Bionic fish robot and motion control method thereof

Abstract

The application discloses a kind of bionic fish robots and its motion control method, it is related to underwater robot technical field, including head unit, torso unit, tail unit, battery and control unit, head unit includes fish head shell, pectoral fin and dorsal fin, pectoral fin has two and oppositely set in fish head shell left and right sides, and the left and right sides of dorsal fin respectively have inflatable and deflatable air bag, tail unit includes fish tail seat and tail fin.Torso unit includes flexible torso shell, bias adjustment mechanism and swing driving mechanism, and the front and rear ends of flexible torso shell are respectively sealedly connected with fish head shell and fish tail seat, bias adjustment mechanism is arranged inside flexible torso shell by mounting plate, swing driving mechanism is arranged at the execution end of mounting plate and bias adjustment mechanism, and the execution end of swing driving mechanism is connected with fish tail seat.The application significantly improves the high mobility and stability of bionic fish robot in complex underwater environment, and can meet the needs of various underwater operation scenarios.

Description

Technical Field

[0001] This invention relates to the field of underwater biomimetic robot technology, specifically to a biomimetic fish robot and its motion control method. Background Technology

[0002] In recent years, with the increasing demand for underwater environmental exploration, structural inspection, ecological monitoring, and marine ranching management, underwater robotics technology has gradually become an important research direction in the field of marine engineering. However, most mainstream underwater robots are still propeller-driven, and these systems are increasingly revealing limitations in complex environments, such as high energy consumption, high noise, strong disturbance, high risk of entanglement, and insufficient maneuverability. In scenarios such as underwater acoustic monitoring, underwater archaeology, and biological behavior observation, which are sensitive to noise, the eddy noise generated by the propeller can easily interfere with the task itself. In environments with dense structures or narrow spaces, such as wind turbine foundations, aquaculture cages, and areas around subsea pipelines, the rotating parts of the propeller are not only prone to entanglement but also pose a high safety risk. Therefore, underwater robots need to develop towards lower noise, higher maneuverability, less disturbance, and higher safety, and traditional propulsion methods can no longer meet this trend.

[0003] Bionic fish robots, mimicking the propulsion mechanisms of real fish, exhibit significant advantages in low noise, high energy efficiency, and strong maneuverability, thus becoming an important direction in underwater robotics research in recent years. Fish in nature use body wave propulsion and multi-fin coordination to flexibly turn, precisely adjust their posture, and hover or slowly approach targets in confined environments; their high degree of freedom of movement is of great value for engineering applications. However, most existing bionic fish robots use a single tail fin as their primary propulsion method. While this allows for relatively efficient linear swimming, it is still insufficient in terms of posture adjustment, vertical control, hovering ability, and small-radius turning. In particular, relying solely on tail fin oscillation makes it difficult to generate stable lift, limiting the vertical movement or depth control of bionic fish robots. Using differential tail fin oscillation for turning often results in a large turning radius and slow motion response, making it unsuitable for complex engineering environments. Furthermore, many bionic fish robots use multiple servos connected in series to drive multiple body segments to generate body waves; this structure is complex and difficult to synchronize, affecting long-term engineering reliability. In contrast, natural fish employ a coordinated mechanism of pectoral fins, dorsal fins, and tail fins in complex movements. For example, pectoral fins can generate lift or lateral force through rotation in different directions, which can be used for hovering, deceleration, braking, and fine attitude control; some fish selectively open or change the shape of their dorsal fins when turning, thereby increasing local drag or creating additional hydrodynamics, making turning movements more flexible and effective. These characteristics have not yet been fully utilized in existing biomimetic robots, limiting the practical performance of biomimetic fish robots in engineering environments.

[0004] In terms of control methods, the Central Pattern Generator (CPG), as a control structure that simulates the rhythmic behavior of a biological nervous system, can generate stable, continuous, and parameter-adjustable output signals, making it very suitable for driving biomimetic fish robots and enabling them to produce more standard fish-like body waves. CPGs not only have advantages such as strong anti-disturbance capabilities and smooth, natural control, but can also form various stable motion patterns by adjusting phase, amplitude, and frequency, allowing biomimetic fish robots to smoothly switch between actions such as straight swimming, sharp turns, and slow turns. However, traditional CPG control lacks environmental adaptability, typically relying on fixed parameters or manually designed patterns, making it difficult to automatically select the optimal motion state based on real-time sensor information, which is detrimental to achieving autonomous behavior in complex underwater scenarios. With the development of deep learning and reinforcement learning technologies, it has become possible to provide biomimetic fish robots with intelligent and highly autonomous decision-making mechanisms. Reinforcement learning, based on a "trial and error-feedback" approach, finds strategies and can actively select the optimal action mode based on environmental input, thereby enabling biomimetic fish robots to possess intelligent behaviors such as path tracking and dynamic obstacle avoidance. Similar to CPG, reinforcement learning is also inspired by biology, but its advantage lies in its self-learning and self-optimization capabilities, continuously improving decision-making quality in unknown environments. In practical engineering, relying solely on reinforcement learning to directly output continuous control signals is not only difficult to train but may also lead to unstable outputs, which is detrimental to the precise body wave control of biomimetic fish robots. Therefore, using reinforcement learning as a high-level decision-maker responsible for "selecting motion patterns," while CPG is responsible for "generating specific motion rhythms," presents a clear and highly feasible hierarchical architecture. By using reinforcement learning to output fine-tuning parameters and adjusting the final output motion pattern, the biomimetic fish robot can have more motion options, resulting in superior performance in path tracking and target tracking. Summary of the Invention

[0005] To address the shortcomings of the existing technologies, this invention aims to propose a biomimetic fish robot to solve the problems of weak posture adjustment ability, difficulty in vertical control, insufficient hovering stability, and limited small-radius turning ability of existing biomimetic fish robots, thereby improving the robot's motion flexibility and posture control ability in complex underwater environments.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: A biomimetic fish robot includes a head unit, a torso unit, a tail unit, a battery, and a control unit. The head unit includes a fish head shell, pectoral fins, and a dorsal fin. The fish head shell is a rigid shell with an axisymmetric structure and a closed rear end. The torso unit and the tail unit are arranged sequentially behind the fish head shell.

[0007] There are two pectoral fins, which are located opposite each other on the left and right sides of the fish head shell. Two pectoral fin drive components are symmetrically arranged inside the fish head shell. The actuator of each pectoral fin drive component is connected to the pectoral fin on the same side and drives the pectoral fin to rotate to adjust its angle.

[0008] The dorsal fin is located on the top of the fish head shell, with an air sac on each of its left and right sides. Inside the fish head shell are two inflation / deflation components, which independently inflate or deflate the two air sacs.

[0009] The tail unit includes a fish tail seat and a tail fin. The fish tail seat is a cone-shaped sealed shell with an axisymmetric structure, and the tail fin is vertically fixed at the rear end of the fish tail seat.

[0010] The torso unit includes a flexible torso shell, an offset adjustment mechanism, and a swing drive mechanism. The flexible torso shell is an axisymmetric cylindrical structure, with its front and rear ends sealed and connected to the fish head shell and the fish tail seat, respectively. The offset adjustment mechanism is located inside the flexible torso shell via a mounting plate, which is fixedly located at the rear end of the fish head shell.

[0011] The swing drive mechanism is located at the execution end of the mounting plate and the bias adjustment mechanism. The execution end of the swing drive mechanism is connected to the fish tail seat. In the working state, the bias adjustment mechanism causes the flexible body shell and the tail unit to be biased to one side of the central axis of the fish head shell through the swing drive mechanism. At the same time, the swing drive mechanism drives the tail unit to swing regularly.

