A bionic robot obstacle avoidance control method for deep sea net cage culture monitoring

Abstract

The application discloses a kind of bionic robot obstacle avoidance control methods for deep-sea net cage culture monitoring.The method uses hierarchical modular design, first perception layer synchronously acquires multi-source sensor information and carries out pre-processing, to construct standardized environment state vector;Decision-making layer is based on the dynamic threat evaluation method based on improved virtual potential field, based on the safety cost of action to evaluate its threat.Then, construct the multi-objective comprehensive cost function of fusion safety cost, energy cost and smooth cost;Through the hybrid intelligent decision maker based on multi-objective comprehensive cost function, the optimal action decision is output;Finally, the abstract action instruction of decision output is mapped and converted into specific bottom control signal.The application realizes safe, smooth and energy-efficient real-time obstacle avoidance without constructing global environment map, effectively solves the problem that traditional method is easy to fall into local minimum, control chattering and poor adaptability under limited computing power.

Description

Technical Field

[0001] This invention belongs to the field of autonomous control technology for underwater robots, specifically relating to a control method for real-time obstacle avoidance decision-making based on multi-source information fusion and multi-objective cost functions for biomimetic underwater robots with limited computing resources. Background Technology

[0002] When performing monitoring tasks in deep-sea cage aquaculture, biomimetic underwater robots need to autonomously navigate and collect data in complex environments involving net structures, aquaculture equipment, and fish schools. Reliable onboard safety is essential, with autonomous obstacle avoidance being particularly critical. Unlike general path planning methods that rely on high-precision, high-power sensors (such as forward-looking sonar and 3D LiDAR) and global map construction, small, biomimetic underwater robots generally face engineering constraints such as limited payload, restricted sensor power consumption in the high-pressure environment of the deep sea, and insufficient computing power of embedded processors. Therefore, researching a lightweight, reactive, real-time obstacle avoidance strategy has significant practical engineering value.

[0003] Traditional artificial potential field methods are widely used due to their computational simplicity, but their inherent drawbacks, such as local minima, stagnation or oscillation near the target point due to the cancellation of attraction and repulsion, and inconvenience in discrete motion control systems, limit their performance in complex environments. Furthermore, existing methods often rely on single sensors, making it difficult to provide robust sensing information in dynamic, unstructured underwater environments; while multi-sensor fusion methods are computationally complex and difficult to run in real-time on embedded platforms. Therefore, for deep-sea cage aquaculture monitoring scenarios, a lightweight obstacle avoidance control method is needed that can fuse limited sensing information, make rapid intelligent decisions, and output smooth, stable, and safe actions. Summary of the Invention

[0004] To achieve safe, smooth, and energy-efficient real-time obstacle avoidance without constructing a global map, and to address the technical problems of poor perception reliability, susceptibility to local minima, and decision-making oscillation in existing technologies under limited computing power, this invention provides a biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture. The biomimetic robot includes at least: a monocular camera mounted on the longitudinal central axis of its head, an ultrasonic sensor, an inertial measurement unit, and a depth sensor, as well as a thruster or articulated servo motor for driving motion.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] In a first aspect, embodiments of this application provide a biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture, the specific implementation steps of which are as follows:

[0007] Step 1: Synchronously collect information from multiple sensor sources and perform preprocessing to obtain preprocessed multi-source sensing data.

[0008] Step 2: Constructing the environmental state vector: Based on the preprocessed multi-source sensing data, construct a standardized environmental state vector.

[0009] Step 3: Action-oriented dynamic threat assessment: The threat level is assessed based on the security cost of actions using a dynamic threat assessment method based on an improved virtual potential field.

[0010] Step 4: Multi-objective integrated cost calculation: After obtaining the safety cost of each candidate action, energy consumption cost and smoothing cost are further introduced to construct a multi-objective integrated cost function.

[0011] Step 5: Output the optimal action decision: The optimal action decision is output through the designed hybrid intelligent decision-maker based on the multi-objective integrated cost function.

[0012] Step Six: Generation and Execution of Low-Level Control Commands: The abstract action commands output by the decision are mapped and converted into specific low-level control signals, including thruster thrust commands and joint servo motor angle commands. The robot is driven to perform corresponding obstacle avoidance maneuvers through the low-level control signals to complete closed-loop obstacle avoidance control.

[0013] In one possible implementation, the synchronous acquisition and preprocessing of multi-source sensor information is specifically performed as follows:

[0014] During each operational sensing cycle, raw data from the ultrasonic sensor, monocular camera, inertial measurement unit (IMU), and depth sensor onboard the biomimetic underwater robot are simultaneously acquired. The data acquired by the ultrasonic sensor is the precise radial distance from its beam center to the nearest reflecting surface. The data acquired by the depth sensor is the robot's actual distance from the horizontal plane. The monocular camera acquires a two-dimensional intensity image of the scene in front of the robot. The inertial measurement unit (IMU) acquires the robot's three-axis attitude angles (including roll angle). Pitch angle Yaw angle ) and triaxial angular velocity.

[0015] The preprocessing process is as follows: First, the two-dimensional intensity image is binarized, and then morphological opening is performed to eliminate noise points and small holes, resulting in a smoothed binary image. Finally, the maximum connected contour in the binary image is extracted and fitted with the minimum bounding rectangle to obtain the center coordinates, pixel width, and height of the obstacle in the image pixel coordinate system.

[0016] In one possible implementation, the environmental state vector is constructed as follows: based on the preprocessed multi-source perception data, the fused distance, horizontal azimuth angle and vertical azimuth angle of the obstacle relative to the robot are obtained by fusing forward distance information and solving azimuth information, and combined with the robot's actual distance from the horizontal plane and yaw angle, a standardized environmental state vector S is constructed.

