enables closed-loop correction of the entire chain from operation results to control commands, allowing the system to autonomously cope with various uncertainties in the operation process and achieve unmanned autonomous operation.

[0114] 6. Multi-task integrated collaborative control principle

[0115] This invention, based on a unified spatiotemporal reference with hardware synchronization, allocates synchronized runtime sequences to the motion control module, the acquisition control module, and the closed-loop feedback correction module. The control cycle for all three modules is 1ms, ensuring complete runtime synchronization. Throughout the robotic arm's acquisition operation, the acquisition control module synchronizes the robotic arm's action timing, trajectory, and expected disturbance to the motion control module in real time via an internal bus. The motion control module, based on the robotic arm's action state and combined with the robot's real-time pose feedback, anticipates the body's posture disturbance caused by the robotic arm's movements and dynamically adjusts the thrust distribution and output of the bionic thrusters, achieving a combination of feedforward compensation and feedback adjustment to maintain the robot's six-degree-of-freedom posture stability. This principle achieves deep collaboration between body motion control and end-effector acquisition operations, solving the problems of timing discrepancies and uncompensated disturbances between the two control links, and realizing integrated collaborative control of multiple tasks.

[0116] 7. Full-process emergency protection principle

[0117] The emergency protection procedure of this invention is in a real-time monitoring state throughout the entire closed-loop control process, with a monitoring period of 1ms. The triggering conditions include:

[0118] (1) The state deviation of the work result exceeds the preset correctable threshold range;

[0119] (2) Any sensor fails, and no valid data output exceeds 200ms;

[0120] (3) The thruster unit malfunctions, making attitude stabilization control impossible;

[0121] (4) The robotic arm's execution unit is stuck and cannot complete the preset action;

[0122] (5) Communication interruption between the robot and the remote monitoring terminal exceeds 5 seconds;

[0123] (6) The distance between the robot and the obstacle is less than the preset safe distance, posing a risk of collision;

[0124] (7) The robot's depth exceeds the preset safe depth range. As long as any of the above triggering conditions are met, the system will immediately stop the harvesting operation, control the flexible gripper to maintain the current gripping state, control the robot to automatically return to the preset safe point, and simultaneously upload fault data and operation logs to achieve full-process safety protection.

[0125] To make the present invention more fully disclosed, more specific embodiments are described below.

[0126] Example 1

[0127] This embodiment is applied in a shallow sea cucumber farming area with a water depth of 15-20 meters. The visibility in the working environment is 3-5 meters, with a low-speed reciprocating current ranging from 0.3-0.5 m / s. The farming area is a seabed-enclosed aquaculture area with relatively flat terrain and a few farming reefs and obstacles. The target for harvesting is live sea cucumbers from bottom-seeded farming. The underwater harvesting robot used in this embodiment has an open-frame structure, constructed entirely of 6061 aluminum alloy and high-density polyethylene. Its overall dimensions are 800mm × 600mm × 500mm, and its weight in air is 35kg. With a zero-buoyancy design underwater, it possesses excellent underwater stability and a pressure resistance rating of 30 meters, meeting the operational requirements of shallow sea farming areas. The perception and navigation layer comprises a multimodal sensor array. The forward-looking imaging sonar uses a single-beam forward-looking sonar with a detection range of 0.5-50 meters. The binocular underwater vision camera is a 1080P resolution sealed underwater camera equipped with a 20W underwater LED supplemental light. The inertial measurement unit uses an industrial-grade MEMSMU with a sampling frequency of 1000Hz and zero-bias stability of 1° / h. The depth pressure sensor has a measurement range of 0-30 meters and an accuracy of ±0.1%FS. The clamping force feedback sensor uses a 3-point array thin-film pressure sensor with a range of 0-50N. The control system layer's main controller uses an NXP i.MX 8QuadMax industrial-grade multi-core processor, running the Linux PREEMPT_RT real-time operating system. The high-speed industrial communication bus uses the EtherCAT bus to ensure real-time data transmission. The bionic thruster unit of the mechanical actuator layer uses four horizontal thrusters and two vertical thrusters, with a maximum thrust of 200N per unit, achieving six degrees of freedom motion control. The robotic arm actuator unit uses a six-degree-of-freedom underwater robotic arm with an operating radius of 0.8 meters and an end-effector repeatability of ±0.5mm. The flexible gripper unit uses a pneumatically driven silicone flexible gripper, suitable for grasping soft marine delicacies such as sea cucumbers.

[0128] The specific operation process of this embodiment is as follows: After the system is powered on, the operator sends the preset 10 sea cucumber harvesting points and the operation area to the robot control system layer through the shore-based remote monitoring terminal. The perception and navigation layer completes the calibration and initialization of all sensors. Through the 1kHz synchronous trigger signal of the main controller, it controls the multimodal sensor group to synchronously collect sonar point cloud data, visual image data, IMU attitude data, and depth data of the underwater environment. After completing spatiotemporal registration, the data is uploaded to the control system layer in real time. The fusion positioning and navigation module of the control system layer adopts a tightly coupled fusion algorithm based on factor graph optimization to perform fusion calculation on the multimodal data, construct a three-dimensional grid map of the aquaculture area, complete the robot's six-degree-of-freedom pose calculation, and based on the preset harvesting points, adopts an improved... The algorithm generates a globally collision-free operating path, avoiding obstacles such as aquaculture reefs. Based on the global operating path and combined with the robot's real-time pose feedback, the motion control module adopts a fusion control strategy of adaptive PID + fuzzy neural network to calculate the thrust drive parameters of the thrusters, generate motion control commands and send them to the mechanical execution layer, controlling the robot to navigate smoothly along the planned path. At the same time, it compensates for attitude disturbances caused by reciprocating currents in real time. After reaching the target harvesting point, the robot is controlled to achieve dynamic and precise hovering, with hovering position accuracy controlled within ±5cm.