[0012] The battery and control unit are both installed inside the fish head shell. The control unit includes a main control board, IMU, depth sensor, Doppler velocimeter, sonar, and vision sensor.

[0013] Furthermore, the front end of the fish head shell has a hemispherical flow guide, the inner side of which is connected to the inside of the fish head shell, and a square opening is provided at the top of the fish head shell.

[0014] The square opening of the fish head shell is equipped with a waterproof cover. The waterproof cover is detachably fixed and sealed to the outer edge of the square opening of the fish head shell. The top surface of the waterproof cover and the surface of the fish head shell are streamlined curved surfaces.

[0015] Furthermore, the dorsal fin is vertically arranged along the central axis of the fish head shell, and the bottom of the dorsal fin is fixedly connected to the upper surface of the waterproof cover plate as one piece.

[0016] The air bladder is an elastic diaphragm made of rubber, the shape of which is consistent with the shape of the dorsal fin sidewall. The annular edge of the air bladder is fixedly and sealed to the dorsal fin sidewall, so that a closed space is formed between the inner wall of the air bladder and the dorsal fin sidewall.

[0017] The dorsal fin has a mounting hole on each of its left and right sides. The inflation / deflation assembly includes an electric air pump and an air guide tube. The electric air pump is fixed inside the fish head shell by a bracket.

[0018] One end of the air duct is connected to the outlet of the electric air pump via a high-speed solenoid valve, and the other end is fixedly inserted into the mounting hole of the dorsal fin with its outer side wall sealed to the mounting hole of the dorsal fin. The end face of the air duct is flush with the side wall of the dorsal fin. The signal terminals of the electric air pump and the high-speed solenoid valve are respectively connected to the main control board for communication.

[0019] Furthermore, the pectoral fin is a flat plate that is approximately a right-angled triangle, with one right-angled side of the pectoral fin close to the fish head shell. The front side of the pectoral fin is an outwardly convex arc-shaped oblique side, and two axial holes are symmetrically opened on the left and right sides of the fish head shell.

[0020] The pectoral fin drive assembly includes a servo motor and a horizontal shaft. The horizontal shaft passes through a shaft hole on the same side and is rotated and sealed with the fish head shell. One end of the horizontal shaft is fixedly connected to the straight edge of the pectoral fin near the fish head shell on the same side.

[0021] The first servo motor is fixed inside the fish head shell by the second bracket. The output shaft of the first servo motor is coaxially and fixedly connected to the other end of the horizontal axis. The signal terminal of the first servo motor is connected to the main control board for communication.

[0022] Furthermore, the flexible torso shell includes a tubular bladder made of silicone rubber, with a spring skeleton fixedly embedded inside the tubular bladder.

[0023] The spring frame is made of elastic steel wire wound in a spiral manner. In its natural state, the overall shape of the spring frame is consistent with the shape of the inflated cylindrical bladder.

[0024] The rear end face of the fish head shell has a protrusion 1 integrated therewith, and the front end face of the fish tail base has a protrusion 2 integrated therewith.

[0025] The front end of the flexible torso shell is fixedly sleeved on the outside of boss one, and its rear end is sleeved on the outside of boss two. The outer surface of the flexible torso shell smoothly transitions with the surfaces of the fish head shell and the fish tail seat, respectively.

[0026] Furthermore, the mounting plate is arranged horizontally, and its front end is fixedly connected to the center of the rear sidewall of the fish head shell.

[0027] The bias adjustment mechanism includes a motor, a crank, and an L-shaped rocker arm. The motor is fixed above the mounting plate, and the crank and L-shaped rocker arm are both located below the mounting plate. One end of the crank is fixed to the output shaft of the motor, and the corner of the L-shaped rocker arm is hinged to the front end of the mounting plate via a vertical shaft.

[0028] The other end of the crank is movably connected to one end of the L-shaped rocker arm via a connecting rod. In operation, the motor drives the L-shaped rocker arm to rotate relative to the mounting plate via the crank and connecting rod to adjust the angle of the L-shaped rocker arm.

[0029] Furthermore, the L-shaped rocker arm is a flat plate structure and is parallel to the mounting plate. The swing drive mechanism includes a second servo motor, a drive shaft, a rotating arm, a cross rod, and a U-shaped bracket. The second servo motor is fixed to the upper surface of the mounting plate. The drive shaft is located above the L-shaped rocker arm through a bearing seat. The front end of the drive shaft is connected to the output shaft of the second servo motor through a universal coupling.

[0030] The rotating arm is located on one side of the drive shaft. One end of the rotating arm is fixedly connected to the rear end of the drive shaft. The motor drives the drive shaft to rotate through a universal coupling, and the drive shaft drives the rotating arm to rotate around the axis of the drive shaft.

[0031] A fixed plate parallel to the L-shaped rocker arm is provided above it, and the front end of the fixed plate is fixedly connected to the top of the bearing seat.

[0032] The cross member is located between the L-shaped rocker arm and the fixed plate. The other end of the rotating arm is connected to the cross member through a U-shaped bracket. The upper and lower ends of the cross member are connected to the fishtail seat through a rigid rod.

[0033] Furthermore, the other end of the rotating arm is provided with a circular insertion hole that is inclined relative to the drive shaft, and both the left and right ends of the horizontal section of the cross rod are rotatably connected to the U-shaped bracket.

[0034] The front side of the U-shaped bracket has a plug shaft. One end of the plug shaft is fixedly connected to the outer wall of the U-shaped bracket, and the other end is inserted into a circular plug hole and slidably engaged with the other end of the rotating arm.

[0035] The upper and lower ends of the vertical section of the cross member pass through the fixed plate and the L-shaped rocker arm respectively, and rotate with the fixed plate and the L-shaped rocker arm. The two rigid rods are arranged in parallel, and the front end of each rigid rod is fixedly connected to the corresponding end of the vertical section of the cross member, and the rear end is fixedly connected to the front end of the fishtail seat.

[0036] Another objective of this invention is to provide a motion control method.

[0037] A motion control method, based on the aforementioned biomimetic fish robot, includes the following steps: Step 1: First, the control unit acquires the real-time attitude angle, three-axis acceleration, depth and sonar data of the bionic fish robot. After data preprocessing and decoupling and splicing, the acquired data is combined with the current motion state to generate a unified feature vector. Step 2: The feature vector is synchronously input into the hierarchical decision system, which includes a high-level strategic reinforcement learning network and a low-level tactical reinforcement learning network. The high-level strategic reinforcement learning network outputs discretely based on the feature vector to generate macroscopic motion pattern commands, which are then sent to the low-level tactical reinforcement learning network and the subsequent CPG mapping mechanism. The low-level tactical reinforcement learning network combines the feature vector with the motion pattern commands from the high level to continuously output and generate oscillator parameters for precise fine-tuning. Step 3: The motion mode command and continuous oscillator parameters are input to the mapping mechanism of the central mode generator. After smoothing interpolation and phase continuity processing inside, they are applied to the Hopf oscillator to generate a continuous and smooth rhythm control signal. Step 4: The control signal is precisely distributed to servo motor 1, motor, servo motor 2 and high-speed solenoid valve through the decoupling mapping module, thereby realizing the driving of the bionic fish robot in complex underwater environment.