[0017] In one possible implementation, the forward distance information fusion is specifically as follows: forward distance information is fused based on data obtained from a monocular camera and an ultrasonic sensor, and visual ranging confidence and ultrasonic ranging confidence are defined separately for weighted calculation of the fused distance:

[0018] Visual ranging confidence The definition is as follows:

[0019]

[0020] in, It is the pixel width of the obstacle in the image pixel coordinate system. It is the width of the two-dimensional intensity image captured by the camera. The x-coordinate of the image center. The parameter is used to adjust the weight of the horizontal concentration.

[0021] Ultrasonic confidence The value is 1 when the obstacle is within the range of ultrasonic detection, and 0 otherwise.

[0022] Monocular visual ranging value Based on the pinhole model and prior physical width calculate:

[0023]

[0024] in, This refers to the camera's focal length.

[0025] Precise radial distance acquired by the ultrasonic sensor As the ranging value measured by the ultrasonic sensor. In summary, the fused distance... The result was obtained by weighted averaging based on visual ranging confidence and ultrasonic confidence:

[0026]

[0027] in, It is a small constant added to prevent division by zero.

[0028] In one possible implementation, the orientation information is calculated as follows: First, the coordinates of the center position of the obstacle's pixel range are calculated using the minimum bounding rectangle method. And the distance from the minimum bounding rectangle of the obstacle to the perimeter of the two-dimensional intensity image. Then the horizontal azimuth angle and vertical azimuth Calculated using the following formula:

[0029]

[0030] in, A positive value indicates that the obstacle is located slightly to the left in front of the robot, while a negative value indicates that it is slightly to the right in front of the robot. A positive value indicates that the obstacle is located slightly above the front of the robot, while a negative value indicates that it is located slightly below the front.

[0031] In one possible implementation, the dynamic threat assessment method includes three sub-steps: constructing an improved virtual potential field model, prospective pose prediction, and security cost calculation.

[0032] Sub-step 1: Construct an improved virtual potential field model

[0033] First, define the improved repulsive potential function. :

[0034]

[0035] in, The fusion distance from the robot to the obstacle. The radius of influence of the potential field. This is the repulsive gain coefficient, used to adjust the overall strength of the potential field. This is the distance adjustment index. When... When, it degenerates into a traditional quadratic potential field; when At that time, potential energy It grows faster when it decreases.

[0036] Sub-step 2, Forward pose prediction:

[0037] Define a set of discrete obstacle avoidance actions F, L, R, U, and D represent five candidate maneuvers: going straight, turning left, turning right, surfacing, and diving, respectively. For each non-straight-go candidate maneuver... Based on the current environmental state vector S, the changes in each component of the state vector after the action is performed are predicted through pre-established kinematic relationships, thus obtaining a set of state changes.

[0038] Sub-step 3: Calculation of security costs:

[0039] (1) Basic security cost calculation: Based on the predicted set of state changes, the execution action is calculated. The predicted distance is then quantified. Substituting this predicted distance into the improved repulsive potential function yields the predicted repulsive potential energy after performing the action, which serves as the basic security cost. .

[0040] (2) Geometric safety cost addition: The robot is simplified into a rectangular bounding box of known size. The distance from each corner point of the bounding box to the obstacle after the action is performed is calculated, and the minimum distance is taken. We calculate the repulsive potential energy as an additional geometric safety cost.

[0041] Therefore, the full cost of security Defined as:

[0042]

[0043] in, It is a penalty coefficient greater than 1.

[0044] In one possible implementation, the multi-objective comprehensive cost calculation specifically operates as follows: After obtaining the safety cost of each candidate action... Then, energy consumption costs were further introduced. and smoothing cost A multi-objective comprehensive cost function is constructed to achieve a comprehensive evaluation of candidate actions.

[0045] Among them, the energy consumption cost The definition is as follows:

[0046]

[0047] Among them, for smoothing cost Defined as the current candidate action Actions performed in the previous cycle Functions of the degree of difference:

[0048]

[0049] This indicates switching between different types of actions. It indicates the switching between opposite actions of the same type.

[0050] Finally, the multi-objective integrated cost function is constructed:

[0051]

[0052] in, These are the weight coefficients for the costs of each dimension, satisfying... . This represents the value after normalizing the original cost.

[0053] In one possible implementation, the hybrid intelligent decision-maker first performs a rapid threat assessment. If a threat is triggered, it enters a detailed obstacle avoidance process, which includes action pre-screening and comprehensive cost calculation.

[0054] Sub-step 1: Rapid Threat Assessment and Decision Triggering: After obtaining the environmental state vector S, a rapid assessment is first performed, and the fused distance is... With the preset safety threshold Compare them.

[0055] like If no immediate collision threat is detected, the decision-maker will directly output a "go straight" action, generate and execute underlying control commands, and end the current decision cycle.

[0056] like If the robot is detected, it is determined that it has entered the threat range of the obstacle, triggering a detailed obstacle avoidance decision-making process.

[0057] Sub-step 2: Action pre-screening: Before entering the comprehensive cost calculation, based on the current environment state vector S, an action pre-screening mechanism is performed according to the following rules to exclude actions that are obviously unreasonable or physically infeasible in the current state:

[0058] If the current height above the water surface is less than the minimum safe height, then an ascent is ruled out; if the current height above the water surface is greater than the maximum safe height, then a descent is ruled out.

[0059] After pre-screening, the remaining actions constitute a subset of candidate actions for the current period. Subsequent steps only calculate the comprehensive cost for the actions in the subset of candidate actions.

[0060] Sub-step 3: Based on the multi-objective comprehensive cost calculation method, calculate the comprehensive cost of each candidate action subset, and select the action with the smallest comprehensive cost as the output action of the current decision cycle.

[0061] If multiple actions have the same cost, the action that is the same as the action in the previous cycle should be selected first to maintain decision continuity; if all actions are different from the previous cycle, the action should be selected according to the priority of left turn, right turn, surfacing, and diving.

