Bionic robotic fish swimming control method based on data driving

Abstract

Fish in nature shows excellent swimming ability, however, biomimetic robotic fish developed and inspired by aquatic organisms is still different from real fish in motion performance. In order to improve the swimming efficiency and the control precision of the bionic robotic fish, the invention provides a swimming control method based on data driving. The method comprises the following steps: firstly, synchronously acquiring swimming videos and environmental parameters of real fishes through a multi-angle high-definition camera system and an underwater sensor, and constructing a time-space aligned multi-modal data set; then, based on an improved DeepLabCut algorithm and a graph convolutional network, key points of a fish body are extracted, and a joint angle sequence is reconstructed; a driving mechanism of the bionic robotic fish is established by combining a three-joint bionic fish motion model and a propulsion curve equation. The linear propulsion effects under different fish joint angle input are compared and analyzed through the simulation platform, the propulsion efficiency and stability are evaluated, and the key influence of joint parameters on the overall motion performance is revealed. Experimental results show that the method can effectively reproduce movement styles of various fishes, and dynamic optimization of a control strategy is realized through self-supervised learning. The control method gets rid of a traditional modeling mode based on a fish body wave function, provides a new normal form for driving the control system by using real visual data, and has good real-time performance, robustness and expansibility.

Description

Technical Field

[0001] The present invention belongs to the technical field of bionic fish, and in particular relates to a data-driven bionic robotic fish swimming control method. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] In recent years, bionic robotic fish, a key research area in underwater robotics, have garnered widespread attention due to their superior maneuverability, high energy efficiency, and environmental adaptability. Compared to traditional propeller-driven underwater vehicles, bionic robotic fish, by simulating the propulsion mechanisms and swimming patterns of real fish, exhibit more flexible, energy-efficient, and biomimetic motion characteristics. They are widely used in scenarios such as environmental monitoring, underwater exploration, and aquatic animal observation.

[0004] Traditional robotic fish control strategies are often based on simplified physical models and empirical parameter adjustments, making it difficult to fully capture the highly nonlinear dynamics of fish swimming. Furthermore, these model-driven approaches often exhibit weaknesses such as insufficient robustness and limited adaptability to environmental changes or differences in robot structure.

[0005] To overcome these limitations, data-driven control methods have become a research hotspot. These methods leverage large-scale swimming data collected from real fish experiments or robotic motion tests to uncover the patterns and inherent dynamics of fish motion. By leveraging machine learning, system identification, and deep learning techniques, they construct precise motion models and design adaptive control strategies, enabling robotic fish to learn and optimize swimming patterns, thereby improving swimming efficiency, stability, and biomimetic fidelity.

[0006] With the deep integration of bionics and artificial intelligence technology, data-driven swimming control technology is gradually becoming the core driving force of the new generation of intelligent underwater robots, laying a solid foundation for achieving higher performance and more adaptable underwater bionic platforms.

[0007] The present invention starts from the reasons for the robotic fish swimming control scheme:

[0008] To address these challenges, this study employed a data-driven strategy, extracting key metrics such as swimming trajectory and tailbeat frequency from real fish swimming videos. This approach enables the construction of a more accurate dynamic model of the robotic fish, providing a new data-driven strategy for biomimetic robot design. Based on video-captured data, this study obtained the positional information of six key points on the robotic fish from the swimming videos. Using MATLAB, the angles of three joints were extracted and a data-driven dynamic model was constructed. Unlike traditional theoretical derivation methods, this approach is closer to the actual swimming process. Using the MATLAB platform, a linear swimming simulation system for the robotic fish was established and visually demonstrated. During the simulation, the dynamic behavior of each joint was simulated in real time, propulsion efficiency was continuously evaluated, and the impact of angle changes on propulsion performance was intuitively reflected. By comparing the linear motion effects generated by different angle input data, the study analyzed the differences in propulsion efficiency and motion stability among different fish species and verified the critical influence of joint angle changes on overall motion.

[0009] In order to achieve the above object, the present invention adopts the following technical solutions:

[0010] A data-driven bionic robotic fish swimming control method comprises the following steps:

[0011] In a standard water tank or ocean test environment, multi-angle high-definition stereo cameras and multiple types of underwater sensors such as piezoelectric, temperature, and pressure are configured to synchronously collect real fish swimming videos and water environment parameters to form a multimodal dataset aligned in time and space.