[0038] Furthermore, the high-level strategic reinforcement learning network performs macroscopic motion mode switching by mapping global perception information to a predefined discrete state machine probability distribution. make For high-level strategic reinforcement learning networks The input state space at time t is defined as: ; in, For the heading error of the biomimetic fish robot, The lateral deviation of the trajectory of the biomimetic fish robot. For the depth of the biomimetic fish robot, and These represent the distance and azimuth angle of the biomimetic fish robot relative to the obstacle; High-level strategic reinforcement learning networks use the Softmax activation function to output discrete motion patterns at the current time step. The probability distribution is used to output seven movement modes, including left turn, sharp left turn, straight, right turn, sharp right turn, ascend, and descend. The mode with the highest probability is selected as the output command. ; in, and To enhance the weights and biases of the learning network for high-level strategic reinforcement; The underlying tactical reinforcement learning network is responsible for outputting continuously fine-tuned parameters under the selected movement pattern; make For the underlying tactical reinforcement learning network in The input state space at each time step needs to integrate the instructions from the high-level strategic reinforcement learning network and the physical rhythms of the low-level tactical reinforcement learning network. The input state space is as follows: ; in, and For the linear velocity and angular velocity of the biomimetic fish robot, The attitude angle vector measured by the IMU. This represents the instantaneous phase of the current central pattern generator; Through the computation of the underlying tactical reinforcement learning network, the final output is a continuous fine-tuning parameter vector: ; It is the amplitude compensation coefficient of the central mode generator, which acts on the intrinsic amplitude parameter of the Hopf oscillator and is used to fine-tune the amplitude of the tail fin swing when subjected to backflow or downstream disturbances. The CPG frequency compensation coefficient is used to fine-tune the oscillation frequency of the tail fin in order to achieve linear and smooth adjustment of swimming speed. The tail offset angle adjustment is directly mapped to the motor, which provides a small lateral force to correct heading deviations without switching steering modes. The angle of attack and fine-tuning of the left and right pectoral fins are applied to servo motor one, which compensates for roll and pitch attitude stability by changing the rotation angle of the pectoral fins around the vertical axis. The underlying tactical reinforcement learning network uses the Tanh activation function to output a continuously fine-tuned parameter vector. for: ; in This is the range constraint matrix for the physical actuator. and The weights and biases of the underlying tactical reinforcement learning network.

[0039] By adopting the above technical solution, the beneficial technical effects of the present invention are as follows: (1) This invention generates controllable lift and lateral force in water by independently adjusting the tilt angle of the two pectoral fins, which reproduces the natural fish's floating, diving and precise motion control in complex environments, improves maneuverability and flexibility, and provides important support for the high maneuverability and stability of biomimetic fish robots in engineering applications.

[0040] (2) When maneuvering is required, the air bladder on the side of the dorsal fin can be inflated quickly to participate in the steering and attitude adjustment control of the bionic fish robot. The deformation of the dorsal fin can effectively amplify the hydrodynamic asymmetry effect, so that the fish body still has good steering response ability under low speed conditions, which helps to suppress the attitude fluctuation caused by the large swing of the tail, so as to assist in completing the steering or assist in body stability.

[0041] (3) The architecture employing a high-level strategic reinforcement learning network and a low-level tactical reinforcement learning network aims to achieve precise control of discrete motion mode switching and continuous fine-tuning parameters within the same network framework. This balances the macroscopic speed of decision-making (mode switching) and the microscopic precision of actions (parameter fine-tuning). Macroscopic speed refers to the algorithm's ability to quickly output the discrete policy network, i.e., to rapidly provide the next motion mode. Microscopic precision refers to the algorithm's ability to achieve more precise control by further outputting fine-tuning parameters. This invention solves the problem of motion oscillation and non-convergence caused by the direct output of low-level signals in traditional reinforcement learning. Attached Figure Description

[0042] Figure 1 This is a three-dimensional structural diagram of a biomimetic fish robot according to the present invention.

[0043] Figure 2 This is a schematic diagram of the structure of the head unit of the present invention.

[0044] Figure 3 This is a schematic diagram of the structure of the pectoral fin, pectoral fin drive assembly, dorsal fin and inflation / deflation assembly of the present invention.

[0045] Figure 4 This is a schematic diagram of the structure of the present invention after removing the flexible torso shell.

[0046] Figure 5 This is a schematic diagram of the structure of the combination of the bias adjustment mechanism, the swing drive mechanism and the tail unit of the present invention.

[0047] Figure 6 yes Figure 5 The bottom view of the assembly shown.

[0048] Figure 7 This is a flowchart of the motion control method of the present invention.

[0049] Figure 8 This is a timing diagram for CPG phase compensation based on hardware latency characteristics.

[0050] Figure 9 This is a comparison chart showing the output phase continuity with and without a smooth interpolation algorithm.

[0051] Figure 10 This is a schematic diagram of the trajectory tracking workflow of a biomimetic fish robot.

[0052] The diagram shows: 1. Head unit; 11. Fish head shell; 111. Flow deflector; 112. Boss 1; 12. Pectoral fin; 121. Servo 1; 122. Horizontal axis; 13. Dorsal fin; 14. Waterproof cover; 15. Airbag; 16. Electric air pump; 17. Air duct; 18. High-speed solenoid valve; 19. Bracket 1; 2. Torso unit; 21. Torso shell; 22. Mounting plate; 3. Tail unit; 31. Fish tail seat; 32. Tail fin; 33. Boss 2; 41. Motor; 42. Crank; 43. L-shaped rocker arm; 44. Vertical axis; 45. Connecting rod; 46. Fixing plate; 51. Servo 2; 52. Drive shaft; 53. Rotating arm; 54. Cross rod; 55. U-shaped bracket; 551. Plug-in shaft; 56. Universal coupling; 57. Bearing seat; 58. Rigid rod. Detailed Implementation

[0053] The embodiments of the present invention will be described in further detail below with reference to the accompanying drawings and examples. The following examples are for illustrative purposes only and should not be construed as limiting the scope of the invention.

[0054] Example 1, combined with Figures 1 to 6 A biomimetic fish robot includes a head unit 1, a torso unit 2, a tail unit 3, a battery, and a control unit. The head unit 1 includes a fish head shell 11, a pectoral fin 12, and a dorsal fin 13. The fish head shell 11 is a rigid shell with an axisymmetric structure, and the rear end of the fish head shell is a closed structure. The torso unit 2 and the tail unit 3 are arranged sequentially behind the fish head shell 11.

[0055] The front end of the fish head shell 11 has a hemispherical flow guide 111, which is an integral structure with the fish head shell 11. The inner side of the flow guide 111 communicates with the interior of the fish head shell 11. A square opening is provided at the top of the fish head shell 11. A waterproof cover 14 is provided at the square opening of the fish head shell 11. The waterproof cover 14 is detachably fixed and sealed to the outer edge of the square opening of the fish head shell 11. The top surface of the waterproof cover 14 is a streamlined curved surface that smoothly transitions to the upper surface of the fish head shell 11. The battery uses a rechargeable lithium battery of existing technology. Both the battery and the control unit are installed inside the fish head shell 11. The control unit includes a main control board, IMU, depth sensor, Doppler velocimeter, sonar, and vision sensor.

[0056] There are two pectoral fins 12, which are positioned opposite each other on the left and right sides of the outer shell 11 of the fish head. Two pectoral fin drive components are symmetrically arranged inside the outer shell 11 of the fish head. The actuator of each pectoral fin drive component is connected to the pectoral fin 12 on the same side and drives the pectoral fin 12 to rotate, thereby adjusting the tilt angle of the pectoral fin 12.

