Air-ground amphibious robot autonomous navigation method and control system

Abstract

The invention discloses an air-ground amphibious robot autonomous navigation method and control system, and the method comprises the steps: collecting images, including an RGB image and a depth image in the same scene; generating a flat terrain mask graph M of the RGB image; fusing the flat terrain mask map M and the depth image, and carrying out environment scene discrimination, including a land flat area, a low obstacle area and an air crossing area; according to an environment scene judgment result and a global target position in the flat terrain mask graph M, determining a local target point advancing to a global target for a short time; in various environment scenes, a motion skill control strategy for the robot to go to a local target point and finally arrive at a global target point is determined, and execution control of the motion skill of the air-ground amphibious robot is performed based on a control system until an automatic navigation task is completed. The navigation efficiency, the control stability and the energy consumption management of the air-ground amphibious robot in a complex and changeable scene are improved by fusing the sensing judgment environment and dynamically adjusting the path and the skill.

Description

Technical Field

[0001] The present invention relates to the technical field of autonomous navigation of mobile robots, and in particular to an autonomous navigation method and control system for a land and air amphibious robot. Background Art

[0002] As multimodal mobile platforms, amphibious robots combine the ground locomotion of wheeled, tracked, or multi-legged structures with the flight capabilities of quadcopters or multi-rotor systems. These robots can flexibly switch between locomotion techniques on flat ground, in obstacle-ridden areas, and in the air, enabling autonomous navigation and operation across terrain. However, despite the gradual maturity of their hardware architecture, current amphibious robots still face several key challenges in intelligent navigation. These include the lack of efficient and reliable environmental understanding mechanisms in actual tasks, making it difficult to rely solely on images to determine the current environmental scenario (fully traversable, partially traversable, or completely impassable). Motion control strategies often lose control due to inaccurate modeling in complex scenarios. Path planning methods often rely on a single motion model and lack comprehensive consideration of multi-skill scheduling and energy optimization, making it impossible to ensure a balance between mission success rate and energy consumption for the robot.

[0003] Environmental perception remains a key challenge hindering the advancement of mobile robot perception systems. This is particularly true in complex, dynamic, and diverse real-world scenarios. Accurately identifying and understanding the surrounding environment is crucial for tasks such as path planning, obstacle avoidance, and multimodal behavior switching. Existing methods based on RGB image classification have become a mainstream approach for building visual perception modules due to their simple structure and fast computational speed. These methods typically employ end-to-end classification training of environmental images using convolutional neural networks to achieve recognition of different scenes or object categories. However, these methods still have significant limitations. RGB images inherently contain a large amount of redundant information, resulting in lengthy model training and high computational resource consumption, particularly when processing high-resolution images. Furthermore, relying solely on the two-dimensional pixel distribution of RGB images lacks an effective representation of the scene's spatial structure and depth, making them susceptible to recognition errors in the presence of occlusion, varying illumination, or interference from multiple objects. This compromises the robustness and accuracy of overall environmental perception.

[0004] In mobile robot navigation systems, current navigation methods rely on establishing an accurate global map and performing global path planning based on this map to obtain the optimal path from the starting point to the destination. These methods offer high navigation accuracy and path interpretability, provided the environment is stable and information is complete. They can provide clear and continuous motion instructions to the robot and are widely used in relatively static scenarios such as indoor service and industrial logistics. However, in practical deployments, building and maintaining a global map is computationally intensive, requiring high computing power and storage resources, making it difficult to adapt to resource-constrained mobile platforms. Furthermore, when the external environment undergoes dynamic changes (such as obstacle movement or structural changes), the system must rebuild the map and replan the path, severely impacting the real-time and continuity of the task. Furthermore, these methods are highly dependent on positioning accuracy. Accumulating positioning errors or degrading sensor accuracy can lead to failure of global path planning.

[0005] In summary, current amphibious robots still face multiple technical bottlenecks when performing multimodal tasks, including unstable perception, heavy navigation burdens, and insufficient intelligent mode switching. Systematic breakthroughs are urgently needed to support their efficient and autonomous operation in complex environments. Regarding environmental perception, existing visual perception methods struggle to effectively capture the spatial structure and depth information of a scene. This is particularly true under non-ideal conditions such as occlusion, varying illumination, or multiple target interference. This can lead to recognition bias, which in turn reduces the robustness and accuracy of the perception system and prevents it from providing accurate environmental understanding for subsequent mission decision-making and path planning. Regarding navigation, current mainstream methods rely on global map construction and path optimization. While these methods offer high accuracy in static scenes, their high computational overhead and resource requirements limit their widespread application in dynamically changing environments and on resource-constrained platforms. More critically, current amphibious robots still rely primarily on manual judgment when switching between motor skills, lacking effective intelligent decision-making mechanisms. This makes it difficult to achieve adaptive switching based on mission requirements and environmental conditions, resulting in a dynamic trade-off between performance and energy consumption during mission execution. These problems jointly restrict the further improvement of the intelligence level of amphibious robots on land and in the air, and there is an urgent need to achieve deep integration and innovation in perception modeling, lightweight navigation and autonomous decision-making mechanisms. Summary of the Invention

[0006] Purpose of the invention: The purpose of the present invention is to provide an autonomous navigation method and control system for an amphibious robot to enhance the autonomous navigation capability of the amphibious robot in complex and changing environments.

