Variable-structure water-air amphibious robot and motion control method

Abstract

The invention discloses a variable-structure water-air amphibious robot and a motion control method. The robot comprises a water-air amphibious robot body, a variable-structure wing system and an electric control system. The robot body comprises a sealed cabin, a propeller fixing assembly, a fluid shell and a variable support tail cover. The variable-structure wing system comprises a driving gear transmission assembly, a two-section folding wing and a wing folding linkage assembly; the electric control system comprises a control core element, a visual module, a power supply module, a GPS module and an underwater depthometer module which are arranged in the cabin body; the motion control method is based on a deep reinforcement learning control method, control generalization ability is enhanced through teacher and student strategies and random domain, and the motion control method has more advantages in scenes such as inaccurate modeling and large external disturbance change. The structure of the variable-structure water-air amphibious robot integrates the structures of a rotorcraft, a fixed-wing aircraft, an underwater robot and the like, the modes of the wings can be switched, and efficient movement operation of the water-air amphibious robot is achieved.

Description

Technical Field

[0001] This invention relates to the fields of rotary-wing and fixed-wing unmanned aerial vehicles and underwater robots, specifically to a variable-structure amphibious robot suitable for scenarios such as coastal patrol, terrain exploration, logistics transportation, and monitoring and rescue. Background Technology

[0002] As research into unmanned basic application technologies such as unmanned aerial vehicles (UAVs) and underwater vehicles (UUVs) matures, the operational space of aircraft and marine vehicles is being continuously expanded and is becoming increasingly intertwined and integrated. The development of new aerial robots is gradually moving towards integrated air-sea cross-media aircraft.

[0003] In existing structural research on amphibious robots, they can be basically divided into fixed-wing, multi-rotor, hybrid-wing, biomimetic-wing, and swept-wing configurations. The basic principle is to minimize the watertight impact and underwater fluid resistance by using the entry and exit points as the minimum cross-section. Therefore, variable-structure amphibious robots not only combine the rapid deployment and takeoff of rotorcraft, the long-endurance flight of fixed-wing aircraft, and the exploration and stealth capabilities of underwater robots, but also, through their ability to traverse both air and water media, can perform more complex and specialized cross-domain exploration tasks. Amphibious robots have broad application prospects and development potential, such as underwater rescue, underwater structural damage detection, joint atmospheric and oceanic observation, rapid emergency response, long-range detection and tracking of coastal patrols, terrain exploration, logistics transportation, and monitoring and rescue.

[0004] In terms of motion control algorithms for cross-media multimodal applications, although traditional model-based control methods are relatively common, deep reinforcement learning-based control methods have been studied and experimentally verified by many researchers in more complex scenarios for amphibious missions that require high robustness, high applicability, and low modeling accuracy. Moreover, the controllers have excellent performance. However, the existing reinforcement learning-based research results in the field of amphibious robot control have not been able to effectively solve the problems of large changes in amphibious motion modes and large media spans. Summary of the Invention

[0005] Given the current lack of a good combination of the performance of rotorcraft, fixed-wing aircraft and underwater robots in the field of amphibious robots, as well as the shortcomings of existing amphibious robots such as excessive size, low flexibility, poor adaptability and high cost, this invention proposes a variable structure amphibious robot system and motion control method. This invention aims to provide a more efficient, economical and stable amphibious robot system solution.

[0006] The present invention provides a variable structure amphibious robot, comprising an amphibious robot body, a variable structure wing system, and an electronic control system;

[0007] Furthermore, the robot body includes a sealed chamber, a thruster fixing assembly, a fluid outer shell, and a variable support tail cover. The sealed chamber is divided into an electrical control chamber and a battery chamber, which are aligned and fixed front to back by the thruster fixing assembly, forming a two-cylinder configuration that is larger at the front and smaller at the rear. Both chambers are connected after being watertightened with waterproof material. The outer perimeter of the body in contact with the medium is supplemented by a hollow fluid outer shell and a variable support tail cover for flow rectification, resulting in a streamlined design with a larger front end and a smaller rear end. This reduces fluid resistance during aerial and underwater movement without significantly increasing the body's weight.

[0008] Furthermore, the electrical control compartment and battery compartment of the sealed chamber are mechanically and statically sealed near the thruster fixed assembly using O-rings and transparent circular sealing covers, respectively. Multiple circumferential bolts are used to connect and compress the sealing cover to ensure underwater waterproof performance. The electrical watertight connectors penetrate the transparent sealing cover to achieve electrical link connection between the inside and outside of the body. The underwater depth gauge module is also installed and connected using this method.

[0009] Furthermore, the propulsion fixing assembly includes a polygonal carbon fiber plate-shaped bracket connector, a ducted fan fixing component, and a ducted fan propulsion unit; wherein the rotor blade configuration of the ducted fan propulsion unit is that rotor 1 and rotor 3 are counterclockwise blades, and rotor 2 and rotor 4 are clockwise blades.

[0010] Furthermore, the fluid outer shell and variable support tail cover include an annular-wing coupling rectifier, a porous annular component, and a support tail cover assembly; the annular-wing coupling rectifier is nested on the outside of the support connector for rectifying air and water media to reduce motion resistance; the support tail cover assembly passively opens when in contact with the ground, becoming an unfolded landing support to maintain the stability of the robot body during vertical descent.