[0062] Secondly, embodiments of this application provide a biomimetic robot obstacle avoidance control system for monitoring deep-sea cage aquaculture, comprising the following modules:

[0063] Data acquisition module: synchronously collects information from multiple sources of sensors and performs preprocessing to obtain preprocessed multi-source sensing data.

[0064] Environmental state vector construction module: Based on preprocessed multi-source sensing data, construct standardized environmental state vectors.

[0065] Dynamic Threat Assessment Module: This module assesses the threat level of an action based on the security cost of the action using a dynamic threat assessment method based on an improved virtual potential field.

[0066] Cost Calculation Module: Cost calculation is performed based on a multi-objective comprehensive cost function constructed from security cost, energy consumption cost, and smoothing cost.

[0067] Decision output module: The optimal action decision is output through a hybrid intelligent decision-maker based on a multi-objective integrated cost function.

[0068] Command generation and execution module: Maps and converts the abstract action commands output by the decision into specific low-level control signals, including thruster thrust commands and joint servo motor angle commands; drives the robot to perform corresponding obstacle avoidance maneuvers through the low-level control signals, and completes closed-loop obstacle avoidance control.

[0069] Thirdly, embodiments of this application provide an electronic device, including a processor and a memory;

[0070] The memory is used to store computer programs.

[0071] When the processor executes the program stored in the memory, it implements any of the bionic robot obstacle avoidance control methods described in this application.

[0072] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements any of the bionic robot obstacle avoidance control methods described in this application.

[0073] Fifthly, embodiments of this application provide a computer program product containing instructions that, when run on a computer, cause the computer to execute any of the bionic robot obstacle avoidance control methods described in this application.

[0074] The beneficial effects of this invention are as follows:

[0075] 1. To address the issues of poor robustness of single-sensor perception and complex computation of multi-source fusion, this invention proposes a lightweight confidence-weighted fusion strategy. This strategy significantly reduces computational complexity while ensuring fusion accuracy, enabling the algorithm to run in real time at frequencies above 20Hz on embedded platforms such as Jetson Nano, effectively solving the problem of poor perception reliability under limited computing power.

[0076] 2. In response to the problem that traditional potential field methods are prone to getting trapped in local minima and the target is unreachable, this invention proposes an action-oriented dynamic threat assessment mechanism based on an improved potential field method. The mechanism transforms the obstacle avoidance decision problem from "continuous potential field optimization" to "discrete action cost comparison", which fundamentally avoids the inherent problem of traditional potential field methods being prone to getting trapped in local minima and ensures that the robot can smoothly reach the target position.

[0077] 3. Existing obstacle avoidance methods mostly involve continuous velocity space programming (linear velocity + angular velocity). The output results require additional mapping and transformation before they can be applied to the discrete control mode of biomimetic robots, and this transformation process is prone to introducing errors. This invention directly uses the inherent discrete movements of the biomimetic underwater robot (straight ahead, left turn, right turn, surfacing, diving) as the decision space. A one-to-one correspondence is formed between the decision output and the actuator's drive commands, eliminating the need for intermediate transformation steps. This integrated "decision-execution" design simplifies the control chain, reduces error accumulation, and makes the algorithm structure more concise and easier to implement in engineering. Attached Figure Description

[0078] Figure 1 This is a schematic diagram illustrating the calculation of the horizontal azimuth angle between the robot and the obstacle in an embodiment of the present invention.

[0079] Figure 2 This is a schematic diagram of an underwater experimental platform according to an embodiment of the present invention.

[0080] Figure 3 This is a top-down view of the process of turning left to avoid obstacles according to an embodiment of the present invention.

[0081] Figure 4 This is a graph showing the variation of roll angle and yaw angle during left-turn obstacle avoidance in an embodiment of the present invention.

[0082] Figure 5 This is a top-down view of the obstacle avoidance process during a right turn, according to an embodiment of the present invention.

[0083] Figure 6 This is a graph showing the changes in roll angle and yaw angle during right-turn obstacle avoidance in an embodiment of the present invention.

[0084] Figure 7 This is a diagram of the floating obstacle avoidance process according to an embodiment of the present invention, where (a) is a top view and (b) is a side view.

[0085] Figure 8 This is a graph showing the changes in pitch angle and depth during obstacle avoidance in an embodiment of the present invention.

[0086] Figure 9 This is a diagram of the obstacle avoidance process during the diving in an embodiment of the present invention, where (a) is a top view and (b) is a side view.

[0087] Figure 10This is a graph showing the pitch angle and depth variation during obstacle avoidance in an embodiment of the present invention.

[0088] Figure 11 This is a diagram of the integrated obstacle avoidance control architecture of the "perception-decision-execution" three-layer structure proposed in this invention. Detailed Implementation

[0089] The present invention will now be described in further detail with reference to the accompanying drawings and embodiments. It should be noted that the following embodiments are only used to more clearly illustrate the technical solutions of the present invention and should not be construed as limiting the scope of protection of the present invention.

[0090] The proposed "perception-decision-execution" three-layer integrated obstacle avoidance control architecture adopts a layered modular design, including: a perception layer, a decision layer, and a control layer. The perception layer is responsible for periodically collecting and fusing multi-source data at a fixed frequency, outputting a standardized environmental state vector. The decision layer, composed of a dynamic threat assessment module and a multi-objective optimization decision-maker connected in series, is responsible for calculating the comprehensive cost of each candidate action and outputting the optimal action. The control layer transforms abstract action commands into drive signals for specific execution mechanisms through control law calculations. The specific implementation steps of the above control architecture within one cycle are as follows:

[0091] Step 1: Synchronous Acquisition and Preprocessing of Multi-Source Sensor Information: During each operational sensing cycle, raw data from the ultrasonic sensor, monocular camera, inertial measurement unit, and depth sensor mounted on the biomimetic underwater robot are acquired synchronously; the raw data is preprocessed to obtain preprocessed multi-source sensing data.