[0129] After the robot hovers and stabilizes, the binocular vision camera in the perception and navigation layer identifies the sea cucumber target entering the working area. Through sonar and vision fusion localization, the robot obtains the sea cucumber's three-dimensional spatial coordinates and uploads them to the control system layer. Based on the target localization result, the harvesting control module uses a fifth-order polynomial interpolation algorithm to plan the motion trajectory and grasping sequence of the robotic arm's end effector, generating harvesting control commands which are then sent to the mechanical execution layer. This controls the robotic arm to move the flexible gripper above the sea cucumber target, and then controls the gripper to slowly close and grasp the sea cucumber. During the grasping process, the gripping force feedback sensor collects gripping force data in real time and transmits it back to the harvesting control module. The preset safe gripping threshold for sea cucumbers is 5N, and the safe overload threshold is 8N. When the gripping force reaches 5N, the gripper's gripping and feeding action is immediately stopped to maintain the current gripping force and avoid injuring the sea cucumber.

[0130] After the mechanical execution layer completes a single grasping action, the perception and navigation layer immediately collects feedback data on the current robot pose, robotic arm pose, gripper force, and sea cucumber grasping status, and transmits this data back to the control system layer. The closed-loop feedback correction module compares the feedback data with the preset expected operating state and calculates the state deviation. The preset correctable thresholds are: robot hovering position deviation ≤ 10cm, posture deviation ≤ 3°, gripping force deviation ≤ ±1N, indicating successful target grasping. If the deviation is within the correctable threshold range, the closed-loop feedback correction module adjusts subsequent control parameters, corrects control commands, and controls the robotic arm to retract to a safe position. The robot then travels to the next harvesting point and repeats the entire closed-loop control process. If the deviation exceeds the correctable threshold, such as when the sea cucumber falls off the grasp, an emergency protection program is triggered, controlling the robotic arm to return to its initial position. The robot then hovers and stabilizes again, re-identifies the target, and executes the harvesting process. If three consecutive harvesting attempts fail, the robot returns to a safe point on the water surface, and the operation log and deviation data are uploaded to the remote monitoring terminal.

[0131] Throughout the entire operation, the control system layer coordinates the operation sequence of motion control and acquisition control based on a unified spatiotemporal reference synchronized with hardware. During the entire process of the robotic arm's movement, the motion control module monitors the robot's posture changes in real time, dynamically adjusts the thrust output of the thrusters, compensates for the posture disturbances caused by the robotic arm's movements, and always maintains the robot's hovering stability, thus achieving integrated and coordinated control of the robot's movement and end-effector acquisition.

[0132] Example 2

[0133] The application scenario for this embodiment is a deep-sea rocky reef area with a water depth of 80-100 meters. The operating environment has a water visibility of less than 1 meter, irregular turbulence, and a current velocity range of 0.6-0.8 m / s. The seabed topography is complex, with numerous obstacles such as reefs and seaweed. The harvesting targets are wild abalone and sea urchins attached to the reefs. The underwater harvesting robot used in this embodiment has a pressure-resistant sealed structure, is made entirely of TC4 titanium alloy, has overall dimensions of 1200mm × 800mm × 700mm, weighs 85kg in air, has a zero-buoyancy design underwater, and a pressure resistance rating of 150 meters, enabling it to withstand the high-pressure environment of the deep sea. The robot body adopts a design where the center of gravity and center of buoyancy coincide, and the horizontal and vertical center of stability height is greater than 5cm, providing good attitude stability. The perception and navigation layer comprises a multimodal sensor array. The forward-looking imaging sonar employs a multi-beam 3D imaging sonar with a detection range of 0.5-80 meters, enabling high-precision 3D point cloud imaging in low-visibility environments. The binocular underwater vision camera is a low-light, high-sensitivity underwater camera with a minimum illumination of 0.001 Lux, equipped with a 30W adjustable underwater supplemental light. The inertial measurement unit (IMU) uses a high-precision fiber optic IMU with a sampling frequency of 2000Hz and zero-bias stability better than 0.5° / h. The depth pressure sensor has a measurement range of 0-150 meters and an accuracy of ±0.05%FS. The clamping force feedback sensor uses a 5-point array thin-film pressure sensor with a range of 0-100N. The control system layer's main controller uses an NXP i.MX 8QuadMax industrial-grade multi-core processor, running the Linux PREEMPT_RT real-time operating system, with an operating temperature range of -20℃ to 60℃. The high-speed industrial communication bus uses the CANopen bus, possessing strong anti-interference capabilities and adapting to the complex electromagnetic environment of the deep sea. The bionic thruster unit of the mechanical actuator layer uses 6 vector thrusters, with a maximum thrust of 300N per unit, to achieve high-precision motion control with six degrees of freedom and stronger resistance to current. The robotic arm actuator unit uses a 7-degree-of-freedom redundant underwater robotic arm with an operating radius of 1.2 meters and an end-effector repeatability of ±0.5mm, which provides stronger obstacle avoidance capabilities. The flexible gripper unit uses a rigid-flexible composite gripper driven by a servo motor, which can stably grasp abalone and sea urchins attached to rocks.

[0134] The specific operation process of this embodiment is as follows: After the system is powered on, the shipborne remote monitoring terminal sends the preset operation area and target type to the robot. The perception and navigation layer completes the calibration and initialization of the sensors under low temperature and high pressure environment. Through the 1kHz synchronous trigger signal of the main controller, it controls the forward-looking imaging sonar to collect three-dimensional point cloud data of the underwater environment, the binocular vision camera to collect near-field image data, the fiber optic IMU to collect high-frequency attitude data, and the depth sensor to collect depth data. After all data is spatiotemporally registered, it is uploaded to the control system layer. The fusion positioning and navigation module of the control system layer adopts a tightly coupled fusion strategy based on factor graph optimization. It mainly uses sonar point cloud data and IMU pre-integrated data, supplemented by visual data, to fuse and solve the robot's six-degree-of-freedom pose, construct a three-dimensional grid map of the deep-sea reef area, and generate a global collision-free search path based on the preset operation area to avoid obstacles such as reefs. The motion control module is based on the planned path and adopts a fusion control strategy of adaptive PID + fuzzy neural network. In response to the strong disturbance of deep-sea turbulence, it calculates the thrust parameters of the thruster in real time and controls the robot to sail smoothly along the planned path. At the same time, it compensates for the attitude disturbance caused by turbulence in real time and achieves dynamic and precise hovering after reaching the work area.