[0011] Furthermore, the variable structure wing system includes a drive gear transmission assembly, a two-section folding wing, and a wing folding linkage assembly; the helical gear transmission component consists of a 16-tooth 0.8-module curved bevel gear and a 24-tooth 0.8-module curved bevel gear, with a transmission ratio of 3:2 and a transmission shaft angle of 54.74 degrees. The drive source uses a 180-degree micro waterproof servo motor, with the servo motor output shaft connected to the 16-tooth curved bevel gear. The transmission shaft is integrated with the inner wing to enable the wing parts to cruise in the air in a fixed-wing mode in the horizontal direction when deployed. The wing rotation angle is 120 degrees. In the folded state, the wing is retracted in the vertical direction to minimize the frontal cross-section of the body, ensuring minimal fluid resistance during water entry, exit, and underwater movement.

[0012] Furthermore, the two-section folding wing includes an inner wing component and an outer wing component, which together form a hinge structure that allows for 180-degree rotation between each other. The rotation axis is perpendicular to the windward section of the wing and offset downward by a unit distance of axial radius. When unfolded, the inner and outer wing components can be spliced ​​together to form a seamless complete wing. The complete wing is forward-swept with a forward sweep angle of 3 degrees, which is used to provide lift during fixed-wing cruise and underwater gliding. When folded, the outer wing component folds 180 degrees inward and then folds parallel to the inside of the inner wing component, realizing the function of folding the two wing sections into one section with the smallest possible cross section. The wing structure remains in a folded state during the robot's vertical take-off and landing, entry and exit from the water surface, and normal underwater movement.

[0013] Furthermore, the electronic control system includes a control core component, a vision module, a power module, a GPS module, and an underwater depth gauge module, all housed within the cabin. The control core component controls the robot's movements and includes an onboard computer that deploys control algorithms at the upper level and flight control hardware that interacts with components such as the thrusters, GPS module, and underwater depth gauge module at the lower level. The vision module is used to acquire image information of the surrounding environment. The power module supplies power to the robot's various electrical components. The GPS module acquires the robot's position information. The underwater depth gauge module acquires depth information during underwater movement.

[0014] Furthermore, the electronic control system is installed in the electronic control compartment through a vertical two-layer internal support structure, the battery module is installed in the battery compartment, and the vision module is installed near the head to achieve a better visual field. The separate design of the battery compartment and the battery compartment allows for quick battery swapping without disassembling the electronic control compartment during daily use of the robot.

[0015] Furthermore, the core control component is used to control the robot's movements. It includes an onboard computer that deploys the control algorithm at the upper level and flight control hardware that interacts with components such as the thruster, GPS module, and underwater depth gauge module at the lower level. The onboard computer uses a carrier board with a Linux system that can run ROS. It communicates with the GPS module and underwater depth gauge via serial port and with the thruster via PWM control. The flight control hardware itself has a 9-axis IMU module that provides measurement information, which communicates with the onboard computer via Mavros.

[0016] Furthermore, the wing linkage folding transmission of the variable structure wing system includes a transmission rope, an end fixing component, and a pivot spring. One end of the transmission rope is connected to the outer wing component, and the other end passes through the cylindrical hollow groove of the inner wing and is connected to the transmission shaft mounting bracket, continuously spanning the wing rotation axis and the inner and outer wing rotation axes. The pivot spring is used for the rotation of the inner and outer wings to achieve an initial overlapping state. The variable structure wing system pulls the transmission rope at the same time as the inner wing unfolds, causing the outer wing component to unfold simultaneously. When the inner wing folds, the transmission rope is released, and the outer wing component returns to the overlapping state due to the elastic force of the pivot spring. This achieves synchronous linkage rotation of the outer wing by pulling the transmission rope while the inner wing is driven by the bevel gear transmission of the servo motor.

[0017] This invention also discloses a method for the movement of a variable-structure amphibious robot using the above-mentioned method, characterized by comprising the following steps:

[0018] S1. The robot takes off vertically from the horizontal plane in quadcopter mode with its wings folded.

[0019] S2. After the robot reaches a certain altitude, its wings unfold and it transitions to a level flight state. It then cruises in the air in a fixed wing mode, with the thrusters providing flight power, yaw and pitch moments.

[0020] S3. After reaching the water surface, the robot glides head-down towards the water surface and changes into a folded state before contacting the water, and then enters the water in the folded state.

[0021] S4. After the robot enters the water, the thruster rotor can run in both directions at low speed. The forward and reverse thrust and torque generated by the rotor are used as the power source for the robot to move underwater. During normal underwater movement, the wings remain folded. When deceleration is required or water resistance is generated, the wings will unfold at different angles. When fully unfolded, it will glide underwater in underwater glider mode.

[0022] S5. The robot accelerates vertically upward in a folded state at a certain depth underwater, and relies on inertia to break out of the water surface by impact. After the thruster rotor is completely out of the water, the rotor switches to high-speed rotation mode to generate air thrust, and then leaves the water surface in quadcopter mode.

[0023] S7. After leaving the water, the robot's wings are deployed, and it performs fixed-wing horizontal cruising.

[0024] S6. When the robot approaches the ground, its wings convert to a folded state and it flies in quadcopter mode. When the passive landing gear at the tail touches the ground, it unfolds and completes a vertical landing.