[0057] Specifically, the pectoral fin 12 is an approximately right-angled triangular plate, with one right-angled side of the pectoral fin 12 close to the outer wall of the fish head shell 11. The front side of the pectoral fin 12 is a convex arc-shaped bevel. Two shaft holes are symmetrically opened on the left and right sides of the fish head shell 11. The pectoral fin drive assembly includes a servo motor 121 and a horizontal shaft 122. The horizontal shaft 122 passes through the shaft hole on the same side and is in a rotational sealing fit with the fish head shell 11. A dynamic sealing assembly is provided between the inner wall of the shaft hole and the horizontal shaft 122. The dynamic sealing assembly adopts existing technology and can effectively prevent water from the outside of the fish head shell 11 from entering its internal cavity through the shaft hole. One end of the horizontal shaft 122 is fixedly connected to the straight edge of the pectoral fin 12 on the same side near the fish head shell 11. In the working state, each pectoral fin 12 rotates together with the corresponding horizontal shaft 122.

[0058] The servo motor 121 is fixed inside the fish head shell 11 by the bracket 2. The output shaft of the servo motor 121 is coaxially and fixedly connected to the other end of the horizontal axis 122. In addition, the battery supplies power to the servo motor 121. The signal terminal of the servo motor 121 is connected to the main control board for communication. The main control board controls the rotation direction and rotation angle of the servo motor 121 through command signals. By changing the rotation angle, rotation direction, and holding time of the pectoral fins, various control effects such as ascent, descent, or attitude stabilization can be achieved under different working conditions. When the pectoral fins on both sides rotate symmetrically, stable lift or downforce can be generated to achieve the ascent, descent, or constant depth control of the bionic fish robot. When the pectoral fins on both sides rotate asymmetrically, additional lateral force and rolling torque can be introduced, allowing the bionic fish robot to complete attitude tilt adjustment or assisted steering.

[0059] The dorsal fin 13 is installed on the top of the fish head shell 11, and there is an air bladder 15 on each of its left and right sides. The fish head shell 11 has two inflation / deflation components inside, and the two inflation / deflation components independently inflate or deflate the two air bladders 15.

[0060] Specifically, the dorsal fin 13 is vertically arranged along the central axis of the fish head shell 11, and the bottom of the dorsal fin 13 is fixedly connected to the upper surface of the waterproof cover plate 14. The air bladder 15 is an elastic diaphragm made of rubber, and its shape is consistent with the shape of the side wall of the dorsal fin 13. The annular edge of the air bladder 15 is fixedly and sealed to the side wall of the dorsal fin 13, so that a closed space is formed between the inner wall of the air bladder 15 and the side wall of the dorsal fin 13.

[0061] The dorsal fin 13 has a mounting hole on each of its left and right sides. The inflation / deflation assembly includes an electric air pump 16 and an air guide tube 17. The electric air pump 16 is fixed inside the fish head shell 11 by a bracket 19. One end of the air guide tube 17 is connected to the outlet end of the electric air pump 16 through a high-speed solenoid valve 18, and the other end is fixedly inserted into the mounting hole of the dorsal fin 13, with its outer wall sealingly fitted to the mounting hole. The end face of the air guide tube 17 is flush with the side wall of the dorsal fin 13. The high-speed solenoid valve 18 is a high-speed solenoid valve with flow metering function. The signal terminals of the electric air pump 16 and the high-speed solenoid valve 18 are respectively connected to the main control board for communication. The two electric air pumps 16 can inflate and deflate the two air bladders 15 respectively. After inflation, the air bladders 15 bulge outwards. During the swimming process, the water flow will generate a lateral force on the outer wall of the inflated air bladders 15, which is used to assist the bionic fish robot in turning. After the airbag 15 is deflated, it can completely adhere to the side wall of the dorsal fin 13 under a small negative pressure.

[0062] In overall motion control, the pectoral fin 12 serves as the angle adjustment actuator for the biomimetic fish robot, working synergistically with the tail wave propulsion and dorsal fin-assisted steering. Compared to traditional fixed dorsal fins, the deformable lateral dorsal fins cause less interference with water flow during straight-line swimming, while quickly participating in control when maneuvering is required, thus achieving a balance between propulsion efficiency and maneuverability. From a biomimetic perspective, the design of the air sacs 15 on both sides of the dorsal fin 13 originates from the mechanism by which natural fish utilize the dorsal fin's function during turning and posture adjustment. Many fish adjust water flow separation and drag distribution by changing the shape of their dorsal fins to assist in turning or stabilizing their bodies.

[0063] The tail unit 3 includes a fish tail seat 31 and a tail fin 32. The fish tail seat 31 is a cone-shaped sealed shell with an axisymmetric structure, and the tail fin 32 is vertically fixed at the rear end of the fish tail seat 31.

[0064] The torso unit 2 includes a flexible torso shell 21, an offset adjustment mechanism, and a swing drive mechanism. The flexible torso shell 21 is an axisymmetric cylindrical structure, with its front and rear ends sealed to the fish head shell 11 and the fish tail seat 31, respectively. The flexible torso shell 21 includes a cylindrical bladder made of silicone rubber, with a spring skeleton fixedly embedded inside. The spring skeleton is made of elastic steel wire wound in a spiral manner, and in its natural state, the overall shape of the spring skeleton is consistent with the shape of the inflated cylindrical bladder.

[0065] The rear end face of the fish head shell 11 has an integral boss 112. The outer contour of the boss 112 is consistent with the outer contour shape of the rear end face of the fish head shell 11 and is located inside the outer contour of the rear end face of the fish head shell 11. In addition, the front end face of the fish tail seat 31 has an integral boss 33. The outer contour of the boss 33 is consistent with the outer contour shape of the front end face of the fish tail seat 31 and is located inside the outer contour of the front end face of the fish tail seat 31.

[0066] The outer contour of the front end face of the flexible torso shell 21 is the same as the outer contour of the rear end face of the fish head shell 11. The front end of the flexible torso shell 21 is fixedly fitted onto the outside of the boss 112, and its inner sidewall is sealed to the circumferential sidewall of the boss 112. In addition, the outer contour of the rear end face of the flexible torso shell 21 is the same as the outer contour of the front end face of the fish tail seat 31. The rear end of the flexible torso shell 21 is fitted onto the outside of the boss 23, and its inner sidewall is sealed to the circumferential sidewall of the boss 23. The outer surface of the flexible torso shell 21 smoothly transitions to the surfaces of the fish head shell 11 and the fish tail seat 31, respectively.

[0067] In addition, another identical inflation / deflation assembly is installed inside the fish head shell 11. The high-speed solenoid valve 18 of this inflation / deflation assembly is connected to the interior of the flexible body shell 21 through a rigid conduit. This rigid conduit is used for inflation and deflation of the interior of the flexible body shell 21. A pressure sensor located inside the flexible body shell 21 is connected to the main control board. The pressure inside the flexible body shell 21 can be adjusted through the rigid conduit according to the change of external water pressure.

[0068] The bias adjustment mechanism is installed inside the flexible body shell 21 via a mounting plate 22, which is fixedly installed at the rear end of the fish head shell 11. The mounting plate 22 is horizontally arranged, and its front end is fixedly connected to the center of the rear side wall of the fish head shell 11. The swing drive mechanism is located at the execution end of the mounting plate 22 and the bias adjustment mechanism. The execution end of the swing drive mechanism is fixedly connected to the front end of the fish tail seat 31. In the working state, the bias adjustment mechanism causes the flexible body shell 21 and the tail unit 3 to be biased towards one side of the central axis of the fish head shell 11 through the swing drive mechanism. At the same time, the swing drive mechanism drives the tail unit 3 to perform regular reciprocating swing.