[0135] Due to extremely low visibility in the water, the perception and navigation layer uses multibeam imaging sonar to identify abalone and sea urchin targets attached to the reefs. Combined with the precise identification by a near-field vision camera, it achieves accurate target positioning and shape feature extraction, which is then uploaded to the control system layer. Based on the target positioning results and the reef terrain, the harvesting control module plans the obstacle avoidance trajectory of the robotic arm, generates harvesting control commands, and sends them to the mechanical execution layer. This controls the robotic arm to move the flexible gripper to the target position. The rigid structure of the gripper completes the peeling of the target, followed by stable grasping using the flexible structure. During the grasping process, the gripping force feedback sensor collects gripping force data in real time. The preset safe gripping threshold is 15N, and the safe overload threshold is 20N. When the gripping force reaches the safe threshold, the gripper feed stops, maintaining the gripping force to avoid damage to the target and the gripper.

[0136] After a single harvesting action is completed, the perception and navigation layer collects feedback data on the operation results. The control system layer compares this with the expected operation status, calculates the state deviation, and the closed-loop feedback correction module, based on the deviation and real-time turbulence disturbance parameters, corrects subsequent motion control and harvesting control commands in real time. This compensates for target displacement and body attitude disturbances caused by turbulence, and re-executes the entire closed-loop control process. If the deviation exceeds the correctable threshold, such as target stripping failure or the robot colliding with reefs, an emergency protection program is triggered. The robotic arm immediately stops its movements, controls the robot to retreat to a safe position, and replans the path and harvesting action. If a serious malfunction occurs, the robot automatically floats to a safe point on the water surface, while simultaneously uploading fault data and operation logs.

[0137] Throughout the entire operation, the control system layer synchronously coordinates the timing of motion control and acquisition control. During the entire process of the robotic arm's movement, the motion control module adjusts the thrust of the thrusters in real time to compensate for the body posture disturbance caused by the robotic arm's movement, ensuring the robot's hovering stability in complex turbulent environments and realizing multi-task integrated collaborative control.

[0138] Example 3

[0139] This embodiment is applied in nearshore circular aquaculture cages at a water depth of 5-10 meters. The operating environment is characterized by turbid water with a visibility of 1-2 meters. Live grouper are cultured inside the cages, with a water flow velocity of 0.2-0.4 m / s. The operating space is limited, with the cages measuring 10 m in diameter and 8 m in depth. Collisions with the cage structure must be avoided while ensuring the survival rate of the harvested fish. The underwater harvesting robot used in this embodiment features a miniaturized, compact, open-frame structure with overall dimensions of 450 mm × 450 mm × 280 mm and a weight of 12 kg in air. Its zero-buoyancy design allows for flexible movement within the cages, and it has a pressure resistance rating of 15 meters. The multimodal sensor array in the perception and navigation layer includes a forward-looking obstacle avoidance sonar, a binocular near-field vision camera, a MEMSMU, a depth sensor, and a clamping force feedback sensor. The forward-looking obstacle avoidance sonar uses a single-beam high-frequency sonar with a detection range of 0.1-20 meters, used to detect cage walls and obstacles to avoid collisions. The binocular near-field vision camera uses a miniaturized 720P resolution underwater camera for close-range identification of live fish and shrimp targets. The inertial measurement unit uses a miniaturized MEMSMU with a sampling frequency of 500Hz and zero-bias stability of 2° / h. The depth sensor has a measurement range of 0-15 meters and an accuracy of ±0.1%FS. The clamping force feedback sensor uses a flexible thin-film pressure sensor with a range of 0-20N and high sensitivity. The main controller in the control system layer uses a miniaturized NXP i.MX 8QuadMax industrial-grade multi-core processor, running the Linux PREMIPT_RT real-time operating system, featuring low power consumption and high real-time performance. The high-speed industrial communication bus uses the EtherCAT bus to ensure rapid response to control commands. The bionic thruster unit of the mechanical actuator layer uses four small underwater thrusters, each with a maximum thrust of 100N, achieving four degrees of freedom motion control and possessing flexible steering and hovering capabilities. The robotic arm actuator unit uses a four-degree-of-freedom small underwater robotic arm with an operating radius of 0.5 meters and an end-effector repeatability of ±1mm. The flexible gripper unit uses a soft pneumatic gripper made of food-grade silicone, which can achieve flexible gripping of live fish and shrimp, avoiding damage to the fish and shrimp and ensuring the survival rate of the live animals.

[0140] The specific operation process of this embodiment is as follows: After the system is powered on, the remote monitoring terminal of the net cage platform sends the preset net cage harvesting area and target fish and shrimp specifications to the robot. The perception and navigation layer completes the calibration and initialization of the sensors, and simultaneously collects sonar obstacle avoidance data, visual image data, IMU attitude data, and depth data inside the net cage. After completing spatiotemporal registration, the data is uploaded to the control system layer. The fusion positioning and navigation module of the control system layer performs fusion calculation on the multimodal data, constructs a three-dimensional environment map inside the net cage, completes the robot's pose calculation, and generates a globally collision-free path based on the preset harvesting area to avoid the net cage frame and netting structure. The motion control module, based on the planned path, adopts a fusion control strategy to calculate the thrust parameters of the thrusters, controls the robot to navigate smoothly inside the net cage, and achieves precise hovering after reaching the harvesting area. At the same time, it compensates for water flow disturbances in real time to avoid colliding with the net cage.

[0141] The perception and navigation layer uses a fusion of vision and sonar to identify live grouper targets that meet specifications within the net cage. It employs a Kalman filter algorithm to track the target's movement trajectory in real time, uploading the target's real-time position and speed to the control system layer. Based on the target's real-time movement trajectory, the harvesting control module uses a model predictive control algorithm to plan the robotic arm's dynamic following trajectory and grasping timing. It generates harvesting control commands and sends them to the mechanical execution layer, controlling the robotic arm to dynamically follow the fish / shrimp target with its soft grippers, completing the grasping action at the optimal moment. During the grasping process, a gripping force feedback sensor collects gripping force data in real time. The preset safe gripping threshold is 3N, and the safe overload threshold is 5N, strictly controlling the gripping force within a safe range to avoid damage to the fish / shrimp and ensure a high survival rate.