[0025] This invention also discloses a deep reinforcement learning control method for the motion of a variable-structure amphibious robot, characterized by comprising the following steps:

[0026] A reinforcement learning-based control method for the motion of a variable-structure amphibious robot, as described in any one of claims 1-8, is characterized by comprising the following steps:

[0027] S1. Establish a dynamic model for robot motion control. Since the robot's main drive source consists of two wing-driven servo motors and four propeller motors, to simplify the data flow of reinforcement learning, it is modeled as a 6-DOF rigid body driven by four motors. The deployment and folding of the wings only affect factors such as air lift, drag, or underwater drag-lift. Therefore, to meet the needs of large-scale training, a lightweight computational dynamic simulation model equation is established, expressed as follows:

[0028]

[0029] Where ⊙ represents a quaternion rotation, p WB For position, q WB For posture, v WB For velocity, ω B Angular velocity, g W For gravity, f p For thrust, f i For air or underwater resistance, τ p For duct torque;

[0030] S2. Establish a single motor thrust model: f i =K f A t A t ∈[0,1]; coefficient K in the thrust model f The value is obtained based on tests in real-world motion on different media, and is a second-order nonlinear curve function;

[0031] S3. A robot reinforcement learning control algorithm framework is built based on the deep reinforcement learning PPO algorithm, including the loss function J(θ) and its gradient. The formula is as follows:

[0032]

[0033] in, Let R(τ) be the expectation operator, and a be the return. t For action, s t For state, π θ Let be the action probability function;

[0034] The generalized advantage estimation (GAE) method is used to reduce variance in the training reward feedback to construct the advantage function. To reconstruct the reward signal:

[0035]

[0036] Where, γ t′-t r is the power of the discount factor. t' As a reward, V ω (s t ) is the state value function;

[0037] Proximal ratio pruning loss is used to limit the policy update amplitude L. CLIP (θ) is as follows:

[0038]

[0039] Where, r t (θ) represents the probability ratio between the old and new strategies;

[0040] Value loss function L VF (θ) is used to optimize the strategy, as follows:

[0041] L VF (θ)=(V θ -V target )

[0042]

[0043] Among them, V θ For current value, V target For target value;

[0044] Add student-imitation teacher strategies, use KL divergence judgment and knowledge distillation for fusion training, and finally combine them to generate a new TS-PPO advantage function and objective function:

[0045]

[0046] in, The old dominance function, γ is the discount factor. For knowledge distillation loss, r t (θ) represents the probability ratio, clip() represents the clipping function, and λ represents the weighting coefficient;

[0047] Then, the policy network and value network are iteratively updated based on the Actor-Critic method, and the network structure is an MPL multilayer fully connected network (512, 256, 128).

[0048] Define the state input vector: s = [p x ,p y ,p z ,q w ,qx ,q y ,q z ,v x ,v y ,v z ,w x ,w y ,w z ] T p x ,p y ,p z For position, q w ,q x ,q y ,q z For the four elements of posture, v x ,v y ,v z For linear velocity, w x ,w y ,w z Let be the angular velocity; define the output action: A = [a1, a2, a3, a4] T Where a1, a2, a3, and a4 represent the thrust magnitudes of the four ducted thrusters;

[0049] S4. Enhance the control generalization ability using the randomization method. The specific randomization equation is expressed as follows:

[0050]

[0051] Among them, M flow For hydrodynamic disturbance functions, For IMU attitude noise, For depth gauge noise, M represents body mass buoyancy, I represents inertial buoyancy, and F represents... d For random external force, F air For aerodynamic external force, T fail This is the thrust failure vector;

[0052] S5. Build a reinforcement learning simulation platform in the IsaacGym environment to perform parallel training of multiple agents, with the target location as an additional state input: S h =[p hx ,p hy ,p hz ] T p hx ,p hy ,p hz Let XYZ be the position in space, a 3D vector;

[0053] Among them, the total S during simulation A The input has a total of 13 + 3 = 16 dimensions:

[0054] SA =S h +S

[0055] Reward function settings: designed to guide the aircraft to the target location, the total value R is defined as follows:

[0056] R = r p +r p (r l +r w )+r a +r s

[0057] Location reward r p (d t (This represents the Euclidean distance to the target):

[0058]

[0059] Verticality reward r l Suppress rotation reward r w (l upright factor, w rotation factor):

[0060]

[0061] Collision penalty r a :

[0062]

[0063] Contact target reward r s :

[0064]

[0065] The training uses MLP connection layer parameters (512, 256, 128), and after training, a lightweight neural network model is generated.

[0066] S6. Run a lightweight neural network model on the onboard computer board of the robot's core control component, where the state input and output interact with ROS topic information, use Mavros to perform thrust mapping and sensor interaction with the lower-level machine, and finally realize the deployment of the robot's reinforcement learning controller.

[0067] The beneficial effects of this invention are:

[0068] (1) The variable structure amphibious robot body configuration of the present invention integrates the structures of rotorcraft, fixed wing, and underwater robot, and uses foldable wings to complete the form switching. In the folded state, it can take off and land vertically on the ground like a rotorcraft without relying on the runway. It has higher flexibility in takeoff and small-range movement. When conducting long-range cruise in the air, it can switch to the unfolded state like a fixed wing and use the lift generated by the wings to offset a certain amount of its own weight, thereby extending the flight range. In addition, when facing the strong impact or retention force when crossing the water surface, the impact can be minimized by entering and exiting the water with the smallest cross section. Therefore, the robot body can efficiently complete cross-medium tasks, such as underwater rescue, underwater structure damage detection, joint atmospheric and ocean observation, rapid emergency response, remote detection and tracking of coastal patrols, terrain exploration, logistics transportation and monitoring and rescue.