[0092] The data acquired by the ultrasonic sensor is the precise radial distance from its beam center to the nearest reflecting surface. This serves as the primary basis for determining the robot's distance from obstacles. Data collected by the depth sensor represents the robot's true distance from the horizontal plane. This serves as a reference for the robot's actions. A two-dimensional intensity image of the scene in front of the robot is acquired using a monocular camera. The robot's three-axis attitude angles (roll) are obtained using an inertial measurement unit (IMU). , up and down ,yaw ) and triaxial angular velocity ω.

[0093] The preprocessing procedure is as follows: First, the two-dimensional intensity image is binarized. Then, morphological opening operations are performed to eliminate noise points and small holes, resulting in a smoothed binary image. Finally, the maximum connected contour in the binary image is extracted and fitted with the minimum bounding rectangle to obtain the center coordinates of the obstacle in the image pixel coordinate system. and pixel width With height .

[0094] Step 2, Environmental State Vector Construction: Based on the preprocessed multi-source perception data, the fused distance, horizontal azimuth, and vertical azimuth of the obstacle relative to the robot are obtained through forward distance information fusion and orientation information calculation. Combined with the robot's actual distance from the horizontal plane and yaw angle, a standardized environmental state vector S is constructed.

[0095] First, forward distance information is fused based on data obtained from the monocular camera and ultrasonic sensor. Specifically, the fused distance *d* between the robot and the obstacle ahead is achieved through the complementary use of high precision ultrasonic measurement and wide field of view visual measurement. Visual ranging confidence and ultrasonic ranging confidence are defined separately for the weighted calculation of the fused distance.

[0096] Visual ranging confidence The definition is as follows:

[0097]

[0098] in, It is the pixel width of the obstacle in the image pixel coordinate system obtained in step one. It is the width of the two-dimensional intensity image captured by the camera (this value is equivalent to the pre-processed binarized image). The x-coordinate of the image center. The parameters are used to adjust the horizontal concentration weights. When the obstacle is centered and the pixel size is large, the visual ranging confidence level... The value is high.

[0099] Ultrasonic confidence The value is 1 when the obstacle is within the range of ultrasonic detection, and 0 otherwise.

[0100] Monocular visual ranging value Based on the pinhole model and prior physical width calculate:

[0101]

[0102] in, This refers to the camera's focal length.

[0103] Precise radial distance acquired by the ultrasonic sensor As the ranging value measured by the ultrasonic sensor. In summary, the fused distance... The result was obtained by weighted averaging based on visual ranging confidence and ultrasonic confidence:

[0104]

[0105] in, This is a small constant ε added to prevent division by zero (the value of the small constant ε ranges from 0.001 to 0.01, preferably 0.005). This strategy ensures that precise calculations are prioritized when facing an obstacle. When approaching at an angle, the transition to visual estimation is smooth. .

[0106] Then, the orientation information is calculated, and the calculation method is as follows: Figure 1 As shown, the monocular camera is fixed at the center of the robot's head, so the robot can be considered to be at the center of the image. First, the coordinates of the center position of the obstacle's pixel range are calculated using the minimum bounding rectangle method. And the distance from the minimum bounding rectangle of the obstacle to the perimeter of the two-dimensional intensity image. Then the horizontal azimuth angle and vertical azimuth Calculated using the following formula:

[0107]

[0108] in, A positive value indicates that the obstacle is located slightly to the left in front of the robot, while a negative value indicates that it is slightly to the right in front of the robot. A positive value indicates that the obstacle is located slightly above the front of the robot, while a negative value indicates that it is located slightly below the front.

[0109] Based on the above processing steps and combined with the robot's actual distance from the horizontal plane... Yaw angle obtained from inertial measurement unit (IMU) Within each operational sensing cycle, a complete environmental state vector is output. :

[0110]

[0111] Step 3: Action-Oriented Dynamic Threat Assessment: Design a dynamic threat assessment method based on an improved virtual potential field to assess the threat level based on the security cost of actions. The dynamic threat assessment method includes three sub-steps: constructing an improved virtual potential field model, prospective pose prediction, and security cost calculation.

[0112] Sub-step 1: Construct an improved virtual potential field model

[0113] First, define the improved repulsive potential function. :

[0114]

[0115] in, The fusion distance from the robot to the obstacle. The radius of influence of the potential field. This is the repulsive gain coefficient, used to adjust the overall strength of the potential field. This is the distance adjustment index. When... When, it degenerates into a traditional quadratic potential field; when At that time, potential energy Growth is faster when it decreases, → When it approaches infinity at 0, it ensures that the robot will not collide with obstacles.

[0116] Compared with traditional quadratic potential fields, this invention introduces a 1 / d term instead of a 1 / d² term, which makes the potential energy grow more steeply when it is near an obstacle, generating a stronger repulsion warning and prompting the robot to take avoidance actions earlier and more decisively.

[0117] Sub-step 2, Forward pose prediction:

[0118] Define a set of discrete obstacle avoidance actions F, L, R, U, and D represent five candidate maneuvers: going straight, turning left, turning right, surfacing, and diving, respectively. For each non-straight-go candidate maneuver... Based on the current environmental state vector S, the changes in each component of the state vector after performing the action are predicted through pre-established kinematic relationships, thus obtaining a set of state changes. ,in For fusion distance The change Horizontal azimuth The transformation quantity, Vertical azimuth The change The actual height above the water surface The change Yaw angle The change The distance from the minimum bounding rectangle of the obstacle to the perimeter of the two-dimensional intensity image. The change in quantity.

[0119] Sub-step 3: Calculation of security costs:

[0120] (1) Calculation of basic security cost: based on the set of state changes predicted in sub-step 2 Calculate the action to be executed Predicted distance Substituting this predicted distance into the improved repulsive potential function yields the predicted repulsive potential energy after performing the action, which serves as the basic security cost. :

[0121]

[0122] This prediction repulsive potential energy Characterizes the action to be performed and duration Then, the magnitude of the collision risk faced by the robot.