[0012] Based on the improved DeepLabCut, cross-domain transfer learning is introduced, and the labeled data of synthetic bionic fish and different fish species are used to design a key point detection network based on the temporal attention mechanism and graph convolutional network (GCN).

[0013] The key point detection network structure is as follows:

[0014] Feature extraction layer: Use lightweight convolutional neural networks such as ResNet-50 to extract the spatial feature vector of each frame image.

[0015] Graph convolution module: Builds an undirected graph structure based on the biological structure of the fish body, where each key point is regarded as a graph node. The connection relationship between nodes is established according to the physiological structure of the fish joints, and the spatial dependency between key points is captured through graph convolution operations.

[0016] Temporal Attention Module: This module introduces an attention mechanism to process dynamic changes between frames, constructs an attention matrix in the temporal dimension, learns key frame weights from consecutive frames, and adaptively improves the ability to recognize highly dynamic areas.

[0017] Output layer: Outputs the coordinate positions of all key points in each frame of the image for subsequent motion modeling and control system input.

[0018] During operation, the bionic robotic fish maintains a real-time connection with a host computer via a built-in communication module. The host computer is responsible for sending control commands, receiving sensor data, and updating control strategies in real time. Furthermore, to meet the needs of multi-robot collaborative operations (such as formation cruising, target search, and underwater monitoring), the present invention has designed a multi-robot formation control module that supports autonomous collaborative control among multiple bionic robotic fish.

[0019] As an optional implementation, Matlab is used for simulation to build a three-joint dynamic model of the bionic robotic fish, including angular velocity and propulsion efficiency. The specific process is as follows:

[0020] Step 101: Introduce real fish key point motion data as input boundary conditions to simulate the swimming process of the fish.

[0021] Step 102: Use Stateflow and the Control System Toolbox in MATLAB to design a PID controller and an adaptive reinforcement learning controller to control the phase and amplitude of the three joints.

[0022] Step 103: You can choose to load a neural network model for control strategy verification, which is compatible with the policy network exported by Deep Learning Toolbox.

[0023] Step 104: Data visualization is performed through MATLAB scripts to analyze performance indicators such as propulsion speed and acceleration; a GIF animation can be output to show the bionic fish simulation process.

[0024] Step 105: Through the parameter sweep function, systematically verify the impact of different elastic coefficients, angular frequencies, control strategies and other factors on propulsion performance; support the export of simulation data for comparison with actual hardware.

[0025] Step 106: In an environment without hardware dependency, quickly evaluate the effect of the control algorithm to provide a reliable priori evaluation basis for controller deployment and actual underwater experiments.

[0026] As an optional implementation, two high frame rate (≥150 fps) monocular cameras, one pressure sensor, and one flow rate sensor are equipped.

[0027] As an optional embodiment, the controller is based on the PPO algorithm, and the action is a joint angle instruction.

[0028] As an optional implementation, key point extraction only monitors six key points such as the head, torso, and tail, using an improved DeepLabCut network. The specific process is:

[0029] Step 201: Use two synchronized cameras to collect videos, each video is 10 seconds long and has a frame rate of 24 frames per second.

[0030] Step 202: Use the k-means method to extract 100 images from each video, for a total of 300 images as our dataset.

[0031] Step 203: Select ResNet_50 as the baseline network. Here, our evaluation metric is RMSE (Root Mean Square Error), which is calculated by taking the square root of the average of the squares of the differences between the predicted values ​​and the actual values. The calculation formula is:

[0032] ;

[0033] Step 204: Multi-camera calibration To achieve accurate 3D estimation, the first necessary step is camera calibration. The conversion equation from the world coordinate system to the pixel coordinate system is:

[0034] ;

[0035] ;

[0036] Step 205: Camera distortion is generally divided into radial distortion and tangential distortion. The final correction point combining radial distortion and tangential distortion is:

[0037] ;

[0038] Step 206: During the camera calibration phase, a bundle adjustment-based method is used to optimize the camera calibration process.

[0039] Step 207: After calibration, perform 3D pose estimation using the inverse mapping from the world coordinate system to the pixel coordinate system:

[0040] ;

[0041] As an optional implementation, the controller adopts a fusion architecture of MPC and PPO, which automatically switches when the underwater environment changes.