[0069] Specifically, the bias adjustment mechanism includes a motor 41, a crank 42, and an L-shaped rocker arm 43. The motor 41 is fixed above the mounting plate 22, while the crank 42 and the L-shaped rocker arm 43 are both located below the mounting plate 22. One end of the crank 42 is fixed to the output shaft of the motor 41, and the corner of the L-shaped rocker arm 43 is hinged to the front end of the mounting plate 22 via a vertical shaft 44. The motor 41 is powered by the battery. Furthermore, the signal terminal of the motor 41 is communicatively connected to the main control board, which controls the rotation direction and angle of the output shaft of the motor 41 via command signals.

[0070] The other end of the crank 42 is movably connected to one end of the L-shaped rocker arm 43 via a connecting rod 45. In operation, the motor 41 drives the crank 42 to rotate around the output shaft of the motor 41, and the other end of the crank 42 drives the L-shaped rocker arm 43 to rotate around the vertical axis 44 via the connecting rod 45, so as to adjust the angle of the L-shaped rocker arm 43 relative to the mounting plate 22, thereby adjusting the body unit to offset to one side or return to its original position relative to the fish head shell 11. When the bionic fish robot receives a task to perform turning, the motor 41 drives the L-shaped rocker arm 43 to rotate to a predetermined angle via the connecting rod 45, pulling the flexible body shell 21 and the fish tail base 31 to maintain a controllable offset posture, thus achieving turning during forward movement.

[0071] Specifically, the L-shaped rocker arm 43 is a flat plate structure and is parallel to the mounting plate 22. The swing drive mechanism includes a second servo motor 51, a drive shaft 52, a rotating arm 53, a cross rod 54, and a U-shaped bracket 55. The motor 41 is fixed to the upper surface of the mounting plate 22. The drive shaft 52 is located above the L-shaped rocker arm 43 via a bearing seat 57. The front end of the drive shaft 52 is connected to the output shaft of the second servo motor 51 via a universal coupling 56. The battery supplies power to the second servo motor 51. The signal terminal of the second servo motor 51 is connected to the main control board for communication. The main control board controls the rotation direction and speed of the output shaft of the second servo motor 51 via command signals. The output shaft of the second servo motor 51 drives the drive shaft 52 to rotate via the universal coupling 56. Furthermore, the front end of the drive shaft 52 has an integral spline sleeve, and inside the spline sleeve is a coaxial spline shaft. The spline shaft and the drive shaft 52 are linearly slidingly engaged. The front end of the spline shaft is fixedly connected to one end of the universal coupling 56, and the other end of the universal coupling 56 is fixedly connected to the output shaft of the second servo motor 51.

[0072] The rotating arm 53 is located on one side of the drive shaft 52, and one end of the rotating arm 53 is fixedly connected to the rear end of the drive shaft 52. During the process of the servo motor 51 driving the drive shaft 52 to rotate through the universal coupling 56, the drive shaft 52 drives the rotating arm 53 to rotate around the axis of the drive shaft 52. A fixed plate 46 parallel to the L-shaped rocker arm 43 is provided above it, and the front end of the fixed plate 46 is fixedly connected to the top of the bearing seat 57.

[0073] The cross member 54 is located between the L-shaped rocker arm 43 and the fixed plate 46. The other end of the rotating arm 53 is connected to the cross member 54 via a U-shaped bracket 55. Both the upper and lower ends of the cross member 54 are connected to the fishtail seat 31 via a rigid rod 58. The other end of the rotating arm 53 has a circular insertion hole that is inclined relative to the drive shaft 52. Both the left and right ends of the horizontal section of the cross member 54 are rotatably connected to the U-shaped bracket 55. The front side of the U-shaped bracket 55 has a plug-in shaft 551. One end of the plug-in shaft 551 is fixedly connected to the outer wall of the U-shaped bracket 55, and the other end passes through the circular insertion hole and slides into the other end of the rotating arm 53.

[0074] The upper and lower ends of the vertical section of the cross rod 54 pass through the fixed plate 46 and the L-shaped rocker arm 43 respectively, and are rotatably engaged with the fixed plate 46 and the L-shaped rocker arm 43. The two rigid rods 58 are arranged in parallel, with the front end of each rigid rod 58 fixedly connected to the corresponding end of the vertical section of the cross rod 54, and the rear end fixedly connected to the front end of the fish tail seat 31. When the rotating arm 53 rotates around the drive shaft 52, the rotating arm 53 drives the cross rod 54 to rotate around the axis of its vertical section through the U-shaped bracket 55. The cross rod 54 drives the fish tail seat 31 and the tail fin 32 to swing back and forth around the vertical section of the cross rod 54. The swinging tail fin 32 provides the forward driving force for the bionic fish robot. The swinging frequency of the tail fin 32 is determined by the rotation speed of the servo motor 2 51.

[0075] During underwater movement, the rotation of the output shaft of servo motor 51 is transmitted to the tail fin 32 via universal coupling 56, drive shaft 52, U-shaped bracket 55, and cross rod 54, causing the tail fin 32 to rhythmically reciprocate. The spline joint structure of universal coupling 56 and drive shaft 52 not only ensures the rigidity and smoothness of power transmission but also gives the tail fin flexibility in water movement, enabling it to generate body waves of different amplitudes and frequencies in response to control commands, thus providing the main forward driving force for the bionic fish robot.

[0076] Example 2, combined with Figures 7 to 10 A motion control method, based on the biomimetic fish robot described in Example 1, includes the following steps: Step 1: First, the control unit acquires the real-time attitude angle, three-axis acceleration, depth and sonar data of the bionic fish robot. The acquired data is preprocessed and combined with the current motion state to generate a unified feature vector.

[0077] Step 2: The feature vector from Step 1 is synchronously input into the hierarchical decision system, which includes a high-level strategic reinforcement learning network and a low-level tactical reinforcement learning network. The high-level strategic reinforcement learning network outputs discretely based on the feature vector to generate macroscopic motion pattern commands, which are then sent to the mapping mechanism of the low-level tactical reinforcement learning network and the subsequent central pattern generator. The low-level tactical reinforcement learning network combines the feature vector with the motion pattern commands sent from the high level to continuously output and generate oscillator parameters for precise fine-tuning.

[0078] Most current reinforcement learning applications in robotics are either discrete (action A or B) or continuous (outputting specific numerical values), often resulting in stiff movements or difficulty in convergence. This invention employs a reinforcement learning module based on a Hybrid Action Space architecture, aiming to achieve precise control of both discrete motion mode switching and continuous parameter fine-tuning within the same network framework. This balances the macroscopic speed of decision-making (mode switching) with the microscopic precision of action (parameter fine-tuning). Macroscopic speed refers to the algorithm's ability to quickly output a discrete policy network, i.e., rapidly determine the next motion mode, while microscopic precision refers to the algorithm's ability to achieve more precise control through further output of fine-tuning parameters.

[0079] The high-level strategic reinforcement learning network is responsible for planning the "macroscopic motion pattern" of the global path. This network takes the global state (such as target heading error and trajectory deviation) as input, uses the Softmax activation function to output a discrete probability distribution of predefined motion patterns (such as forward, left turn, right turn, ascent, descent, etc.), and selects the maximum probability to send to the lower-level network. The lower-level tactical reinforcement learning network is responsible for "micro-parameter fine-tuning" in the fluid environment. This network takes the local states of the inner and outer loops (such as mode commands issued by the high-level network, current actual linear velocity and angular velocity) as input, uses the Tanh activation function to output a continuous set of fine-tuning parameter vectors. This vector is directly mapped to the bias and amplitude command combinations in the oscillation equation of the central pattern generator (CPG). This network is trained independently of the high-level network and is specifically designed to compensate for sudden changes in fluid resistance during the movement of the physical prototype, solving the problems of motor current spikes and inconsistent motion posture caused by hard switching of discrete modes, significantly improving the smoothness of the biomimetic robotic fish's movement and the accuracy of micro-control in real water.