[0069] (2) The variable structure amphibious robot body of the present invention is made of lightweight waterproof materials such as acrylic, aluminum alloy, PLA, nylon, etc., in order to balance the contradiction between the water tightness of the underwater robot and the lightweight of the aircraft. The robot body is small in size, and a thin-walled fairing structure is designed to achieve a streamlined shape. The wing structure is completed by additive manufacturing. Finally, a single-layer wall thickness and 3% space spiral filling are used to achieve ultra-light weight while ensuring basic strength. The tail fairing is also designed as a passive landing gear, reducing the weight of additional parts with a multi-functional integrated structure.

[0070] (3) The variable structure amphibious robot of the present invention uses a split electronic control compartment and a battery compartment, which can realize the replacement of the battery compartment when the battery needs to be replaced quickly, without the need to operate the electronic control compartment. Furthermore, the propulsion system uses a ducted fan as the power source. The ducted fan is an intermediate state between an air rotor and an underwater rotor, which avoids the problem that the rotation radius of the air rotor is too long, resulting in excessive drag of the underwater rotor. This realizes a dual-purpose propulsion device that can be used in both air and water media.

[0071] (4) The variable structure amphibious robot of the present invention uses a variable structure wing system and designs a folding wing with a spatial transmission shaft angle of 54.74 degrees. It can achieve the folding of the wing from the horizontal state to the vertical state with a single degree of freedom. It designs a segmented wing configuration so that a larger wing area can be unfolded in a smaller folding space to achieve greater lift in the air. Furthermore, it designs a passive linkage folding method through a transmission rope and spring structure. Finally, it can perform the overall folding of the wing and the folding of the inner and outer wings using only the rudder drive source on the wing.

[0072] (5) The variable structure amphibious robot of the present invention uses complete electronic control components, including vision, main control, lower computer, sensors and other components, which are installed in the electronic control cabin in a highly efficient stacking manner, so that it can carry advanced algorithms, realize the robot's overall intelligent and autonomous movement, and better realize autonomous exploration tasks in more complex aerial and underwater environments.

[0073] (6) The deep reinforcement learning-based control method used in the variable structure amphibious robot of the present invention addresses the problems that traditional control methods (PID, MPC) struggle with, such as unknown media disturbances and inaccurate robot modeling. Deep reinforcement learning control can achieve more robust and efficient control effects, especially when dealing with the motion of the robot in a variable structure state. Based on the PPO algorithm, a teacher-student imitation learning strategy and a randomized domain method are added. The robot's own electronic control system is equipped with a ROS system for reinforcement learning control deployment, providing a new motion control method and process for the field of amphibious robots. Attached Figure Description

[0074] The present invention will be further described below with reference to the accompanying drawings and embodiments:

[0075] Figure 1 This is an isometric view of the overall structure of the robot of the present invention.

[0076] Figure 2 This is a front view of the overall structure of the robot of the present invention.

[0077] Figure 3 This is a top view of the overall structure of the robot of the present invention.

[0078] Figure 4 This is a side sectional view of the overall structure of the robot of the present invention.

[0079] Figure 5 This is an isometric view of the wing-folded state of the amphibious robot body of the present invention.

[0080] Figure 6 This is a schematic diagram of the variable structure wing system of the amphibious robot of the present invention.

[0081] Figure 7 This is a schematic diagram of the internal electronic control system of the amphibious robot of the present invention.

[0082] Figure 8 This is a schematic diagram of the propulsion assembly of the amphibious robot of the present invention.

[0083] Figure 9 This is a schematic diagram of the internal sealed cabin structure of the amphibious robot of the present invention.

[0084] Figure 10This is a schematic diagram of the fluid housing and variable support tail cover of the present invention.

[0085] Figure 11 This is a schematic diagram of the wing drive gear transmission assembly of the present invention.

[0086] Figure 12 This is a schematic diagram of the two-section folding wing of the present invention.

[0087] Figure 13 This is a schematic diagram of the wing folding linkage mechanism of the present invention.

[0088] Figure 14 This is an isometric sectional view of the electronic control system of the present invention.

[0089] Figure 15 This is a schematic diagram of the method for moving the variable-structure amphibious robot of the present invention.

[0090] Figure 16 This is a schematic diagram of the coordinate system definition and simulation environment for the amphibious robot of the present invention.

[0091] Figure 17 This is a data graph showing the simulation results of reinforcement learning training for the amphibious robot of the present invention. Detailed Implementation

[0092] Example 1

[0093] like Figure 1 As shown, this embodiment describes a variable-structure amphibious robot, in its wing-deployed state, as follows: Figure 1-4 As shown, it includes the amphibious robot body 1, the variable structure wing system 2, and the electronic control system 3;

[0094] like Figure 5 The diagram shown is a schematic of the robot's wing folding state. The robot body 1 includes a sealed cabin 11, a thruster fixing assembly 12, a fluid shell 13, and a variable support tail cover 14.

[0095] like Figure 6 As shown, the variable structure wing system 2 includes a drive gear transmission assembly 21, a two-section folding wing 22, and a wing folding linkage assembly 23;

[0096] like Figure 7 As shown, the electronic control system 3 includes a control core component 31, a vision module 32, a power module 33, a GPS module 34, and an underwater depth gauge module 35, all housed within the cabin.