[0042] As an optional implementation, the bionic robotic fish can maintain a real-time connection with a host computer during operation via a built-in communication module. This communication module, which can be a wireless communication module, an underwater acoustic communication module, or other communication device suitable for underwater environments, is used to enable data exchange with the host computer. The host computer is responsible for sending motion control commands, receiving data collected by the bionic robotic fish's sensors, and updating control strategies and parameters in real time based on task requirements.

[0043] As an optional implementation, to meet multi-robot collaborative operation scenarios such as formation cruising, target search, and underwater monitoring, the present invention has designed a multi-robot formation control module. This module supports autonomous collaborative control between multiple bionic robotic fish through a distributed communication protocol. Each robotic fish can dynamically adjust its own motion trajectory based on the status information received from neighboring robotic fish, achieving formation maintenance, task coordination, and obstacle avoidance, thereby improving the overall task execution efficiency and robustness of the system. The formation strategy can adopt the following control mechanism combination:

[0044] Master-slave control structure: A master bionic fish acts as the leader, issuing path planning or mission instructions, and other fish act as slave units to execute the instructions;

[0045] Centralized control strategy: The host computer coordinates the position, speed, and behavior of each bionic fish based on global task planning;

[0046] Distributed collaborative algorithm: Each bionic robotic fish independently calculates and controls output based on local perception information and neighbor status, achieving decentralized formation behavior.

[0047] As an alternative implementation, the present invention extracts key motion parameters, including tail beat frequency and posture rhythm, based on the swimming styles of real fish exhibited in various videos. During the modeling process, dedicated control templates are created for different fish species (such as tuna, grass carp, and carp). These templates are mapped to the three-joint drive model of the robotic fish using a neural network or data fitting function, resulting in the bionic robotic fish exhibiting dynamic characteristics consistent with the original fish.

[0048] Furthermore, considering the physical differences between the bionic robotic fish structure and that of biological organisms in terms of processing materials, servo response capabilities, and connecting rod layout, the present invention introduces a "structural mapping correction module" to automatically adjust the scale, frequency, and dynamic range of the input data during the control strategy generation process to ensure that the data of real fish can be effectively absorbed and executed by bionic robotic fish with different structures.

[0049] In addition, the control system adopts a self-supervised learning mechanism. The system monitors the swimming trajectory and posture data of the robot fish in real time through sensors, and matches it with the input video reference action. It automatically adjusts the control weights and motion parameters according to the error, so that the robot fish can achieve behavioral adaptation and evolutionary control strategy updates during operation.

[0050] The beneficial effects of the present invention are:

[0051] The present invention proposes a data-driven bionic robotic fish swimming control method, constructs a highly integrated system architecture from perception, modeling to control, and significantly improves the motion consistency, robustness and intelligence level of the bionic robotic fish in simulation platforms and real waters.

[0052] This method integrates cutting-edge technologies such as multimodal perception, deep reinforcement learning, and embedded control. By introducing a cross-domain transfer learning mechanism, it integrates video data from different fish species with synthetic bionic fish data to establish a unified key point annotation system. Combined with an improved graph neural network model and a temporal attention mechanism, it achieves highly robust detection of key moving parts of fish, accurately identifying fish structure and dynamic features even in the presence of rapid posture changes or occlusion.

[0053] After identifying key points, the system integrates high-dimensional temporal information, such as velocity and angular acceleration, extracted from the visual trajectory and environmental sensors. Key motion parameters of different fish species in the video are extracted, and control templates are constructed for each. These templates are then mapped to the three-joint drive model of the bionic robotic fish using a neural network or data fitting method, enabling dynamic motion reconstruction of different fish styles.

[0054] In view of the physical differences between the servo response of the bionic robotic fish and that of real fish, the scale, frequency and dynamic range of the input signal are automatically adjusted. The control system monitors the posture trajectory of the robotic fish in real time through sensors, dynamically matches it with the reference behavior of the target fish, and adjusts the control weights and drive parameters online based on error feedback to achieve adaptive evolution of the control strategy and behavior optimization.