[0080] High-level strategic reinforcement learning networks perform macroscopic motion mode switching by mapping global perception information to a predefined discrete state machine probability distribution.

[0081] make For high-level strategic reinforcement learning networks The input state space at time t is defined as: .

[0082] in, For the heading error of the biomimetic fish robot, The lateral deviation of the trajectory of the biomimetic fish robot. For the depth of the biomimetic fish robot, and These represent the distance and azimuth angle of the biomimetic fish robot relative to the obstacle.

[0083] High-level strategic reinforcement learning networks use the Softmax activation function to output discrete motion patterns at the current time step. The probability distribution is used to output seven movement modes, including left turn, sharp left turn, straight, right turn, sharp right turn, ascend, and descend. The mode with the highest probability is selected as the output command. .

[0084] in, and To enhance the weights and biases of the learning network for high-level strategic reinforcement.

[0085] The underlying tactical reinforcement learning network is responsible for outputting continuously fine-tuned parameters under the selected movement pattern.

[0086] make For the underlying tactical reinforcement learning network in The input state space at each time step needs to integrate the instructions from the high-level strategic reinforcement learning network and the physical rhythms of the low-level tactical reinforcement learning network. The input state space is as follows: .

[0087] in, and For the linear velocity and angular velocity of the biomimetic fish robot, The attitude angle vector measured by the IMU. This represents the instantaneous phase of the current central mode generator.

[0088] Through the computation of the underlying tactical reinforcement learning network, the final output is a continuous fine-tuning parameter vector: .

[0089] It is the amplitude compensation coefficient of the central mode generator, which acts on the intrinsic amplitude parameter of the Hopf oscillator and is used to fine-tune the amplitude of the tail fin swing when subjected to backflow or downstream disturbances. The CPG frequency compensation coefficient is used to fine-tune the oscillation frequency of the tail fin in order to achieve linear and smooth adjustment of swimming speed. The tail offset angle adjustment is directly mapped to the motor, which provides a small lateral force to correct heading deviations without switching steering modes. The angle of attack and fine-tuning of the left and right pectoral fins are applied to servo motor one, which compensates for roll and pitch attitude stability by changing the rotation angle of the pectoral fins around the vertical axis.

[0090] The underlying tactical reinforcement learning network uses the Tanh activation function to output a continuously fine-tuned parameter vector. for: .

[0091] in This is the range constraint matrix for the physical actuator. and The weights and biases of the underlying tactical reinforcement learning network.

[0092] If the aforementioned high- and low-level dual-channel networks are jointly trained end-to-end, gradient conflicts and environmental non-stationarity issues are easily triggered due to the different action space properties of discrete decision-making and continuous parameter tuning, making it difficult for the algorithm to converge. To address this, this invention proposes a two-stage training method based on goal decoupling. This method uses the maximization of expected cumulative discount reward to achieve the optimal evolution of the policy.

[0093] Phase 1 (Dedicated Training for the Underlying Tactical Network): Freezing the high-level strategic decision-making network. Discrete motion pattern instructions are issued to the underlying network either randomly or in a fixed sequence. Without obtaining global path coordinates, the underlying tactical network undergoes iterative training with the optimization objective of "approaching the theoretical baseline values ​​of the current pattern with actual kinematic parameters." This phase aims to train the underlying network into a stable and accurate continuous parameter actuator, enabling it to perfectly replicate any predefined pattern.

[0094] Phase Two (High-Level Network Dedicated Training): The trained weights of the underlying tactical network are locked and used as black-box actuators. Training of the high-level network is then initiated. The high-level network outputs discrete mode instructions based on the global state and calls the underlying black-box actuators to drive the biomimetic fish robot's physical model. The optimization objectives in this phase shift to "global trajectory tracking accuracy" and "minimizing mode switching frequency."

[0095] This two-stage training method achieves physical and logical separation of responsibility between "navigation" and "driving," eliminates the gradient shirking problem caused by the dual-layer action space, and significantly shortens the convergence cycle of the reinforcement learning algorithm for the biomimetic fish robot in simulated and real aquatic environments.

[0096] Its global objective function is defined as: .

[0097] In this formula, Indicated by Strategies with parameters The expected cumulative return under the following conditions Represents the expectation operator For the trained control strategy, This is the maximum number of steps per training session. For the current time step The discount factor is used to balance the relationship between immediate rewards and long-term rewards. For a moment The total reward value obtained by the system.

[0098] To address the issue of gradient interference between high and low layers, this invention employs a phased decoupled training strategy. The first phase involves dedicated training of the underlying tactical reinforcement learning network, during which the parameters of the higher-level networks are frozen, and the underlying reward function is defined as: .

[0099] In this formula, Positive weighting coefficients and For real-time kinematic vectors, and and Command mode The theoretical reference vector below; The second norm of a vector; acceleration vector The time derivative is acceleration, and penalizing this term can induce the network to generate a smooth output signal; The terms are used to constrain and control energy consumption.

[0100] After the underlying network converges, the second stage begins, which involves locking the weights of the underlying network and starting training of the high-level strategic network. At this point, the high-level reward function is defined as: .

[0101] In the formula This is the proportionality coefficient; For a moment With time The difference in distance between the bionic fish and the target point is rewarded if the distance decreases. This represents the absolute value of the heading error. For indicator functions, when the mode is switched (i.e. and The value is 1 when the values ​​are different from the given value, and 0 otherwise. This penalty can effectively suppress instruction oscillation. The obstacle avoidance penalty term has the following specific function form: .

[0102] The punishment is Triggered by time, in the formula To determine the severity of the punishment, To maintain a safe distance, The distance to obstacles is measured in real time. This target isolation-based training method allows the two-layer network to optimize in parallel within their respective reward boundaries, eliminating gradient deferral caused by the dual action space and significantly shortening the convergence cycle of the algorithm in complex dynamic waters.

[0103] Step 3: The motion mode command and continuous oscillator parameters are input to the mapping mechanism of the central mode generator. After smoothing interpolation and phase continuity processing inside, they are applied to the Hopf oscillator to generate a continuous and smooth rhythm control signal.

[0104] Step 4: The control signal is precisely distributed to servo motor 1, motor, servo motor 2 and high-speed solenoid valve through the decoupling mapping module, thereby realizing the driving of the bionic fish robot in complex underwater environment.

[0105] To ensure the dynamic stability of the biomimetic fish robot during multi-mode switching in complex underwater environments and to address the potential for abrupt motion changes due to discrete decision-making, this invention designs a dual smoothing mechanism that incorporates phase continuity constraints and parameter smoothing interpolation. First, when the output motion mode changes (e.g., switching from "straight-line propulsion" to "left turn"), the system strictly adheres to the phase continuity constraint principle. The CPG network does not reset the instantaneous state (phase and radial displacement) of the Hopf oscillator; instead, it forcibly uses the final state value from the previous moment as the initial value for the new mode evolution. Utilizing the limiting cycle attraction characteristic of the Hopf oscillator, the motion state evolves continuously along the time axis towards the new target limiting cycle in a second-order manner.