[0123] (2) Additional Geometric Safety Costs: Considering the non-zero geometric shape of the robot, only considering the center-of-gravity distance may result in scraping and collisions to the robot body even though the center of gravity is safe. Therefore, the robot is simplified into a rectangular bounding box of known dimensions. The distances from each corner point of the bounding box (especially the front corner closest to the obstacle) to the obstacle after the action is performed are calculated, and the minimum distance is taken. The repulsive potential energy is calculated at the most dangerous point as an additional geometric safety cost.

[0124] Therefore, the full cost of security It can be defined as:

[0125]

[0126] in, It is a penalty coefficient greater than 1, used to amplify the additional risks caused by the robot's shape. This evaluation mechanism gives the decision-making system a certain "spatial awareness," enabling it to avoid situations where the center of mass is safe but the robot body is scraped or collided.

[0127] Step 4: Multi-objective comprehensive cost calculation: After obtaining the safety cost of each candidate action... Then, energy consumption costs were further introduced. and smoothing cost A multi-objective comprehensive cost function is constructed to achieve a comprehensive evaluation of candidate actions.

[0128] Among them, the energy consumption cost The energy consumption is calibrated based on the estimated energy consumption or time cost of the thrusters performing different actions. Straight-line motion has the lowest energy cost, while turning and vertical plane maneuvers have relatively higher energy costs. The energy cost is defined as follows:

[0129]

[0130] Among them, for smoothing cost To avoid robot motion jitter, actuator wear, and energy waste caused by high-frequency switching of control commands, it is necessary to encourage the smoothness of action sequences over time. The smoothness cost is defined as the current candidate action. Actions performed in the previous cycle Functions of the degree of difference:

[0131]

[0132] This indicates switching between different types of movements, such as switching from vertical plane movements (surfacing, diving) to horizontal plane movements (turning left, turning right). This indicates the switching between opposite actions of the same type, such as switching between surfacing and diving, or between turning left and turning right. This significantly reduces arbitrary decision-making, avoids large swaying of the robot, and makes the robot's movements more consistent and stable.

[0133] Finally, the multi-objective integrated cost function is constructed: Integrated Cost Function The design aims to quantitatively evaluate the overall performance of an action across three key dimensions: Safety, Energy, and Smoothness, which are defined as follows:

[0134]

[0135] in, These are the weight coefficients for the costs of each dimension, satisfying... . This represents the value after normalizing the original cost to eliminate differences in the dimensions and orders of magnitude of different cost terms.

[0136] Step 5: Optimal Action Decision Output: Design a hybrid intelligent decision-maker based on a multi-objective comprehensive cost function. This involves advanced and rapid threat assessment; if a threat is triggered, a detailed obstacle avoidance process is initiated, including action pre-screening and comprehensive cost calculation.

[0137] Sub-step 1: Rapid Threat Assessment and Decision Triggering: After obtaining the environmental state vector S, a rapid assessment is first performed, and the fused distance is... With the preset safety threshold Compare them.

[0138] like If no immediate collision threat is detected, the decision-maker will directly output a "go straight" action, generate and execute underlying control commands, and end the current decision cycle to reduce computational load and improve response efficiency.

[0139] like If the robot is detected, it is determined that it has entered the threat range of the obstacle, triggering a detailed obstacle avoidance decision-making process.

[0140] Sub-step 2: Action pre-screening: Before entering the comprehensive cost calculation, based on the current environment state vector S, an action pre-screening mechanism is performed according to the following rules to exclude actions that are obviously unreasonable or physically infeasible in the current state:

[0141] The current height h above the water surface is less than the minimum safe height h. minIf there is no room to rise above the surface, then the attempt to rise is ruled out; if the current height h above the surface is greater than the maximum safe height h... max If the cage has already reached the bottom and there is no room for it to descend further, then the descent action is ruled out.

[0142] After pre-screening, the remaining actions constitute a subset A of candidate actions for the current cycle. sub ⊆ A, subsequent steps only apply to A. sub The calculation of the overall cost of actions in the process is used to reduce computational overhead.

[0143] Sub-step 3: Based on the multi-objective comprehensive cost calculation method in step 4, perform the calculation on the candidate action subset A. sub The comprehensive cost of each candidate action is calculated. Choose the action with the lowest overall cost. This is the output action of the current decision-making cycle.

[0144] If multiple actions have the same cost, the action that is the same as the action in the previous cycle should be selected first to maintain decision continuity; if all actions are different from the previous cycle, the action should be selected according to the priority of left turn, right turn, surfacing, and diving.

[0145] Step Six: Generation and Execution of Low-Level Control Commands: The abstract action command a* output by the decision is mapped and converted (by calculation using control laws such as PID control) into specific low-level control signals. These low-level control signals include thruster thrust commands and joint servo motor angle commands. The robot is then driven to perform corresponding obstacle avoidance maneuvers through these low-level control signals, thus completing closed-loop obstacle avoidance control.

[0146] The embodiments of the present invention are carried out in a pool environment, such as... Figure 2 As shown, the pool is 5 meters long, 4 meters wide, and 2 meters deep. The robot communicates with the host computer via a lightweight, zero-buoyancy cable and is in a free-swimming state; the cable's drag force is negligible. The experiment was conducted in a still water environment, without introducing artificial water flow or external disturbances such as waves. The parameters of the beaver-inspired underwater robot platform are as described above, and the core processor is a Jetson Nano. Initially, a 1 m × 0.5 m black acrylic sheet was fixed in the center of the pool as an obstacle.

[0147] The key parameter settings for the decision model in the implementation example are as follows: radius of influence of potential field. Safety threshold Repulsive gain Adjustment index. The weights of the multi-objective cost function are set as follows: .