[0142] After a single grasping action is completed, the perception and navigation layer collects feedback data on the operation results. The control system layer compares this data with the expected operation status and calculates the deviation, such as target escape, positional deviation, and gripping force deviation. Based on the deviation, the closed-loop feedback correction module corrects the robot's pose control commands and the robotic arm's trajectory planning commands in real time, and re-executes the entire closed-loop control process. If the deviation exceeds a correctable threshold, such as when the distance between the robot and the cage wall is less than 20cm, posing a collision risk, an emergency protection program is triggered. The robot's movement is immediately stopped, and it is controlled to retreat to a safe position and replan its path. If five consecutive grasping failures occur, the robot is controlled to return to a safe point at the cage entrance and upload the operation data.

[0143] Throughout the entire operation, the control system layer synchronously coordinates motion control and capture control. As the robotic arm dynamically tracks the target, it adjusts the robot's posture in real time to ensure coordinated movement between the robot body and the end effector, achieving precise capture of dynamic targets. At the same time, it monitors the distance between the robot and the net cage throughout the process to avoid collision risks.

[0144] Example 4

[0145] This embodiment is applied in inland lakes with a water depth of 10-30 meters for emergency salvage operations of underwater wrecked drones and underwater rescue equipment. The operating environment is characterized by turbid water, visibility of less than 1 meter, irregular currents, and uncertain target locations, requiring extensive searching and precise salvage. High reliability and rapid response capabilities are also essential. The underwater harvesting robot used in this embodiment has a modular structure, enabling rapid assembly and disassembly to adapt to different salvage operation needs. Its overall dimensions are 1000mm × 700mm × 600mm, its air weight is 60kg, it features zero buoyancy underwater, a pressure resistance rating of 50 meters, and is equipped with a 433MHz wireless communication module and an underwater acoustic communication module, enabling real-time communication with shore-based monitoring terminals. The multimodal sensor group of the perception and navigation layer includes a forward-looking imaging sonar, a side-scan sonar, a binocular vision camera, a high-precision IMU, a depth sensor, and a clamping force feedback sensor. The side-scan sonar uses a dual-channel side-scan sonar with a scan width of 0-100 meters for scanning a wide underwater environment and achieving full-coverage search. The forward-looking imaging sonar uses a multi-beam imaging sonar for precise imaging of the target area. The binocular vision camera uses a low-light underwater camera for near-field target detail identification. The inertial measurement unit uses an industrial-grade MEMSMU with a sampling frequency of 1000Hz and zero-bias stability of 0.5° / h. The depth sensor has a measurement range of 0-50 meters and an accuracy of ±0.05%FS. The clamping force feedback sensor uses a large-range array pressure sensor with a range of 0-200N, which can adapt to the grasping needs of targets of different weights. The main controller of the control system layer adopts the NXP i.MX 8QuadMax industrial-grade high-performance main controller, which has powerful computing capabilities and can handle massive point cloud data for large-scale searches. The high-speed industrial communication bus adopts the CANopen bus to ensure communication reliability. The bionic thruster unit of the mechanical execution layer adopts 8 vector thrusters, with a maximum thrust of 300N per unit, which has strong power and high resistance to current, and can achieve high-precision motion control of six degrees of freedom. The robotic arm execution unit adopts a 6-degree-of-freedom heavy-duty underwater robotic arm with an operating radius of 1.5 meters and a maximum load of 20kg. The flexible gripper unit adopts an adaptive rigid-flexible composite gripper, which can adapt to salvage targets of different shapes and weights and achieve stable grasping.

[0146] The specific operation process of this embodiment is as follows: After the system is powered on, the shore-based emergency rescue command center sends the preset search area and salvage target features to the robot. The perception and navigation layer completes the calibration and initialization of the sensors, performs a large-scale underwater environment scan using side-scan sonar, collects point cloud data of the target area using forward-looking imaging sonar, collects near-field images using a visual camera, and collects pose data using an IMU and depth sensor. After all data is spatiotemporally registered, it is uploaded to the control system layer. The fusion positioning and navigation module of the control system layer performs fusion calculation on the multimodal data, constructs a large-scale underwater 3D map, completes the global pose calculation of the robot, generates a full-coverage bow-shaped search path based on the preset search area, and controls the robot to complete an underwater net-like search.

[0147] Once the side-scan sonar of the perception and navigation layer identifies a suspected salvage target, the control system layer guides the robot to navigate to the target area. Using forward-looking imaging sonar and a vision camera, it accurately identifies and locates the target. After confirming the wrecked drone target, it plans the robot's navigation path and controls the robot to reach the target location, achieving dynamic and precise hovering. Based on the target's location, shape, and weight, the capture control module plans the robotic arm's salvage trajectory and grasping point, generating capture control commands which are sent to the mechanical execution layer. This controls the robotic arm to move the composite gripper to the target position and controls the gripper to close, completing a stable grasp of the target. During the grasping process, the gripping force feedback sensor collects gripping force data in real time. Based on the target's weight, a corresponding safe gripping threshold of 30N and an overload threshold of 50N are preset to ensure grasping stability, prevent target loss, and avoid gripper overload damage.

[0148] After a single salvage operation is completed, the perception and navigation layer collects feedback data on the operation results. The control system layer compares this data with the expected operational status and calculates the deviation, such as insecure target gripping, positional deviation, or attitude deviation. Based on the deviation, the closed-loop feedback correction module adjusts the gripping force of the gripper, the pose of the robotic arm, and the robot's hovering control commands in real time to ensure the stability of the gripping. If the deviation exceeds the correctable threshold, such as when the target falls off, an emergency protection program is triggered, controlling the robot to search for the target again and execute the salvage process once more. If serious problems such as robot malfunction or attitude instability occur, the robot is immediately and automatically raised to a safe location on the water surface, while simultaneously uploading fault data and operation logs to the shore-based command center in real time.

[0149] Throughout the entire operation, the control system coordinates the entire process of search, navigation, hovering, and salvage in real time, achieving integrated collaborative control of multiple tasks. At the same time, the closed-loop feedback correction module compensates for deviations caused by water flow disturbances and target displacement in real time, ensuring the speed and success rate of emergency salvage operations.