[0055] The MATLAB simulation platform supports a high-fidelity dynamic verification process, which can quickly evaluate the propulsion efficiency, attitude stability, and energy consumption levels under different parameter configurations, significantly shortening the algorithm iteration cycle and actual deployment time, and improving the development efficiency and application reliability of the control system. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] The accompanying drawings, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0057] Figure 1 A block diagram of a robotic fish design for at least one embodiment of the present invention;

[0058] Figure 2 A flowchart of the DEEPLABCUT algorithm according to at least one embodiment of the present invention;

[0059] Figure 3 A schematic diagram of a simplified physical model of a three-joint robotic fish according to at least one embodiment of the present invention;

[0060] Figure 4 A schematic flow chart of a control method according to at least one embodiment of the present invention DETAILED DESCRIPTION

[0061] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0062] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0063] It should be noted that the terms used herein are only for describing specific embodiments and are not intended to limit the exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is intended to include the plural form. In addition, it should be understood that when the terms "comprise" and / or "include" are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.

[0064] A data-driven bionic robotic fish swimming control method comprises the following steps:

[0065] S1: In the predetermined waters, a stereo video camera and underwater sensors are used to synchronously collect the trajectory of key points of the fish and environmental parameters, and align the data in time and space.

[0066] S2: Construct a multi-joint coupling dynamics model based on the MATLAB platform; use the system identification method combined with the least squares loss function to perform online correction of the model's inertia, elasticity, and fluid correction parameters.

[0067] S3: The embedded control unit receives joint status and environmental sensor information, and outputs PWM drive instructions according to the optimization strategy; combined with the fast path planning algorithm, it realizes real-time obstacle detection and dynamic obstacle avoidance.

[0068] S4: Based on the swimming style characteristics of real fish shown in different videos, their key motion parameters are extracted and exclusive control templates are constructed for different fish species. Through neural networks or data fitting functions, the template parameters are mapped to the three-joint drive model of the bionic robotic fish to achieve highly realistic dynamic motion behavior reproduction.

[0069] S5: A structural mapping correction module is introduced to adaptively adjust the scale, frequency, and dynamic range of the input data during the control strategy generation process to compensate for the physical differences between the bionic robotic fish and the real fish in terms of servo response capability and linkage layout, ensuring the effective execution of control instructions.

[0070] S6: Using a self-supervised learning mechanism, the control system dynamically compares the robot fish's posture and motion trajectory data collected in real time by sensors with the target action in the video, calculates the error, and automatically adjusts the control weights and motion parameters to achieve online optimization of the control strategy and adaptive evolution of the behavior.

[0071] Specifically, the multimodal data acquisition described in step S1 involves deploying two or more synchronously triggered stereo cameras (frame rate ≥ 200 fps) in the test pool or ocean. Six key points (head, torso, and tail) are annotated on the fish's surface with micro-markers. The clock synchronization module aligns the timestamps of the data collected by each device, and uses exterior orientation element calibration to spatially align the multi-view data, generating a unified spatiotemporally aligned dataset.

[0072] Specifically, the MATLAB platform model construction and online identification implementation process described in step S2 is:

[0073] Step 101: Create a new model in the MATLAB environment, add a three-degree-of-freedom rotational joint subsystem, and set the initial moment of inertia, damping coefficient, and elastic stiffness parameters.

[0074] Step 102: Import the pre-processed key point angle sequence and environmental flow rate data.

[0075] Step 103: Based on the least squares identification algorithm, call the ident function to calculate the estimated values ​​of the model parameters and automatically update the inertia every 0.5 seconds during the simulation process.

[0076] Step 104: Load the key point and velocity datasets of three fish (Tuna, Carp, and GrassCarp) and perform multiple independent simulation experiments:

[0077] Step 105: For each fish species, other environmental conditions are fixed, only the joint drive input is adjusted, and performance indicators such as propulsion speed, joint power, and posture stability are collected;

[0078] Step 106: Compare the differences in model parameters and movement performance of different fish species; export the results into a multidimensional table for cross-species parameter sensitivity analysis;

[0079] Step 107: Conduct comparative experiments using multimodal data (video, sensor, and simulation):

[0080] Specifically, the embedded control and real-time obstacle avoidance described in step S3 exports the generated MATLAB data into automatic floating-point C code and burns it into the STM32 series microcontroller. The motor is driven by PWM to make the joint move accurately according to the calculated target angle and speed.