[0148] Specific implementation process: The robot is equipped with the "perception-decision-execution" three-layer integrated obstacle avoidance control architecture proposed in this invention. The control architecture adopts a layered modular design, including: a perception layer, a decision layer, and a control layer. After the robot enters the water, it first performs system initialization and loads sensor parameter calibration parameters. Then, the perception layer periodically collects various sensor data at a fixed frequency, and the ultrasonic wave measures the precise radial distance from its beam center direction to the nearest reflecting surface. A monocular camera acquires two-dimensional intensity images. The acquired images undergo preprocessing: the RGB two-dimensional intensity images are converted to the HSV color space, and threshold segmentation is used to obtain binary images. After morphological opening operations to eliminate noise points, the maximum connected contour is extracted and fitted with the minimum bounding rectangle to obtain the pixel width and center coordinates of the obstacle in the image. The IMU outputs the current attitude angle and roll angle. , up and down ,yaw The depth sensor measured the robot's actual distance from the water surface. .

[0149] The confidence scores for visual ranging and ultrasonic ranging are further calculated from the above data, and the fused distance is obtained by weighted averaging. ; Calculate the horizontal and vertical azimuth angles; Construct the environmental state vector .

[0150] The decision layer extracts the fusion distance from the current state vector. , and safety threshold Compare the meters. When Continue straight ahead. Then proceed to the detailed obstacle avoidance decision-making process.

[0151] For a pre-defined set of actions A = {straight ahead F, left turn L, right turn R, ascend U, descend D}, perform action pre-screening, and then select the subset of candidate actions A. sub The comprehensive cost of each obstacle avoidance action is evaluated based on the comprehensive cost function, and the action with the minimum comprehensive cost is selected. As the output action of the current decision cycle, after receiving the action command, the control layer uses the PID control algorithm to calculate the target state quantity, linearly maps it into a PWM signal and sends it to each servo motor for execution, thus ending one obstacle avoidance decision cycle.

[0152] Example 1: Horizontal obstacle avoidance

[0153] (1) Left turn obstacle avoidance: In this example, the robot turns left from the right front of the obstacle (i.e., the initial horizontal azimuth angle). Depart and approach the target along a predetermined straight course. For example... Figure 3As shown, after the robot starts from the starting point, the perception system continuously constructs an environmental state vector. In the initial stage, the fusion distance... Greater than the safety threshold The decision-maker continuously outputs the "go straight" action, and the robot maintains its straight-line navigation.

[0154] When the robot moves to the fusion distance Less than the safety threshold When this happens, a detailed decision-making process is triggered. At this point, the decision-maker bases its decisions on the current environmental state vector. Calculations revealed that the obstacle was located in front of the robot's right side. The decision-maker calculates the overall cost of each candidate action. The result shows that the left turn (L) action has the lowest safety cost (because turning left will guide the robot away from the direction of the obstacle) and the lowest overall cost. Therefore, the "turn left" command is output.

[0155] Figure 4 Recorded real-time robot posture data confirmed the aforementioned behavioral logic. After obstacle avoidance was triggered, the yaw angle began to change due to the robot's left turn, and simultaneously, due to the coupling nature of the robot's motion, the roll angle changed accordingly. Once the robot completed obstacle avoidance and the obstacle was removed from the threat area, the decision-maker output a "go straight" command, the robot's yaw angle stabilized, and it resumed straight-line navigation. Throughout the entire process, the robot safely bypassed the obstacle from the left without any collision.

[0156] (2) Right-turn obstacle avoidance experiment: As a symmetry verification, a right-turn obstacle avoidance experiment was conducted. The robot started from the left front of the obstacle, with an initial horizontal azimuth angle of... (The obstacle is on the left), and other conditions are the same as in the left-turn experiment. The purpose is to verify the stability of the same decision logic under different initial conditions.

[0157] like Figure 5 As shown, the robot starts slightly to the right of the obstacle's centerline. After the obstacle enters a safe distance, the decision-maker performs an evaluation based on a multi-objective cost function. At this point, due to the obstacle's initial azimuth angle... The cost calculation results show that the right turn (R) action is the optimal solution. After the decision, a right turn command is output and executed. Figure 6 The trend of its yaw angle change is shown to be a mirror image of that of the left turn experiment.

[0158] All experiments in Example 1 achieved collision-free obstacle avoidance, and the decision-maker performed stably and consistently under different initial conditions, verifying the effectiveness and robustness of the present invention in horizontal obstacle avoidance scenarios.

[0159] Example 2: Vertical Plane Obstacle Avoidance

[0160] This embodiment was conducted in a water pool environment to verify the obstacle avoidance decision-making and execution capabilities of the present invention in the vertical direction. The experimental obstacle was a 1m × 0.5m black acrylic panel, horizontally fixed at the front of the pool, approximately 0.5 meters above the water surface.

[0161] (1) Floating obstacle avoidance: The robot starts from the upper front of the obstacle, with an initial vertical azimuth angle. ° (The obstacle is located slightly below and in front of the robot), approach the target in a straight line along the set course. (Top-down view as follows) Figure 7 As shown in (a), when the robot is less than the distance to the obstacle The threat assessment condition is triggered. The decision-maker makes a judgment based on the current environmental state vector, and outputs an upward float (U) as the current optimal obstacle avoidance action, after which the robot executes the instruction.

[0162] Figure 8 This is a graph showing the robot's pitch and depth changes. Before the obstacle avoidance command is triggered, the robot's pitch angle remains relatively stable. After the obstacle avoidance maneuver is executed, the robot's pitch angle changes significantly. Combined with the change in depth, this indicates that the robot is rising to avoid the obstacle. Key obstacle avoidance images are shown below. Figure 7 As shown.

[0163] (2) Diving to avoid obstacles: As a symmetrical group, such as Figure 9 In the demonstration experiment of obstacle avoidance, the robot started from the lower front side of the obstacle and approached the target along a set course. After triggering a threat assessment, the robot issued a dive command and bypassed the obstacle from below.