[0150] Comparative Example 1

[0151] This comparative underwater harvesting robot system includes the robot body, a sensing unit, a control unit, and an execution unit. The sensing unit uses only a binocular underwater vision camera as its sole environmental sensing sensor. The control unit uses a traditional 8-bit microcontroller. The execution unit includes an underwater thruster, a robotic arm, and a rigid gripper. The sensing unit, control unit, and execution unit communicate via ordinary serial communication, resulting in low communication speed, high latency, and the lack of a unified spatiotemporal reference for hardware synchronization. The control architecture is purely open-loop control, lacking a closed-loop feedback correction module and a work result feedback mechanism. During operation, the control unit generates fixed control commands based on the operator's preset work path and harvesting actions, and issues them to the execution unit to control the robot to complete navigation and harvesting operations. The sensing unit only collects environmental data once before the start of the operation for initial path planning; it does not collect real-time environmental and work status feedback data during the operation, nor does it compare work results or correct control commands. During the robotic arm's harvesting operation, the control unit does not monitor changes in the robot body's attitude, nor does it dynamically adjust the thruster's thrust, failing to compensate for attitude disturbances caused by the robotic arm's movements or position drift caused by water flow disturbances. The system in this comparison model does not establish a closed-loop control for the entire process, and cannot cope with dynamic factors such as water flow disturbance, target displacement, and body posture disturbance during the operation. During the operation, operators need to intervene remotely in real time to continuously adjust the robot's posture and capture actions. In complex underwater environments, the operation accuracy is extremely low, the capture success rate is poor, and problems such as robot collision with obstacles, damage to the capture target, capture failure, and even the risk of robot loss are very likely to occur.

[0152] Comparative Example 2

[0153] This comparative underwater harvesting robot system includes a robot body, a sensing unit, a control unit, and an execution unit. The sensing unit includes an inertial measurement unit, a depth sensor, and a binocular underwater vision camera. The control unit includes an independent motion control module. The execution unit includes a bionic thruster, a robotic arm, and a rigid gripper. The system only constructs a single-stage closed-loop control for the robot body's motion control. That is, the motion control module adjusts the thruster control commands based on real-time attitude data from the inertial measurement unit and the depth sensor to achieve robot attitude stability. However, it does not construct a closed-loop control for the harvesting operation, nor does it set up a full-process operation result feedback and command correction mechanism. During operation, the robot's navigation and hovering control adopts a closed-loop mode, which can achieve basic attitude stability. However, the harvesting control adopts a pure open-loop mode. The control unit generates fixed harvesting commands based on the preset robotic arm movement trajectory and sends them to the execution unit. It does not collect harvesting operation result feedback data, nor does it set up a gripping force feedback sensor. It cannot achieve closed-loop control of the grasping process, and cannot adjust the harvesting commands according to the actual position and grasping state of the target. Meanwhile, the motion control module and the capture control module use independent controllers, lacking a unified spatiotemporal reference. This results in asynchronous operation sequences, meaning the motion control module cannot perceive the robotic arm's motion status in real time during its movements, and cannot predict and compensate for the body's posture disturbances caused by the arm's actions. This leads to a significant decrease in the robot's hovering accuracy during capture, and even posture deviations. In this comparative system, closed-loop control only covers the single aspect of motion control, failing to cover the capture operation and the entire process's result feedback. It cannot achieve adaptive correction of the capture operation, resulting in a low success rate of target capture in dynamic underwater environments, and is highly susceptible to target damage and capture failure, making it unsuitable for complex operational environments.

[0154] Comparative Example 3

[0155] This comparative underwater harvesting robot system includes a robot body, a sensing unit, a control unit, and an execution unit. The sensing unit includes a forward-looking imaging sonar, a binocular vision camera, an inertial measurement unit, and a depth sensor. The control unit includes a motion control module and a harvesting control module, both of which employ traditional fixed-parameter PID control strategies. The execution unit includes a bionic thruster, a robotic arm, and a flexible gripper. The system employs semi-closed-loop feedback control, but it lacks intelligent disturbance rejection algorithms and a full-process closed-loop feedback correction module. During operation, both the motion control module and the harvesting control module use pre-set fixed PID parameters, adjusting the control output only based on the current deviation value. They cannot adaptively fit the nonlinear disturbance characteristics of the underwater environment and cannot adaptively compensate for dynamic factors such as changes in water flow velocity, target displacement, and robot posture disturbances. When changes occur in the underwater environment, such as increased water flow velocity or target movement, the fixed-parameter PID control cannot adjust the control parameters in real time, leading to a slower system response, a significant decrease in control accuracy, and even control oscillations. This makes it impossible to maintain stable hovering and precise harvesting operations. Furthermore, the system lacks a feedback comparison mechanism for operation results and an adaptive correction mechanism for control commands. When a deviation occurs in a single operation, it cannot automatically correct subsequent control commands, requiring operators to manually readjust PID parameters and operation plans, resulting in extremely low levels of intelligence. The comparative system employs traditional fixed-parameter PID control, which exhibits poor adaptability to nonlinear, strongly coupled, and highly disturbed underwater dynamic systems, exhibiting weak anti-disturbance capabilities. Its operational performance deteriorates significantly in complex dynamic underwater environments, failing to guarantee operational accuracy and success rates.

[0156] Comparative Example 4

[0157] This comparative underwater harvesting robot system includes the robot body, a sensing unit, a control unit, and an execution unit. The sensing unit uses only a binocular underwater vision camera as its sole sensor for environmental perception, positioning, navigation, and target recognition, without any other sensors such as imaging sonar or a high-precision inertial measurement unit. The control unit includes a navigation module, a motion control module, and a harvesting control module. The execution unit includes an underwater thruster, a robotic arm, and a rigid gripper. The system's positioning, navigation, obstacle avoidance, and target recognition rely entirely on the vision camera. During operation, when underwater visibility is high and light is sufficient, the vision camera can acquire clear environmental image data, enabling basic positioning, navigation, and target recognition. However, when the water is turbid, visibility is low, or underwater light is insufficient, the image quality of the vision camera drops significantly, or even fails to capture images at all, causing the system's positioning, navigation, obstacle avoidance, and target recognition functions to completely fail, and the robot cannot complete the task. Furthermore, using only a single vision sensor prevents the fusion of multi-source data, and positioning and navigation rely entirely on visual odometry. Over long periods of operation, severe cumulative errors occur, leading to a continuous decline in the robot's pose calculation accuracy, deviations in path planning, and even the risk of collisions with obstacles. Furthermore, the system in this comparative example lacks a clamping force feedback sensor at the end of the gripper, making it impossible to achieve precise sensing and closed-loop control during the grasping process. The inability to monitor the clamping force in real time during grasping makes it highly susceptible to problems such as excessive clamping force damaging the target or insufficient clamping force causing the target to detach. The system in this comparative example also suffers from a single sensing method and extremely poor environmental adaptability, operating only in ideal environments with clear water and ample light. It cannot adapt to the complex underwater environments encountered in actual engineering projects, resulting in extremely low operational reliability and stability.