[0081] Specifically, the fish species-specific control template construction and mapping implementation process described in step S4 is as follows:

[0082] Step 201: Obtain motion data of real fish in different videos and use a key point detection algorithm (DeepLabCut) to extract fish posture change information, including core parameters such as tail beat frequency, body bending rhythm, and maximum bending amplitude.

[0083] Step 202: Based on the extracted data, characteristic parameter sets are established for different fish species to form parameter templates representing their swimming styles.

[0084] Step 203: Based on the neural network model or data fitting function, the above template parameters are mapped into the input angle sequence in the three-joint drive model to ensure that the motion pattern of the bionic robotic fish is highly consistent with the original fish in terms of time sequence and spatial amplitude.

[0085] Step 204: Integrate the mapping relationship into the robotic fish control system to drive the bionic robotic fish to perform corresponding actions, thereby achieving highly simulated dynamic behavior reproduction of different fish species.

[0086] Specifically, the parameters of the structural mapping correction module described in step S5 are adjusted to analyze the structural parameters of the bionic robotic fish and establish a differential mapping model between it and a real fish. During the control strategy generation process, the structural mapping correction module is introduced to standardize the input angle, frequency, and motion amplitude, and construct a dynamic adjustment function to adapt it to the current robotic fish hardware constraints.

[0087] Specifically, the error feedback and parameter update implementation process of the self-supervised learning mechanism described in step S6 is as follows:

[0088] Step 301: The control system collects the motion trajectory, posture angle and joint response data of the bionic robotic fish in real time, and obtains the complete state vector through built-in sensors and visual feedback devices.

[0089] Step 302: Dynamically compare the collected state vector with the target fish behavior data extracted from the reference video.

[0090] Step 303: Input the error signal into the self-supervised learning module, and use gradient descent or reinforcement learning strategies to update the control weight parameters and key hyperparameters in the joint drive model online.

[0091] Step 304: As the running time increases, the control system gradually adapts to the specific environment and task conditions, achieving improved simulation accuracy of different fish behaviors and behavioral evolution of the control strategy.

Claims

1. A data-driven bionic robotic fish swimming control method, characterized by: The following steps are involved: In a standard tank environment, a multi-angle high-definition stereo camera and underwater sensors, including temperature and pressure sensors, are deployed to synchronously capture real fish swimming videos to form a spatiotemporally aligned multimodal dataset. Based on the improved DeepLabCut framework, cross-domain transfer learning is introduced. Using labeled data of bionic fish and various fish species, a key point detection network based on the temporal attention mechanism and graph convolutional network (GCN) is constructed to extract the temporal position information of key points on the fish body. Using MATLAB to extract the three joint angles, a data-driven dynamics model was constructed. This method, unlike traditional theoretical derivation methods, is closer to the actual swimming process. Using the platform, a linear swimming simulation system for a robotic fish was established and visually demonstrated. During the simulation, the dynamic behavior of each joint was simulated in real time, the propulsion efficiency was continuously evaluated, and the impact of angle changes on propulsion performance was intuitively reflected. By comparing the linear motion effects produced by input data at different angles, the study analyzed the differences in propulsion efficiency and motion stability among different fish species, verifying the key impact of joint angle changes on overall motion. Construct a MATLAB-based simulation platform to verify the real-time, robustness, and fault recovery performance of the control method on a hardware platform; It supports data interaction and control command transmission with the host computer, and has the ability to form a multi-machine formation. It can realize collaborative path planning and motion synchronization control among multiple bionic robot fish through the network communication module. The control system constructs a fish-specific control template based on the movement style parameters of different fish individuals in the video, and maps the template to the bionic robotic fish controller to achieve accurate reproduction of the movement styles of multiple fish species; The control system uses a self-supervised learning mechanism to autonomously optimize control parameters by comparing the error between the target fish's behavior and the actual output of the bionic robotic fish in real time, achieving online adaptive adjustment of the motion control strategy and continuous performance evolution. A bidirectional feedback closed-loop system is constructed between the video information modality and the control output modality to achieve end-to-end closed-loop control driven by visual information and corrected by feedback behavior, thereby improving the simulation accuracy and environmental adaptability of the control system.

2. The data-driven bionic robotic fish swimming control method according to claim 1, wherein: The DeepLabCut algorithm is used to track the posture of the robotic fish, aiming to accurately capture the movement trajectory and posture changes of the robotic fish in the water. The algorithm allows for 2D and 3D markerless posture tracking. This method does not require any physical markers to be attached to the robotic fish. Instead, a small amount of keyframe data is marked, enabling the deep learning model to infer the posture and position information of the robotic fish throughout the movement process.