[0097] like Figure 8As shown, the main sealed chamber 11 includes an electronically controlled sealed chamber 111, a battery sealed chamber 112, and a transparent circular sealing cover 113. The electronically controlled sealed chamber 111 and the battery sealed chamber 112 are connected by a thruster fixing bracket 121. The electronically controlled sealed chamber 111 is composed of a front plate-shaped circular transparent component 114, a front conical fairing 115, a middle cylindrical hollow transparent component 116, and a rear cylindrical metal seal 117 with four lifting lugs and a sealing ring mounting groove, which are sealed together to form a single chamber. The battery sealed chamber 112 is composed of a front cylindrical metal seal 118 with four lifting lugs and a sealing ring mounting groove, a middle cylindrical hollow transparent component 119, and a rear circular sealing cover 120, which are sealed together to form a single chamber. The transparent circular sealing cover 113 is connected to the electronically controlled sealed chamber 111 and the battery sealed chamber 112 by O-rings and multiple circumferential bolts to achieve a mechanical static seal, ensuring underwater waterproof performance.

[0098] like Figure 9 The diagram shows a thruster fixing assembly 12, which includes a polygonal carbon fiber plate-shaped support connector 121, a ducted fan fixing component 122 with two lifting lugs, and a ducted fan thruster 123. The thruster fixing assembly 12 consists of the support connector 121, the ducted fan fixing component 122, and the ducted fan thruster 123 as a group, which are symmetrically arranged along the horizontal and vertical planes of the electronically controlled sealed cabin 111. A total of 4 groups constitute the entire thrust system to provide the robot with motion power. The ducted fan rotor blade configuration is that rotors 124 and 125 are counterclockwise blades, and rotors 126 and 127 are clockwise blades.

[0099] like Figure 10 The diagram shows the fluid outer shell 13 and the variable support tail cover 14. The fluid outer shell 13 includes a ring-wing coupling rectifier 131 and a porous annular component 132. The ring-wing coupling rectifier is a hollow thin-walled outer shell component composed of a ring shape and four wing-shaped components, which is nested on the outside of the support connector for rectifying air and water media to reduce motion resistance. The variable support tail cover 14 consists of six separate cover plates 141, six tail cover springs 142, and a tail cover base 143. The tail cover plates 141 are passively closed by the springs 142, and passively opened when the tail cover 14 contacts the ground, becoming an unfolded landing support to maintain the stability of the robot body during vertical descent.

[0100] like Figure 11As shown, the drive gear transmission assembly 21 of the variable structure wing system 2 includes a folding drive servo 211, a 16-tooth 0.8-module curved bevel gear 212, a 24-tooth 0.8-module curved bevel gear 213, and a drive shaft mounting bracket 214. The transmission ratio is 3:2, and the drive shaft angle is 54.74 degrees. The input shaft of the drive servo 213 is connected to the 16-tooth curved bevel gear 211, and the curved bevel gear 213 on the output shaft is integrated with the inner wing component 221 as a whole. In addition, the input shaft is axially fixed using bearings and the output shaft is axially fixed using flange bearings 215. Finally, when the wing component is unfolded, it can cruise in the air in a fixed-wing mode in the horizontal direction. The rotation range of the rotation shaft is 120 degrees. After folding, it is retracted in the vertical direction to minimize the cross-section of the main body, so as to ensure minimal fluid resistance when entering, exiting, and moving underwater.

[0101] like Figure 12 As shown, the two-section folding wing 22 includes an inner wing component 221 and an outer wing component 222, which together form a hinge structure that allows for 180-degree rotation between them. The hinge rotation axis is perpendicular to the windward section of the wing and offset downward by a unit distance of axial radius. Two pins and two miniature flange bearings are embedded to achieve shaft assembly. When unfolded, the inner wing component 221 and the outer wing component 222 can be spliced ​​into a seamless complete wing. The complete wing is forward-swept with a forward sweep angle of 3 degrees, which is used to provide lift for fixed-wing cruise flight and underwater gliding. When folded, the outer wing component 222 folds 180 degrees inward and then folds parallel to the inner side 221 of the inner wing component, realizing the function of folding the two wing sections into one section with the robot facing the smallest cross section. The inner wing component 221 has two cylindrical hollow grooves 224 inside, through which the transmission rope 231 for folding linkage passes.

[0102] like Figure 12-13 As shown, the wing folding linkage component 23 of the variable structure wing system 2 includes a transmission rope 231 and a pivot spring 232. One end of the transmission rope 231 is connected to the outer wing component 222, and the other end passes through the cylindrical hollow groove 224 of the inner wing 221 and is connected to the transmission shaft mounting bracket 214, continuously crossing the wing rotation axis and the inner and outer wing hinge rotation axis. The pivot spring 232 is used for the rotation of the inner and outer wings to achieve the initial position of the folded state. The variable structure wing system 2 pulls the transmission rope 231 at the same time as the inner wing 221 unfolds, so that the outer wing component unfolds at the same time. When the inner wing 221 folds, the transmission rope is released, and the outer wing component 222 is restored to the folded state by the elastic force of the pivot spring 232.