[0158] Comparative Example 5

[0159] This comparative underwater harvesting robot system includes the robot body, a sensing unit, a control unit, and an execution unit. The sensing unit includes a forward-looking imaging sonar, a binocular vision camera, an inertial measurement unit, and a depth sensor. The control unit includes completely independent motion control and harvesting control modules, with no communication connection between the two modules and no unified spatiotemporal reference. The execution unit includes a bionic thruster, a robotic arm, and a gripper. The system lacks a closed-loop feedback correction module and a full-process emergency protection procedure. During operation, the motion control module and the harvesting control module operate completely independently, using their own controllers and clock sources, with completely asynchronous operation sequences. The motion control module is only responsible for the robot's navigation and hovering control, while the harvesting control module is only responsible for the motion control of the robotic arm and gripper. There is no data interaction or collaborative control mechanism between the two modules. During the robotic arm's harvesting actions, the motion control module cannot sense the robotic arm's motion state, trajectory, or expected disturbance, cannot adjust the thruster's output in real time, and cannot compensate for the robot's posture disturbances caused by the robotic arm's movements. This leads to severe posture deviations and positional drifts during harvesting, and in severe cases, even the risk of the robot colliding with obstacles or capsizing. Meanwhile, the system lacks a full-process operation status monitoring and deviation comparison mechanism, making it unable to monitor the deviation between the operation results and the expected state in real time. When serious deviations occur, the system malfunctions, or the robot faces collision risks, the emergency protection program cannot be triggered in a timely manner, easily leading to serious accidents such as robotic arm jamming, gripper overload damage, target loss, and robot body collision damage. The system in this comparative example lacks a collaborative control mechanism and an emergency protection mechanism, resulting in extremely poor operational safety and reliability. It cannot adapt to long-term continuous operation in complex underwater environments, has a short equipment lifespan, and a high failure rate.

[0160] Performance index testing:

[0161] To verify the technical effects of the present invention, performance index testing was conducted on the systems of Examples 1-4 and Comparative Examples 1-5. By comparing the test data, the significant technical progress of the present invention was verified.

[0162] 1. The specific testing methods for each indicator are as follows:

[0163] 1. Positioning accuracy test: The underwater ultra-short baseline (USBL) acoustic positioning system was used as the position reference. The reference positioning accuracy was ±2cm. The deviation between the real-time positioning result of the robot and the reference position under different flow velocity conditions was tested. 100 test data were recorded for each condition and the average value was calculated.

[0164] 2. Hovering position accuracy test: Control the robot to hover at the target point for 10 minutes, record the robot's real-time position data at a frequency of 10Hz, and calculate the maximum deviation between the real-time position and the target point. Perform 10 tests for each working condition and calculate the average value.

[0165] 3. Target grasping success rate test: According to the underwater operation performance test method, 100 target grasping operations were carried out under each working condition. The success standard was that the target was stably grasped and retrieved to the robot's safe position. The number of successful graspings was counted and the grasping success rate was calculated.

[0166] 4. Damage rate test of the captured target: After 100 capture operations, count the number of times the captured target suffered physical damage and calculate the damage rate;

[0167] 5. Emergency protection response time test: A high-speed digital oscilloscope was used to collect the time difference between the emergency protection trigger signal and the emergency protection action execution signal. The trigger signal was the level signal with the deviation exceeding the threshold, and the execution signal was the output signal of the thruster return command. The average value of 10 tests was recorded.

[0168] 6. Continuous operation without failure time test: Conduct continuous operation test in a simulated operation environment, record the time during which the robot operates without failure during continuous operation, and the test deadline is 120 hours.

[0169] 2. Test Environment

[0170] The testing environment was the underwater comprehensive testing pool of the National Underwater Robot Quality Inspection and Testing Center. The pool measures 50m×30m×20m and is equipped with a water flow simulation system that can simulate stable and turbulent water flow environments of 0-1m / s. The test water depth was uniformly set at 10 meters and the ambient temperature at 20℃. Tests were conducted under three conditions: still water (0m / s flow), low-speed water flow (0.5m / s), and high-speed water flow (0.8m / s). Under each condition, 100 repeated capture operations were performed on each test object, and the average values ​​of various performance indicators were recorded.

[0171] 3. Test Object

[0172] The test subjects were the systems of Example 1, Example 2, Example 3, and Example 4, as well as the systems of Comparative Example 1, Comparative Example 2, Comparative Example 3, Comparative Example 4, and Comparative Example 5.

[0173] 4. Test Results

[0174] The results of this test are shown in the table below:

[0175]

[0176]

[0177] 5. Test Data Analysis and Conclusions

[0178] As can be seen from the test data in the table above, all performance indicators of Examples 1-4 of the present invention are far superior to the requirements of the national standard GB / T34526-2017, and significantly superior to Comparative Examples 1-5. Specific analysis is as follows:

[0179] First, regarding positioning accuracy and hovering stability, the embodiments of this invention demonstrate positioning and hovering accuracy far exceeding national standards under three conditions: still water, 0.5 m / s, and 0.8 m / s. Furthermore, they represent an order-of-magnitude improvement compared to comparative examples. Under the 0.8 m / s high-speed water flow condition, the systems of comparative examples 1 and 4 completely fail, unable to achieve positioning and hovering. The accuracy of comparative examples 2, 3, and 5 also drops significantly, approaching or exceeding national standard limits. In contrast, the positioning accuracy of the embodiments of this invention remains around ±5 cm, and the hovering accuracy remains within ±5.5 cm, far exceeding national standard requirements. This is because the invention employs a multi-modal sensor tightly coupled fusion navigation and positioning technology. Through multi-source data fusion under a unified spatiotemporal reference with hardware synchronization, the cumulative error of a single sensor is eliminated. Simultaneously, the adaptive anti-disturbance control strategy, which integrates classic and intelligent methods, can compensate for attitude deviations caused by water flow disturbances in real time. In contrast, the systems of the comparative examples either use a single sensor, open-loop control, or fixed-parameter PID control, which cannot cope with complex water flow environments, resulting in a significant decrease in positioning and hovering accuracy.