[0106] Building upon this foundation, to further eliminate the mechanical shock caused by abrupt changes in amplitude or frequency, this invention introduces a parameter smoothing interpolation algorithm based on a sigmoid-like function. During the transition window of mode switching, the control parameters do not change abruptly, but rather gradually transition from the current value to the target value according to a smooth curve. This processing method allows the tail fin swing amplitude to gradually increase or decrease, and the pectoral fin angle of attack to change smoothly, thereby effectively suppressing body vibration and ensuring the robot's posture stability and control continuity when executing high-maneuverability commands. Comparison of effects before and after smoothing processing is shown below. Figure 9 As shown.

[0107] This invention employs a Hopf oscillator with self-excited oscillation characteristics as the mathematical model for the CPG, utilizing its limit cycle characteristics to ensure motion robustness, and transforming the mathematical model into specific control commands for each actuator through a specific mapping mechanism. The differential equation of the oscillator is expressed as: .

[0108] First, for the torso and tail units, the oscillation signal output by the CPG is... The PWM pulse width signal of servo motor 2 is mapped to control the amplitude and frequency of the tail fin's oscillation. Simultaneously, the bias term from the reinforcement learning output is directly mapped to the motor's angular position, achieving decoupled control of propulsion and steering. For the pectoral fin drive assembly, by establishing a functional relationship between the pectoral fin rotation angle and the CPG phase, the oscillator output is mapped to the real-time angle-of-attack command of servo motor 1. This is achieved by adjusting the phase difference between the left and right pectoral fins. This enables the generation of lift or the output of roll torque.

[0109] Furthermore, regarding the dorsal fin, considering the response delay of the aerodynamic system, the system not only maps the steering signal to the on / off state of the high-speed solenoid valve, but also introduces a lead angle compensation mechanism based on hardware delay, such as... Figure 8 As shown. Specifically, for the charge / discharge time constant of the dorsal fin aerodynamic circuit, a leading phase is superimposed on the CPG reference phase. This ensures that the peak moment of aerodynamic deformation of the dorsal fin precisely coincides with the peak moment of hydrodynamic torque generated by the oscillation of the tail fin. Through this targeted mapping strategy, the spatiotemporal consistency of all actuators in the "neuro-mechanical-fluid" loop is ensured, solving the engineering problem of asynchronous response of heterogeneous actuators.

[0110] To verify the practical engineering effectiveness of the control method proposed in this invention, an example of "autonomous inspection and dynamic obstacle avoidance" in a complex underwater environment is used. Figure 10 As shown. In the initial stage of the mission, the biomimetic fish robot is in "efficient cruising mode". The reinforcement learning discrete network selects the straight swimming motion mode, the continuous network outputs low-frequency small-amplitude parameter fine-tuning instructions, and the CPG drives the tail fin to generate high-energy-efficiency propulsion waves.

[0111] When the sensor detects an abrupt change in the path, the state input changes, and the hierarchical decision-making system instantly switches its discrete branch to "turning mode," while the continuous branch outputs a large offset parameter. At this moment, the dorsal fin airbag is immediately inflated to provide auxiliary lateral force, the tail offset motor rapidly rotates to change the tail fin's zero position, and the CPG network smoothly transforms the tail fin oscillation into a high-frequency, large-amplitude oscillation through phase continuity and parameter interpolation algorithms, coordinating differential rotation of the pectoral fins to assist steering. After completing the obstacle avoidance maneuver, it automatically and smoothly transitions back to cruising or hovering observation mode based on environmental feedback. This process fully demonstrates the decision-making advantages of this invention in a hybrid action space: it utilizes a discrete strategy to achieve millisecond-level mode response while ensuring the accuracy and stability of the action through continuous parameter fine-tuning control.

[0112] In the description of this invention, unless otherwise stated, "a plurality of" means two or more; the terms "upper," "lower," "left," "right," "inner," "outer," "front end," "rear end," "head," "tail," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, and are only for the convenience of describing the invention and simplifying the description, and do not indicate or imply that the mechanism or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of the invention. Furthermore, the terms "first," "second," "third," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0113] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "connected" and "linked" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0114] Of course, the above description is not intended to limit the present invention, and the present invention is not limited to the examples given above. Any changes, modifications, additions or substitutions made by those skilled in the art within the scope of the present invention should also fall within the protection scope of the present invention.

Claims

1. A biomimetic fish robot, characterized in that, It includes a head unit, a trunk unit, a tail unit, a battery, and a control unit. The head unit includes a fish head shell, pectoral fins, and a dorsal fin. The fish head shell is a rigid shell with an axisymmetric structure and a closed rear end. The trunk unit and the tail unit are arranged sequentially behind the fish head shell. There are two pectoral fins, which are located opposite each other on the left and right sides of the fish head shell. Two pectoral fin drive components are symmetrically arranged inside the fish head shell. The actuator of each pectoral fin drive component is connected to the pectoral fin on the same side and drives the pectoral fin to rotate to adjust its angle. The dorsal fin is located on the top of the fish head shell, with an air sac on each of its left and right sides. Inside the fish head shell are two inflation / deflation components, which independently inflate or deflate the two air sacs. The tail unit includes a fish tail seat and a tail fin. The fish tail seat is a cone-shaped sealed shell with an axisymmetric structure, and the tail fin is vertically fixed at the rear end of the fish tail seat. The torso unit includes a flexible torso shell, an offset adjustment mechanism, and a swing drive mechanism. The flexible torso shell is an axisymmetric cylindrical structure, with its front and rear ends sealed to the fish head shell and the fish tail seat, respectively. The offset adjustment mechanism is located inside the flexible torso shell via a mounting plate. The swing drive mechanism is located at the execution end of the mounting plate and the bias adjustment mechanism, and the execution end of the swing drive mechanism is connected to the fishtail seat. In operation, the bias adjustment mechanism uses the swing drive mechanism to bias the flexible body shell and tail unit toward one side of the central axis of the fish head shell. At the same time, the swing drive mechanism drives the tail unit to swing back and forth regularly. The battery and control unit are both installed inside the fish head shell. The control unit includes a main control board, IMU, depth sensor, Doppler velocimeter, sonar, and vision sensor.

2. The biomimetic fish robot according to claim 1, characterized in that, The front end of the fish head shell has a hemispherical flow guide, the inside of which is connected to the inside of the fish head shell, and a square opening is provided at the top of the fish head shell. The square opening of the fish head shell is equipped with a waterproof cover plate. The waterproof cover plate is detachably fixed and sealed to the outer edge of the square opening of the fish head shell. The top surface of the waterproof cover plate and the surface of the fish head shell are streamlined curved surfaces.

3. The biomimetic fish robot according to claim 1, characterized in that, The dorsal fin is vertically arranged along the central axis of the fish head shell, and the bottom of the dorsal fin is fixedly connected to the upper surface of the waterproof cover plate as one piece. The air bladder is an elastic diaphragm made of rubber, the shape of which is consistent with the shape of the dorsal fin sidewall. The annular edge of the air bladder is fixedly and sealed to the dorsal fin sidewall, so that a closed space is formed between the inner wall of the air bladder and the dorsal fin sidewall. The dorsal fin has a mounting hole on each of its left and right sides. The inflation / deflation assembly includes an electric air pump and an air guide tube. The electric air pump is fixed inside the fish head shell by a bracket. One end of the air duct is connected to the outlet of the electric air pump via a high-speed solenoid valve, and the other end is fixedly inserted into the mounting hole of the dorsal fin with its outer side wall sealed to the mounting hole of the dorsal fin. The end face of the air duct is flush with the side wall of the dorsal fin. The signal terminals of the electric air pump and the high-speed solenoid valve are respectively connected to the main control board for communication.