[0103] like Figure 14As shown, the core control element 31 of the electronic control system 3 is used to control the robot's movements. It includes an onboard computer 311 that deploys control algorithms at the upper level and flight control hardware 312 that interacts with components such as the ducted thruster 123, GPS module 34, and underwater depth gauge module 35 at the lower level. The vision module 32 is used to collect image information of the surrounding environment. The power module 33 is used to supply power to the robot's various electrical components. The GPS module 34 is used to collect the robot's position information. The underwater depth gauge module 35 is used to collect depth information when moving in water. The vision module 32 and GPS module 34 are installed in the sensor mounting bracket 313, and the onboard computer 311 and flight control hardware 312 are installed in the mounting bracket 314. The mounting bracket 314 is connected to the cylindrical metal seal 118 by bolts. The underwater depth gauge module 35 is installed on the transparent circular sealing cover 113, and three watertight connectors 315 are installed in a circumferential array on the transparent circular sealing cover 113 for internal and external connections of the watertight circuit compartment. The onboard computer uses a Linux system carrier board that can run ROS. It communicates with the GPS module 34 and the underwater depth gauge module 35 via serial port and with the ducted thruster 123 via PWM control. The flight control hardware 312 itself has a 9-axis IMU module that provides measurement information that communicates with the onboard computer 311 via Mavros.

[0104] Example 2

[0105] This embodiment is based on the method for the movement of variable-structure amphibious robots proposed in Embodiment 1, such as... Figure 15 As shown, it includes the following steps:

[0106] S1. The robot takes off vertically from the horizontal plane in quadcopter mode with its wings folded.

[0107] S2. After the robot reaches a certain altitude, its wings unfold and it transitions to a level flight state. It then cruises in the air in a fixed wing mode, with the thrusters providing flight power, yaw and pitch moments.

[0108] S3. After reaching the water surface, the robot glides head-down towards the water surface and changes into a folded state before contacting the water, and then enters the water in the folded state.

[0109] S4. After the robot enters the water, the thruster rotor can run in both directions at low speed. The forward and reverse thrust and torque generated by the rotor are used as the power source for the robot to move underwater. During normal underwater movement, the wings remain folded. When deceleration is required or water resistance is generated, the wings will unfold at different angles. When fully unfolded, it will glide underwater in underwater glider mode.

[0110] S5. The robot accelerates vertically upward in a folded state at a certain depth underwater, and relies on inertia to break out of the water surface by impact. After the thruster rotor is completely out of the water, the rotor switches to high-speed rotation mode to generate air thrust, and then leaves the water surface in quadcopter mode.

[0111] S7. After leaving the water, the robot's wings are deployed, and it performs fixed-wing horizontal cruising.

[0112] S6. When the robot approaches the ground, its wings convert to a folded state and it flies in quadcopter mode. When the passive landing gear at the tail touches the ground, it unfolds and completes a vertical landing.

[0113] Example 3

[0114] This embodiment is based on the reinforcement learning method for motion control of variable structure amphibious robots proposed in Embodiment 1. The method specifically includes the following steps:

[0115] S1. Establish a dynamic model for robot motion control. Since the robot's main drive source consists of two wing-driven servo motors and four propeller motors, to simplify the data flow of reinforcement learning, it is modeled as a 6-DOF rigid body driven by four motors. The deployment and folding of the wings only affect factors such as air lift, drag, or underwater drag-lift. Therefore, to meet the needs of large-scale training, a lightweight computational dynamic simulation model equation is established, expressed as follows:

[0116]

[0117] Where ⊙ represents a quaternion rotation, p WB For position, q WB For posture, v WB For velocity, ω B Angular velocity, g W For gravity, f p For thrust, f i For air or underwater resistance, τ p For duct torque;

[0118] S2. Establish a single motor thrust model: f i =K f A t A t ∈[0,1]; coefficient K in the thrust model f The value is obtained based on tests in real-world motion on different media, and is a second-order nonlinear curve function;

[0119] S3. A robot reinforcement learning control algorithm framework is built based on the deep reinforcement learning PPO algorithm, including the loss function J(θ) and its gradient. The formula is as follows:

[0120]

[0121] in, Let R(τ) be the expectation operator, and a be the return. t For action, s t For state, π θ Let be the action probability function;

[0122] The generalized advantage estimation (GAE) method is used to reduce variance in the training reward feedback to construct the advantage function. To reconstruct the reward signal:

[0123]

[0124] Where, γ t′-t r is the power of the discount factor. t' As a reward, V ω (s t ) is the state value function;

[0125] Proximal ratio pruning loss is used to limit the policy update amplitude L. CLIP (θ) is as follows:

[0126]

[0127] Where, r t (θ) represents the probability ratio between the old and new strategies;

[0128] Value loss function L VF (θ) is used to optimize the strategy, as follows:

[0129] L VF (θ)=(V θ -V target )

[0130]

[0131] Among them, V θ For current value, V target For target value;

[0132] Add student-imitation teacher strategies, use KL divergence judgment and knowledge distillation for fusion training, and finally combine them to generate a new TS-PPO advantage function and objective function:

[0133]

[0134] in, The old dominance function, γ is the discount factor. For knowledge distillation loss, r t(θ) represents the probability ratio, clip() represents the clipping function, and λ represents the weighting coefficient;

[0135] Then, the policy network and value network are iteratively updated based on the Actor-Critic method, and the network structure is an MPL multilayer fully connected network (512, 256, 128).