[0180] Secondly, regarding the target capture success rate, the capture success rate of the embodiments of this invention is far higher than the national standard requirements under different flow rate conditions. It approaches 100% in still water conditions and maintains a success rate of over 91% even in high-speed water flow conditions of 0.8 m / s. In contrast, the capture success rate of the comparative systems drops significantly as the flow rate increases. Comparative Examples 1 and 4 have a capture success rate of 0% in high-speed water flow conditions, making operation completely impossible. This is because the present invention constructs a full-process intelligent closed-loop control architecture, realizing integrated and coordinated control of motion control and capture control. The closed-loop feedback correction module can compensate for deviations caused by water flow disturbances, target displacement, and body posture disturbances in real time. After a deviation occurs in a single operation, it can automatically correct the control command and re-execute the operation process. The comparative systems, however, lack full-process closed-loop control, have poor coordination, and cannot adaptively respond to dynamic environmental changes, resulting in an extremely low capture success rate.

[0181] Third, regarding the damage rate of the captured target, the target damage rate in this embodiment of the invention is only 0.5%-0.8%, far below the national standard requirement of ≤5%, while the damage rates of comparative examples 1, 4, and 5 far exceed the national standard limit, and comparative examples 2 and 3 are also close to or exceed the national standard limit. This is because the capture control module of this invention integrates a clamping force feedback closed-loop control mechanism, realizing refined control of the grasping process. It can accurately control the clamping force within a safe range, achieving flexible and non-destructive grasping. In contrast, most of the comparative systems lack a clamping force feedback closed loop, making it impossible to accurately control the clamping force and easily causing damage to the captured target.

[0182] Fourth, regarding emergency response time and continuous operation without failure time, the emergency response time of this embodiment is only 120-130ms, far below the national standard requirement of ≤500ms. In contrast, the systems in the comparative examples lack a robust emergency protection mechanism, failing to guarantee operational safety. The continuous operation without failure time of this embodiment exceeds 115 hours, more than four times the national standard requirement, significantly superior to the systems in the comparative examples. This is because this invention constructs a full-process status monitoring and multi-scenario emergency protection mechanism, enabling rapid response to operational risks. Furthermore, the three-layer integrated industrial-grade architecture, full-process closed-loop control, and collaborative protection mechanism greatly improve the system's reliability and stability. In contrast, the systems in the comparative examples have a fragmented architecture, lack status monitoring and protection mechanisms, resulting in a high failure rate and poor continuous operation capability.

[0183] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. 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 intelligent closed-loop control system of an underwater harvesting robot, comprising an underwater harvesting robot body, and a perception unit, a control unit and an execution unit carried on the robot body, characterized in that, The sensing unit, control unit, and execution unit are respectively constructed as a sensing and navigation layer, a control system layer, and a mechanical execution layer. The three layers are connected in pairs through a high-speed industrial communication bus for bidirectional full-duplex communication, and together they form an intelligent closed-loop control system for the entire harvesting operation process. The perception and navigation layer is used to collect multimodal sensing data of the underwater working environment and working status, upload the multimodal sensing data to the control system layer in real time, and after the mechanical execution layer completes a single working action, collect the working result feedback data and send it back to the control system layer. The control system layer is mounted in an industrial-grade high-performance main controller inside the pressure-resistant sealed cabin of the robot body. It has built-in a fusion positioning and navigation module, a motion control module, a capture control module, and a closed-loop feedback correction module. Each module interacts bidirectionally with the main controller and is connected to each other through an internal high-speed bus. The control system layer is used to receive multimodal sensing data uploaded by the perception and navigation layer, complete the real-time pose calculation and global operation path planning of the robot through the fusion positioning and navigation module, generate motion control commands for the robot body based on the global operation path and real-time posture feedback through the motion control module, generate capture control commands for the robotic arm and flexible gripper based on the target recognition and positioning results through the capture control module, and send all control commands to the mechanical execution layer through the high-speed industrial communication bus. The control system layer is based on a unified spatiotemporal reference triggered by hardware synchronization. It synchronously coordinates the operation sequence of the motion control module, the acquisition control module and the closed-loop feedback correction module to achieve integrated collaborative control of multiple tasks. The mechanical execution layer includes a bionic thruster unit that is communicatively connected to the motion control module, a robotic arm execution unit and a flexible gripper unit that are communicatively connected to the capture control module, and is used to receive control commands issued by the control system layer to complete the robot's navigation, hovering and target capture actions. The closed-loop feedback correction module is used to compare the operation result feedback data returned by the perception and navigation layer with the preset expected operation state in real time, calculate the state deviation, and correct the motion control command and the capture control command in real time based on the state deviation and the real-time environmental disturbance parameters. It completes the adaptive compensation for three types of dynamic factors: water flow interference, target displacement, and body attitude disturbance, and realizes the closed-loop feedback and adaptive adjustment of the whole process from global path planning and body motion control to end capture operation.

2. The intelligent closed loop control system of the underwater harvesting robot according to claim 1, wherein, The multimodal sensor group of the perception and navigation layer includes a forward-looking imaging sonar, a binocular underwater vision camera, an inertial measurement unit, a depth and pressure sensor, and a clamping force feedback sensor fixedly installed on the flexible gripper's gripping end. The clamping force feedback sensor is used to collect clamping force data in real time during the target grasping process and transmit the clamping force data back to the acquisition control module and closed-loop feedback correction module of the control system layer in real time.

3. The intelligent closed loop control system of the underwater harvesting robot according to claim 1, wherein, Both the motion control module and the acquisition control module of the control system layer adopt a control strategy that integrates classical control and intelligent control. The classical control uses adaptive PID closed-loop control to achieve stable execution of basic actions, while the intelligent control uses a fuzzy neural network anti-disturbance algorithm to achieve adaptive compensation for three types of complex dynamic environments: underwater current disturbances, visibility changes, and target movement.