4. The biomimetic fish robot according to claim 1, characterized in that, The pectoral fin is an approximately right-angled triangular plate, with one right-angled side of the pectoral fin close to the fish head shell. The front side of the pectoral fin is an outwardly convex arc-shaped oblique side, and two axial holes are symmetrically opened on the left and right sides of the fish head shell. The pectoral fin drive assembly includes a servo motor and a horizontal shaft. The horizontal shaft passes through a shaft hole on the same side and is rotated and sealed with the fish head shell. One end of the horizontal shaft is fixedly connected to the straight edge of the pectoral fin on the same side near the fish head shell. The first servo motor is installed inside the fish head shell via the second bracket. The output shaft of the first servo motor is coaxially and fixedly connected to the other end of the horizontal axis. The signal terminal of the first servo motor is connected to the main control board for communication.

5. A biomimetic fish robot according to claim 1, characterized in that, The flexible torso shell includes a tubular bladder made of silicone rubber, with a spring skeleton fixedly embedded inside the bladder; The spring frame is made of elastic steel wire wound in a spiral manner. In its natural state, the overall shape of the spring frame is consistent with the shape of the inflated cylindrical bladder. The rear end face of the fish head shell has a protrusion 1 integral with it, and the front end face of the fish tail base has a protrusion 2 integral with it. The front end of the flexible torso shell is fixedly sleeved on the outside of boss one, and its rear end is sleeved on the outside of boss two. The outer surface of the flexible torso shell smoothly transitions with the surfaces of the fish head shell and the fish tail seat, respectively.

6. A biomimetic fish robot according to claim 1, characterized in that, The mounting plate is arranged horizontally, and its front end is fixedly connected to the center of the rear side wall of the fish head shell. The bias adjustment mechanism includes a motor, a crank, and an L-shaped rocker arm. The motor is fixed above the mounting plate, and the crank and L-shaped rocker arm are both located below the mounting plate. One end of the crank is fixed to the output shaft of the motor, and the corner of the L-shaped rocker arm is hinged to the front end of the mounting plate through a vertical shaft. The other end of the crank is movably connected to one end of the L-shaped rocker arm via a connecting rod. In operation, the motor drives the L-shaped rocker arm to rotate relative to the mounting plate via the crank and connecting rod to adjust the angle of the L-shaped rocker arm.

7. A biomimetic fish robot according to claim 6, characterized in that, The L-shaped rocker arm is a flat plate structure and is parallel to the mounting plate. The swing drive mechanism includes a second servo motor, a drive shaft, a rotating arm, a cross rod, and a U-shaped bracket. The second servo motor is fixed to the upper surface of the mounting plate. The drive shaft is located above the L-shaped rocker arm through a bearing seat. The front end of the drive shaft is connected to the output shaft of the second servo motor through a universal coupling. The rotating arm is located on one side of the drive shaft, and one end of the rotating arm is fixedly connected to the rear end of the drive shaft. A fixed plate parallel to the L-shaped rocker arm is provided above it, and the front end of the fixed plate is fixedly connected to the top of the bearing seat. The cross member is located between the L-shaped rocker arm and the fixed plate. The other end of the rotating arm is connected to the cross member through a U-shaped bracket. The upper and lower ends of the cross member are connected to the fishtail seat through a rigid rod.

8. A biomimetic fish robot according to claim 7, characterized in that, The other end of the rotating arm is provided with a circular insertion hole that is inclined relative to the drive shaft, and both the left and right ends of the horizontal section of the cross rod are rotatably connected to the U-shaped bracket. The front side of the U-shaped bracket has a plug-in shaft. One end of the plug-in shaft is fixedly connected to the outer wall of the U-shaped bracket, and the other end is inserted into a circular plug hole and slidably engaged with the other end of the rotating arm. The upper and lower ends of the vertical section of the cross member pass through the fixed plate and the L-shaped rocker arm respectively, and rotate with the fixed plate and the L-shaped rocker arm. The two rigid rods are arranged in parallel, and the front end of each rigid rod is fixedly connected to the corresponding end of the vertical section of the cross member, and the rear end is fixedly connected to the front end of the fishtail seat.

9. A motion control method, characterized in that, Based on the biomimetic fish robot as described in any one of claims 1 to 8, the control method includes the following steps: Step 1: First, the control unit acquires the real-time attitude angle, three-axis acceleration, depth and sonar data of the bionic fish robot. After data preprocessing and decoupling and splicing, the acquired data is combined with the current motion state to generate a unified feature vector. Step 2: The feature vector is synchronously input into the hierarchical decision system, which includes a high-level strategic reinforcement learning network and a low-level tactical reinforcement learning network. The high-level strategic reinforcement learning network outputs discretely based on the feature vector to generate macroscopic motion pattern commands, which are then sent to the low-level tactical reinforcement learning network and the subsequent CPG mapping mechanism. The low-level tactical reinforcement learning network combines the feature vector with the motion pattern commands from the high level to continuously output and generate oscillator parameters for precise fine-tuning. Step 3: The motion mode command and continuous oscillator parameters are input to the mapping mechanism of the central mode generator. After smoothing interpolation and phase continuity processing inside, they are applied to the Hopf oscillator to generate a continuous and smooth rhythm control signal. Step 4: The control signal is precisely distributed to servo motor 1, motor, servo motor 2 and high-speed solenoid valve through the decoupling mapping module, thereby realizing the driving of the bionic fish robot in complex underwater environment.

10. A motion control method according to claim 9, characterized in that, The high-level strategic reinforcement learning network performs macroscopic motion mode switching by mapping global perception information to a predefined discrete state machine probability distribution. make For high-level strategic reinforcement learning networks The input state space at time t is defined as: ； in, For the heading error of the biomimetic fish robot, The lateral deviation of the trajectory of the biomimetic fish robot. For the depth of the biomimetic fish robot, and These represent the distance and azimuth angle of the biomimetic fish robot relative to the obstacle; High-level strategic reinforcement learning networks use the Softmax activation function to output discrete motion patterns at the current time step. The probability distribution is used to output seven movement modes, including left turn, sharp left turn, straight, right turn, sharp right turn, ascend, and descend. The mode with the highest probability is selected as the output command. ； in, and To enhance the weights and biases of the learning network for high-level strategic reinforcement; The underlying tactical reinforcement learning network is responsible for outputting continuously fine-tuned parameters under the selected movement pattern; make For the underlying tactical reinforcement learning network in The input state space at each time step needs to integrate the instructions from the high-level strategic reinforcement learning network and the physical rhythms of the low-level tactical reinforcement learning network. The input state space is as follows: ； in, and For the linear velocity and angular velocity of the biomimetic fish robot, The attitude angle vector measured by the IMU. This represents the instantaneous phase of the current central pattern generator; Through the computation of the underlying tactical reinforcement learning network, the final output is a continuous fine-tuning parameter vector: ； It is the amplitude compensation coefficient of the central mode generator, which acts on the intrinsic amplitude parameter of the Hopf oscillator and is used to fine-tune the amplitude of the tail fin swing when subjected to backflow or downstream disturbances. The CPG frequency compensation coefficient is used to fine-tune the oscillation frequency of the tail fin in order to achieve linear and smooth adjustment of swimming speed. The tail offset angle adjustment is directly mapped to the motor, which provides a small lateral force to correct heading deviations without switching steering modes. The angle of attack and fine-tuning of the left and right pectoral fins are applied to servo motor one, which compensates for roll and pitch attitude stability by changing the rotation angle of the pectoral fins around the vertical axis. The underlying tactical reinforcement learning network uses the Tanh activation function to output a continuously fine-tuned parameter vector. for: ； in This is the range constraint matrix for the physical actuator. and The weights and biases of the underlying tactical reinforcement learning network.

Images