[0164] Figure 10 This is a diagram showing the changes in pitch and depth. Figure 10 Data change trends and Figure 8 The basic symmetry of the proposed decision-maker was verified, confirming its stability under different initial conditions.

[0165] This application also provides a biomimetic robot obstacle avoidance control system for monitoring deep-sea cage aquaculture, including the following modules:

[0166] Data acquisition module: synchronously collects information from multiple sources of sensors and performs preprocessing to obtain preprocessed multi-source sensing data.

[0167] Environmental state vector construction module: Based on preprocessed multi-source sensing data, construct standardized environmental state vectors.

[0168] Dynamic Threat Assessment Module: This module assesses the threat level of an action based on the security cost of the action using a dynamic threat assessment method based on an improved virtual potential field.

[0169] Cost Calculation Module: Cost calculation is performed based on a multi-objective comprehensive cost function constructed from security cost, energy consumption cost, and smoothing cost.

[0170] Decision output module: The optimal action decision is output through a hybrid intelligent decision-maker based on a multi-objective integrated cost function.

[0171] Command generation and execution module: Maps and converts the abstract action commands output by the decision into specific low-level control signals, including thruster thrust commands and joint servo motor angle commands; drives the robot to perform corresponding obstacle avoidance maneuvers through the low-level control signals, and completes closed-loop obstacle avoidance control.

[0172] This application also provides an electronic device, including a processor and a memory.

[0173] The memory is used to store computer programs.

[0174] When the processor executes a program stored in the memory, it implements any of the methods described in this application.

[0175] In one possible implementation, the electronic device of this application embodiment further includes a communication interface and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus.

[0176] The communication bus mentioned in the above electronic devices can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0177] The communication interface is used for communication between the aforementioned electronic devices and other devices.

[0178] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0179] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0180] In another embodiment provided in this application, a computer-readable storage medium is also provided, which stores a computer program that, when executed by a processor, implements any of the methods described in this application.

[0181] In another embodiment provided in this application, a computer program product containing instructions is also provided, which, when run on a computer, causes the computer to perform any of the methods described in this application.

[0182] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0183] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0184] The various embodiments in this specification are described in a related manner. Each embodiment focuses on the differences from other embodiments, and the same or similar parts between the various embodiments can be referred to each other.

[0185] The above description is merely a preferred embodiment of this application and is not intended to limit the scope of protection of this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application are included within the scope of protection of this application.

Claims

1. A biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture, characterized in that, The specific implementation steps are as follows: Step 1: Synchronously collect information from multiple sources of sensors and perform preprocessing to obtain preprocessed multi-source sensing data; Step 2: Constructing the environmental state vector: Based on the preprocessed multi-source sensing data, construct a standardized environmental state vector; Step 3, Action-Oriented Dynamic Threat Assessment: Using a dynamic threat assessment method based on an improved virtual potential field, the threat level is assessed based on the security cost of the action. Step 4: Multi-objective comprehensive cost calculation: After obtaining the safety cost of each candidate action, energy consumption cost and smoothing cost are further introduced to construct a multi-objective comprehensive cost function; Step 5: Output the optimal action decision: Output the optimal action decision through the designed hybrid intelligent decision-maker based on a multi-objective integrated cost function; Step Six: Generation and Execution of Low-Level Control Commands: The abstract action commands output by the decision are mapped and converted into specific low-level control signals, including thruster thrust commands and joint servo motor angle commands. The robot is driven to perform corresponding obstacle avoidance maneuvers through the low-level control signals to complete closed-loop obstacle avoidance control.

2. The biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 1, characterized in that, The specific steps for synchronous acquisition and preprocessing of multi-source sensor information are as follows: During each operational sensing cycle, raw data from the ultrasonic sensor, monocular camera, inertial measurement unit, and depth sensor mounted on the biomimetic underwater robot are collected synchronously. The data collected by the ultrasonic sensor is the precise radial distance from its beam center to the nearest reflecting surface; the data collected by the depth sensor is the robot's actual distance from the horizontal plane; the monocular camera acquires a two-dimensional intensity image of the scene in front of the robot; and the inertial measurement unit acquires the robot's three-axis attitude angles and three-axis angular velocities. The preprocessing process is as follows: First, the two-dimensional intensity image is binarized, and then morphological opening is performed to eliminate noise points and small holes, resulting in a smoothed binary image. Finally, the maximum connected contour in the binary image is extracted and fitted with the minimum bounding rectangle to obtain the center coordinates, pixel width, and height of the obstacle in the image pixel coordinate system.

3. The biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 2, characterized in that, The specific operation for constructing the environmental state vector is as follows: Based on the preprocessed multi-source perception data, the fused distance, horizontal azimuth angle and vertical azimuth angle of the obstacle relative to the robot are obtained by fusing forward distance information and solving azimuth information. Combined with the robot's actual distance from the horizontal plane and yaw angle, a standardized environmental state vector S is constructed.

4. The biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 3, characterized in that, The forward distance information fusion is performed as follows: Forward distance information is fused based on data obtained from the monocular camera and ultrasonic sensor. Visual ranging confidence and ultrasonic ranging confidence are defined separately for weighted calculation of the fused distance. Visual ranging confidence The definition is as follows: in, It is the pixel width of the obstacle in the image pixel coordinate system. It is the width of the two-dimensional intensity image captured by the camera. The x-coordinate of the image center. Parameters for adjusting the weight of horizontal concentration; Ultrasonic confidence The value is 1 when the obstacle is within the ultrasonic detection range, and 0 otherwise. Monocular visual ranging value Based on the pinhole model and prior physical width calculate: in, The focal length of the camera; Precise radial distance acquired by the ultrasonic sensor As the ranging value of the ultrasonic sensor; in summary, the fused distance The result was obtained by weighted averaging based on visual ranging confidence and ultrasonic confidence: in, It is a small constant added to prevent division by zero.