4. The intelligent closed loop control system of the underwater harvesting robot according to claim 1, wherein, The closed-loop feedback correction module incorporates a deviation comparison unit, an adaptive parameter tuning unit, and a command correction unit that are sequentially bidirectionally connected. The deviation comparison unit is used to calculate the real-time state deviation between the operation result feedback data and the preset expected operation state. The adaptive parameter tuning unit is used to adjust the control tuning parameters of the motion control algorithm and the acquisition control algorithm based on the state deviation and the real-time environmental disturbance parameters collected by the perception and navigation layer. The command correction unit is used to generate corrected motion control commands and acquisition control commands based on the adjusted tuning parameters.

5. The intelligent closed loop control system of the underwater harvesting robot according to claim 1, wherein, Throughout the entire process of the robotic arm execution unit performing the harvesting operation, the control system layer dynamically adjusts the thrust output of the bionic thruster unit based on the real-time pose deviation of the robot body through the motion control module. This compensates for the body posture disturbance caused by the robotic arm's movements in real time, maintains the dynamic hovering stability of the robot body, and realizes multi-task integrated collaborative control of body motion control and end-effector harvesting operations.

6. An intelligent closed-loop control method for an underwater harvesting robot, characterized in that, The intelligent closed-loop control system for the underwater harvesting robot according to any one of claims 1-5 is implemented by including the following steps: S1. Perception Initialization and Data Acquisition: After the system is powered on, the perception and navigation layer completes the calibration and initialization of the multimodal sensor group. Through the synchronous trigger signal of the main controller, the multimodal sensor group is controlled to synchronously acquire real-time sensing data of the underwater working environment, including target point cloud data of forward-looking imaging sonar, environmental image data of binocular underwater vision camera, attitude data of inertial measurement unit, and pressure data of depth sensor. After performing spatiotemporal registration on all sensing data, it is uploaded to the control system layer in real time. S2. Multi-source data fusion and path planning: The control system layer uses fusion positioning and navigation algorithms to perform tight-coupled fusion calculation on the spatiotemporally registered multimodal sensor data, construct a three-dimensional grid map of the underwater operation environment, complete the robot's six-degree-of-freedom real-time pose calculation, and generate a global collision-free operation path and target navigation plan based on the preset operation target. S3. Adaptive motion control and attitude stabilization: The motion control algorithm of the control system layer is based on the generated global operation path and combined with real-time attitude feedback data. It calculates the multi-channel thrust drive parameters of the bionic thruster unit, generates motion control commands and sends them to the mechanical execution layer to control the robot to complete stable navigation and dynamic precise hovering. S4. Target recognition and capture operation control: When the perception and navigation layer recognizes that the target has entered the preset operation range, the capture control algorithm of the control system layer plans the motion trajectory and grasping sequence of the end of the robotic arm based on visual recognition and real-time positioning results. Combined with the clamping force feedback mechanism, it generates capture control commands and sends them to the mechanical execution layer to control the robotic arm execution unit and flexible gripper unit to complete the target capture operation. S5. Feedback on work results and comparison of status: After the mechanical execution layer completes a single work action, the perception and navigation layer immediately collects feedback data on the current work environment, robot pose and work results, and transmits the feedback data back to the control system layer in real time. The control system layer compares the received feedback data with the preset expected work status in real time and calculates the status deviation. S6. Closed-loop correction and operation process determination: Based on the calculated state deviation, the control system layer determines whether the deviation is within the preset correctable threshold range. If the deviation is within the correctable threshold range, the subsequent motion control commands and harvesting control commands are corrected in real time through the closed-loop feedback correction algorithm to compensate for the uncertainties caused by water flow interference, target displacement, and body attitude disturbance. The system then returns to step S1 to re-execute the entire closed-loop control process based on the corrected control target. If the deviation exceeds the correctable threshold range, the emergency protection procedure is triggered. If there is no deviation, the harvesting operation is determined to be completed, and the control process ends.

7. The intelligent closed-loop control method of the underwater harvesting robot according to claim 6, characterized in that, In step S2, the fusion positioning and navigation algorithm adopts a tightly coupled data fusion strategy based on factor graph optimization. It fuses and solves sonar target point cloud data, visual environment image data, inertial measurement unit attitude data and depth sensor pressure data to construct a three-dimensional grid map of the underwater operation environment. Simultaneously, it completes the robot's six-degree-of-freedom real-time pose calculation and global collision-free operation path planning. 8.The intelligent closed-loop control method of the underwater harvesting robot according to claim 6, wherein, In step S4, the clamping force feedback mechanism integrated into the capture control algorithm is as follows: during the process of the flexible gripper clamping the target, the clamping force data is collected in real time through the clamping force feedback sensor. When the clamping force reaches the preset target safety clamping threshold, the clamping feed action of the gripper is stopped immediately to maintain the current clamping force. When the clamping force exceeds the preset safety overload threshold, the gripper is driven to release to the safety clamping range in the reverse direction, so as to realize flexible and non-destructive grasping and overload protection of the target. 9.The intelligent closed-loop control method of the underwater harvesting robot according to claim 6, wherein, In step S5, the preset expected working state includes the expected pose of the robot body with six degrees of freedom, the expected working pose of the end effector of the robotic arm, the expected grasping state of the target and the expected clamping force range; the state deviation includes one or more combinations of position deviation, posture deviation, clamping force deviation and target displacement deviation. 10.The intelligent closed-loop control method of the underwater harvesting robot according to claim 6, wherein, The emergency protection procedure triggered in step S6 specifically includes: the control system layer immediately stopping the harvesting operation, controlling the flexible gripper to maintain the current gripping state, and simultaneously controlling the robot body to return to the preset safe point through the motion control module, and uploading the fault data, deviation data and operation log to the remote monitoring terminal in a synchronized manner; the emergency protection procedure is in a real-time monitoring state throughout the closed-loop control process, and the triggering conditions include operation deviation exceeding the threshold, sensor failure, thruster failure, robotic arm jamming, communication interruption, collision risk, and depth exceeding the range, and it will be executed immediately once any triggering condition is met.

Images