[0136] Define the state input vector: S = [p x ,p y ,p z ,q w ,q x ,q y ,q z ,v x ,v y ,v z ,w x ,w y ,w z ] T p x ,p y ,p z For position, q w ,q x ,q y ,q z For the four elements of posture, v x ,v y ,v z For linear velocity, w x ,w y ,w z Let be the angular velocity; define the output action: A = [a1, a2, a3, a4] T Where a1, a2, a3, and a4 represent the thrust magnitudes of the four ducted thrusters;

[0137] S4. Enhance the control generalization ability using the randomization method. The specific randomization equation is expressed as follows:

[0138]

[0139] Among them, M flow For hydrodynamic disturbance functions, For IMU attitude noise, For depth gauge noise, M represents body mass buoyancy, I represents inertial buoyancy, and F represents... d For random external force, F air For aerodynamic external force, T fail This is the thrust failure vector;

[0140] S5. Build a reinforcement learning simulation platform in the IsaacGym environment to perform parallel training of multiple agents, with the target location as an additional state input: S h=[p hx ,p hy ,p hz ] T , where X, Y, and Z are the spatial positions, and 3 are vectors;

[0141] Among them, the total S during simulation A The input has a total of 13 + 3 = 16 dimensions:

[0142] S A =S h +S

[0143] Reward function settings: designed to guide the aircraft to the target location, the total value R is defined as follows:

[0144] R = r p +r p (r l +r w )+r a +r s

[0145] Location reward r p (d t (This represents the Euclidean distance to the target):

[0146]

[0147] Verticality reward r l Suppress rotation reward r w (l upright factor, w rotation factor):

[0148]

[0149] Collision penalty r a :

[0150]

[0151] Contact target reward r s :

[0152]

[0153] The training uses MLP connection layer parameters (512, 256, 128), and after training, a lightweight neural network model is generated.

[0154] S6. Run a lightweight neural network model on the onboard computer board of the robot's core control component, where the state input and output interact with ROS topic information, use Mavros to perform thrust mapping and sensor interaction with the lower-level machine, and finally realize the deployment of the robot's reinforcement learning controller.

[0155] The experimental data are as follows:

[0156] like Figure 16 The image shows a simulation training exercise for a robot in quadcopter flight mode. Its structure and coordinate system are defined as follows: World coordinate system: w(o W x W y W z W Body coordinate system: B(o) B x B y B z B The training hyperparameters are shown in Table 1.

[0157] Table 1 Training Hyperparameters

[0158]

[0159] Training used MLP connection layers (512, 256, 128), 1024 environments, a maximum of 10,000 steps per epoch, and a maximum of 4,000 epochs. Training was performed using a laptop GPU 4060. The training results are as follows: Figure 17 As shown.

[0160] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the protection scope of the present invention.

Claims

1. A variable structure amphibious robot, characterized in that: The amphibious robot body includes a variable structure wing system, an electric control system, and a water-air amphibious robot body. The robot body includes a sealed cabin, a propeller fixing assembly, a fluid shell, and a variable support tail cover. The variable structure wing system includes a drive gear transmission assembly, a two-section folding wing, and a wing folding linkage assembly. The electric control system includes a control core element, a vision module, a power module, a GPS module, and an underwater depth gauge module.

2. The variable structure water-air amphibious robot according to claim 1, characterized in that: The body sealed cabin includes an electric control sealed cabin, a battery sealed cabin, and a transparent circular sealed cover plate.

3. The variable structure water-air amphibious robot according to claim 1, characterized in that: The propeller fixing assembly includes a polygonal carbon fiber plate-shaped support connecting piece, a duct fan fixing piece, and a duct fan propeller.

4. The variable structure water-air amphibious robot of claim 1, wherein: The fluid shell and the variable support tail cover include a ring-wing coupling fairing, a porous ring, and a support tail cover assembly.

5. The variable structure water-air amphibious robot of claim 1, wherein: The drive gear transmission assembly of the variable structure wing system includes a folding drive steering gear, a curved-tooth bevel gear transmission component, and a transmission shaft mounting bracket. The curved-tooth bevel gear transmission component includes two curved-tooth bevel gears for realizing the horizontal direction of the wing in the unfolded state and the vertical direction of the wing in the folded state. The drive gear transmission assembly has a corresponding mirror image assembly along the vertical plane of the electric control sealed cabin.

6. The variable structure water-air amphibious robot of claim 1, wherein: The two-section folding wing of the variable structure wing system comprises an inner wing component and an outer wing component, which are combined into a hinge structure, connected by a pin and flange bearing to form a rotating pair, the rotating shaft is perpendicular to the wing windward section and is offset downward by a shaft radius unit distance; the inner wing component and the outer wing component can be spliced into a seamless complete wing in the unfolded state, the complete wing is forward-swept for providing lift during flight; the outer wing component is folded 180 degrees towards the inside of the wing and is parallelly folded inside the inner wing component in the folded state; the inner wing component has two cylindrical hollow grooves inside for the transmission rope of the folding linkage to pass through; the two-section folding wing has corresponding mirror image components along the vertical plane of the electrically controlled sealed cabin.