3. The data-driven bionic robotic fish swimming control method according to claim 1, wherein: Next, the fish skeleton is dynamically reconstructed based on the two-dimensional spatial coordinates of the six key points. By calculating the relative positions and angles between the key points, the angle change sequences of dorsal fin-body 1, body 1-body 2, and body 2-tail are obtained, and then the three joint angles corresponding to the robotic fish are extracted.

4. The data-driven bionic robotic fish swimming control method according to claim 1, wherein: The implementation process of the kinematic model construction of the bionic robotic fish swimming control is as follows: Step 101: Based on the fish body's transverse vibration curve, establish the propulsion curve equation: ； in, is the amplitude, is the wave number, is the initial phase; Step 102: Assume that the rotation angle of the first joint is , the corresponding arc length is ； in, is the rotation radius of the first joint; Step 103: Based on the relationship between the propulsion curve in step 101 and the arc length in step 102, derive the kinematic model of the first joint: ； Step 104: Similarly, set the rotation angles of the second and third joints to , using their respective rotation radius Derive its kinematic expression; Step 105: Combine the kinematic equations of the three joints to obtain the three-joint rotation model equations of the bionic robotic fish: ； On the basis of completing the construction of the three-joint kinematic model of the bionic robotic fish, in order to achieve its precise and controllable propulsion behavior, a control dynamics modeling process is further introduced. The control dynamics model not only considers the displacement output caused by the change of joint angle, but also integrates factors such as the acceleration response that is closely related to the propulsion process, and establishes a mapping relationship between the input control signal and the fish body's behavioral response.

5. The data-driven bionic robotic fish swimming control method according to claim 1, wherein: A fish swimming simulation platform was built in the MATLAB environment. By comparing the lateral oscillation trajectory displacement and speed differences of the three fish species, the significant differences in their propulsion strategies were revealed.

6. The data-driven bionic robotic fish swimming control method according to claim 1, wherein: The specific process of building a comprehensive evaluation system to assess the athletic performance of fish is as follows: Step 201: Smoothness is defined as the integral of the second derivative of displacement, representing the acceleration change The degree of quantification is expressed as: ； Step 202: Speed ​​fluctuation reflects the change in speed of the fish during its movement. The greater the fluctuation, the more unstable the fish's movement is. The formula is: ； Step 203: The standard deviation of acceleration is used to evaluate the energy consumption and mechanical properties of fish movement. Stability, acceleration change is expressed as: ; Step 204: The composite motion evaluation index can be expressed as: ； According to the range of speed fluctuation, the weight coefficient is set as: ； The system starts from three key dimensions: smoothness, velocity volatility, and acceleration standard deviation, and comprehensively examines the stability, coordination, and energy consumption level of the fish during movement. It introduces a composite evaluation function, normalizes and weights the multi-dimensional indicators to form a single evaluation quantity, which can quickly determine the quality of the robot fish's movement performance under different control strategies, fish species, and environmental disturbance conditions.

7. The data-driven bionic robotic fish swimming control method according to claim 1, wherein: The multi-machine formation control establishes a unified communication protocol and task scheduling mechanism, allowing each bionic robotic fish to autonomously adjust its motion trajectory, speed and formation structure according to real-time perception information and the status of neighboring nodes, thereby achieving collaborative behaviors such as dynamic obstacle avoidance, target tracking and task division.

8. The data-driven bionic robotic fish swimming control method according to claim 1, characterized in that: This control method constructs a species-specific control parameter template based on the posture feature data of different fish individuals extracted through video analysis, and maps the template to the multi-joint control system of the bionic robotic fish to achieve highly simulated swimming behavior reproduction of different fish styles.

9. The method according to claim 1, wherein: The control system adopts a self-supervised learning mechanism. By comparing the actual motion data of the bionic robotic fish with the data of the target real fish, it automatically updates the control parameters, realizes the dynamic evolution and continuous optimization of the control strategy, and constructs a two-way feedback closed-loop system between the video modality and the control instruction modality. By comparing the visual input and the execution action, the control strategy is adjusted in real time to realize cross-modal linkage control.

Images