5. A biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 4, characterized in that, The orientation information is calculated as follows: First, the coordinates of the center position of the obstacle's pixel range are calculated using the minimum bounding rectangle method. And the distance from the minimum bounding rectangle of the obstacle to the perimeter of the two-dimensional intensity image. Then the horizontal azimuth angle and vertical azimuth Calculated using the following formula: in, A positive value indicates that the obstacle is located slightly to the left in front of the robot, while a negative value indicates that it is slightly to the right in front of the robot. A positive value indicates that the obstacle is located slightly above the front of the robot, while a negative value indicates that it is located slightly below the front.

6. The biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 1, characterized in that, The dynamic threat assessment method includes three sub-steps: constructing an improved virtual potential field model, prospective pose prediction, and security cost calculation. Sub-step 1: Construct an improved virtual potential field model First, define the improved repulsive potential function. : in, The fusion distance from the robot to the obstacle. The radius of influence of the potential field. This is the repulsive gain coefficient, used to adjust the overall strength of the potential field; For distance adjustment index; when When, it degenerates into a traditional quadratic potential field; when At that time, potential energy It grows faster when it decreases; Sub-step 2, Forward pose prediction: Define a set of discrete obstacle avoidance actions Where F, L, R, U, and D represent five candidate actions: going straight, turning left, turning right, surfacing, and diving, respectively; for each non-straight candidate action... Based on the current environmental state vector S, the changes in each component of the state vector after the action is performed are predicted through pre-established kinematic relationships, thus obtaining a set of state changes. Sub-step 3: Calculation of security costs: (1) Basic security cost calculation: Based on the predicted set of state changes, the execution action is calculated. The predicted distance is then used to calculate the predicted repulsive potential energy after performing the action, which serves as the basic security cost. ; (2) Geometric safety cost addition: The robot is simplified into a rectangular bounding box of known size. The distance from each corner point of the bounding box to the obstacle after the action is performed is calculated, and the minimum distance is taken. To calculate the repulsive potential energy as an additional geometric safety cost; Therefore, the full cost of security Defined as: in, It is a penalty coefficient greater than 1.

7. A biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 6, characterized in that, The specific operation for multi-objective comprehensive cost calculation is as follows: After obtaining the safety cost of each candidate action... Then, energy consumption costs were further introduced. and smoothing cost A multi-objective comprehensive cost function is constructed to achieve a comprehensive evaluation of candidate actions; Among them, the energy consumption cost The definition is as follows: Among them, for smoothing cost Defined as the current candidate action Actions performed in the previous cycle Functions of the degree of difference: This indicates switching between different types of actions. It indicates the switching between opposite actions of the same type; Finally, the multi-objective integrated cost function is constructed: in, These are the weight coefficients for the costs of each dimension, satisfying... ; This represents the value after normalizing the original cost.

8. The biomimetic robot obstacle avoidance control method for monitoring deep-sea cage aquaculture according to claim 1, characterized in that, The hybrid intelligent decision-maker first performs a rapid threat assessment. If a threat is triggered, it enters a detailed obstacle avoidance process, which includes action pre-screening and comprehensive cost calculation. Sub-step 1: Rapid Threat Assessment and Decision Triggering: After obtaining the environmental state vector S, a rapid assessment is first performed, and the fused distance is... With the preset safety threshold Compare; like If no immediate collision threat is detected, the decision-maker will directly output a "go straight" action, generate and execute underlying control commands, and end the current decision cycle. like If so, it is determined that the robot has entered the threat range of the obstacle, triggering a detailed obstacle avoidance decision-making process; Sub-step 2: Action pre-screening: Before entering the comprehensive cost calculation, based on the current environment state vector S, an action pre-screening mechanism is performed according to the following rules to exclude actions that are obviously unreasonable or physically infeasible in the current state: If the current height above the water surface is less than the minimum safe height, then an ascent is ruled out; if the current height above the water surface is greater than the maximum safe height, then a descent is ruled out. After pre-screening, the remaining actions constitute a subset of candidate actions for the current period. Subsequent steps only calculate the comprehensive cost for the actions in the subset of candidate actions. Sub-step 3: Based on the multi-objective comprehensive cost calculation method, calculate the comprehensive cost of each candidate action subset, and select the action with the smallest comprehensive cost as the output action of the current decision cycle; If multiple actions have the same cost, the action that is the same as the action in the previous cycle should be selected first to maintain decision continuity; if all actions are different from the previous cycle, the action should be selected according to the priority of left turn, right turn, surfacing, and diving.

9. A biomimetic robot obstacle avoidance control system for monitoring deep-sea cage aquaculture, characterized in that, Includes the following modules: Data acquisition module: synchronously collects information from multiple sources of sensors and performs preprocessing to obtain preprocessed multi-source sensing data; Environment state vector construction module: Based on preprocessed multi-source sensing data, construct standardized environment state vectors; Dynamic Threat Assessment Module: This module assesses the threat level of an action based on the security cost of the action using a dynamic threat assessment method based on an improved virtual potential field. Cost calculation module: Cost calculation is performed based on a multi-objective comprehensive cost function constructed from security cost, energy consumption cost, and smoothing cost; Decision output module: Outputs the optimal action decision through a hybrid intelligent decision-maker based on a multi-objective integrated cost function; Command generation and execution module: Maps and converts the abstract action commands output by the decision into specific low-level control signals, including thruster thrust commands and joint servo motor angle commands; drives the robot to perform corresponding obstacle avoidance maneuvers through the low-level control signals, and completes closed-loop obstacle avoidance control.

10. An electronic device, characterized in that, Including processor and memory; The memory is used to store computer programs; When the processor executes the program stored in the memory, it implements the bionic robot obstacle avoidance control method according to any one of claims 1-8.

Images