7. The variable structure water-air amphibious robot of claim 1, wherein: The wing linkage folding transmission of the variable structure wing system comprises a transmission rope and a rotating shaft spring, one end of the transmission rope is connected to the outer wing component, the other end is connected to the transmission shaft mounting bracket through the cylindrical hollow groove of the inner wing, continuously crossing the wing rotating shaft and the inner and outer wing rotating shaft; the rotating shaft spring is used for the rotation of the inner and outer wing, realizing the initial position of the superposition state; the variable structure wing system pulls the transmission rope at the same time when the inner wing is unfolded, so that the outer wing component is unfolded at the same time, and the transmission rope is released when the inner wing is folded, and the outer wing component is restored to the superposition state by the elastic force of the rotating shaft spring; the wing linkage folding transmission has corresponding mirror image components along the vertical plane of the electrically controlled sealed cabin.

8. The variable structure water-air amphibious robot of claim 1, wherein: The control core element is used for controlling the action of the robot, comprising an on-board computer deploying a control algorithm in an upper position and a flight control hardware interacting with the information of components such as propeller, GPS module and underwater depth gauge module in a lower position; the vision module is used for collecting image information of the surrounding environment; the power module is used for supplying power to each electrical device of the robot; the GPS module is used for collecting position information of the robot body; the underwater depth gauge module is used for collecting depth information when moving in water.

9. A method for sports using the variable structure amphibious robot according to any one of claims 1-8, characterized in that, The method comprises the following steps: S1. The robot vertically takes off from the horizontal plane in the four-rotor mode in the wing folding state; S2. After the robot flies at a certain height, the wings are unfolded to convert to horizontal flight state, and the robot flies in the air in the fixed-wing mode, the propeller provides flight power, yaw and pitch moment; S3. After the robot reaches above the water surface, the head slides towards the water surface, and changes to the folded state before contacting the water surface, and then enters the water in the folded state; S4. After the robot enters the water, the propeller rotor runs in the forward and reverse directions at low speed, and the forward and reverse thrust and torque generated by the rotor are used as the power source for underwater movement of the robot, the wings remain folded in normal underwater movement, and the wings are unfolded at different angles when the speed needs to be reduced or water resistance is generated, and the robot moves underwater in the form of a complete underwater glider after the wings are completely unfolded; S5. The robot accelerates vertically to a certain water depth in the folded state, and relies on inertial force to impact out of the water surface, and when the propeller rotor completely separates from the water surface, the rotor switches to high-speed rotation mode to generate air thrust, and then the robot leaves the water surface in the four-rotor mode. S6. The wing of the robot is converted to an unfolded state after leaving the water surface, and fixed-wing horizontal direction cruise is performed; S7. The wing of the robot is converted to a folded state when approaching the ground, and four-rotor mode flight is performed; when the passive tail landing gear is unfolded after contacting the ground, vertical landing is completed.

10. A control method based on reinforcement learning for the motion of the variable structure amphibious robot according to any one of claims 1-8, characterized in that, The method comprises the following steps: S1. A dynamics model controlled by the robot motion is established, which is modeled as a 6-DOF rigid body driven by 4 motors; in order to realize large-scale training requirements, a lightweight operation dynamics simulation model equation is established as follows: where is the quaternion rotation, p WB is the position, q WB is the attitude, v WB is the velocity, ω B is the angular rate, g W is the gravity, f p is the thrust, f i is the air or underwater drag, τ p is the duct torque; S2. Establish single motor thrust model: f i = K f A t ; A t ∈ [0, 1]; coefficient K f value is obtained according to the test in different medium reality movement, which is a second order nonlinear curve function; S3. Based on the deep reinforcement learning PPO algorithm, the robot reinforcement learning control algorithm framework is established, and the loss function J(θ) and its gradient The formula is as follows: where, is the desired operator, R(τ) is the return, a t is the action, s t is the state, π θ is the action probability function; Add student imitation teacher strategy, use KL divergence criterion and knowledge distillation method for fusion training, and finally generate a new advantage function and objective function wherein, is the old advantage function, γ is a discount factor, is the knowledge distillation loss, r t is the probability ratio, clip() is a clipping function, and λ is a weight coefficient. Define the state input vector: S = [p x ,p y ,p z ,q w ,q x ,q y ,q z ,v x ,v y ,v z ,w x ,w y ,w z ] T , p x , p y , p z are positions, q w , q x , q y , q z are attitude quaternions, v x , v y , v z are linear velocities, and w x , w y , w z are angular velocities. Define output action: A = [a1, a2, a3, a4] T where a1, a2, a3, a4 represent the 4 ducted propeller thrust magnitudes; S4. The random domain method is used to enhance the control generalization ability, and the specific randomization equation is expressed as follows: where M flow is the hydrodynamic disturbance function, is the IMU attitude noise, is the altimeter noise, M is the mass float, I is the inertia float, F d is the random external force, F air is the aerodynamic external force, T fail is the thrust failure vector; S5. Build a reinforcement learning simulation platform in Isaac Gym environment for parallel training of multiple Agents, with the target position as an additional state input: S h = [p hx , p hy , p hz ] T ; set the reward function R, aiming to guide the robot to reach the target position, the total value R is defined as follows: R = r p + r p (r l + r w ) + r a + r s , where r p is the position reward, r l is the verticality reward, r w is the rotation suppression reward, r a is the collision penalty, and r s is the contact target reward; the parameters of the MLP multi-full connection layer used in training are (512, 256, 128), and a lightweight neural network model is obtained after training; S6. A lightweight neural network model is run in the on-board computer board of the control core element of the robot, wherein the state input / output is interacted with the topic information of ROS, Mavros is used to interact with the lower computer for propeller thrust mapping and sensor interaction, and finally the reinforcement learning controller of the robot is deployed.

Images