[0007] Technical solution: To achieve the above-mentioned purpose, the present invention provides an autonomous navigation method for an amphibious robot, comprising the following steps:

[0008] Step 1: Collect images, including RGB images and depth images of the same scene;

[0009] Step 2: Generate a flat terrain mask M of the RGB image;

[0010] Step 3: Fuse the flat terrain mask M with the depth image and perform environmental scene discrimination on the fused image, including "flat land area", "low obstacle area" and "air crossing area";

[0011] Step 4: Based on the environmental scenario discrimination results and the global target position in the flat terrain mask M, determine the local target point for short-term advancement to the global target;

[0012] Step 5: Under various environmental scenarios, determine the robot's motion skill control strategy for moving to the local target point and ultimately reaching the global target point.

[0013] Preferably, the method of fusing the flat terrain mask image M with the depth image and performing environmental scene discrimination on the fused image is: fusing the flat terrain mask image M with the depth image Fusion in the channel dimension forms a dual-channel image The dual-channel image I f The image is input into the improved EfficientNet-B0 classification network to obtain the environmental scene labels of the image: "flat land area", "low obstacle area" and "air crossing area".

[0014] Preferably, the method for determining the local target point for short-term advancement toward the global target based on the environmental scenario discrimination result and the global target position in the flat terrain mask map M is:

[0015] In the flat land area scenario, the traversable area A is extracted from the flat terrain mask map M. flat , in the passable area A flat Select the local target point, expressed as:

[0016]

[0017] Among them, d(p,x r ) represents the candidate local target point p and the robot's current position x r The Euclidean distance between sub (p) is the target point reachability function;

[0018] In the low obstacle area scenario, the low obstacle boundary is extracted from the flat terrain mask M At the edge of a low obstacle Search for a flat and passable area A on one side fall , in the flat and accessible area A fall The best landing point after the robot leaps is selected as the local target point, which is expressed as:

[0019]

[0020] in, Represents the distance between the candidate local target point p and the obstacle boundary, Reachable PostJump (·) is the target point reachability function;

[0021] In the aerial crossing area scenario, the robot takes off and heads towards the global target x g During the flight, the robot continuously determines whether there is a safe landing point P in the visible area. land , as a local target point, if there is no safe landing point P land Then continue to fly, and the local target point is expressed as:

[0022]

[0023] Among them, d(p,x g ) represents the candidate landing point p to the global target position x g The function VisibleFromAir(·) determines whether the candidate landing point p is within the field of view of the flight trajectory and meets the altitude and terrain safety requirements.

[0024] Preferably, the robot's movement skill control strategy for moving to a local target point and ultimately reaching a global target point includes:

[0025] Initialize the robot's movement skills, including walking on land, leaping on land, and flying in the air;

[0026] When the robot is heading towards the global target, it continuously determines whether it has reached the local target and the global target. When heading towards the local target, the robot continuously determines whether it can reach the local target with the current movement skill, or whether there are any obstacles to the current movement skill, until it reaches the global target.

[0027] If the robot cannot reach the local target point under the current motion skill, or the current motion skill is blocked and cannot reach the local target point, the current local target point is canceled, the posture is readjusted to face the global target, and steps 1 to 5 are repeated.

[0028] Preferably, the method for implementing the motor skill control strategy is:

[0029] Step 501: If the initial environment scenario is a flat land area, the land forward strategy is invoked; if it is a low obstacle area, the land leap strategy is invoked; if it is an air crossing area, the air flight strategy is invoked;

[0030] Step 502: While the robot is moving toward the global target, it continuously determines whether it has reached both the local target and the global target. If it has not reached the local target, it continues moving toward the local target using the current motor skill. If it has reached the local target but not the global target, it continues to determine the current environment and the next local target to move toward. If it has reached the global target, the robot completes autonomous navigation.

[0031] Step 503: When traveling to the local target point, the robot determines whether it can reach the local target point under the current movement skill, or whether there is an obstacle to the current movement skill. The judgment function is expressed as:

[0032]

[0033] Among them, P LTP is the local target point. When the movement skill is land forward, P LTP =P sub ; When the sports skill is land leap, P LTP =P jump ; When the movement skill is air crossing, P LTP =P land , if there is no P land , then Switch=0; Reachable(·) represents the target point reachability function, Skill current Indicates motor skills;

[0034] When Switch = 1, it indicates that the robot needs to switch sports skills; when Switch = 0, it indicates that the robot does not need to switch sports skills;

[0035] Step 504: After the motor skill is switched, step 503 is repeated until Switch = 0, and then step 502 is repeated until the robot's autonomous navigation is completed.

[0036] Preferably, the TD3 algorithm is used to generate optimal actions for the three sports skills of land forward movement, land leap, and air flight in corresponding environmental scenarios, including: designing the state space, action space, and reward function of the sports skills, training the TD3 neural network, and using the trained TD3 neural network to obtain the optimal action for the robot to perform in the current sports state based on the current environmental scenario;

[0037] Among them, the TD3 neural network includes an Actor network and two Critic networks At each time step, the Actor network collects a sample quadruple (s, a, r, s') from the interaction with the environment and outputs the currently executed action a t , and calculate the execution action a t The reward value rt ; Continue to interact with the environment and accumulate rewards until the interaction ends and the total reward value is obtained; during several rounds of interaction, the TD3 neural network is continuously updated according to the previous training experience and the designed loss function until the training is completed.

[0038] Preferably, the state space of the land forward movement skill is represented as:

[0039]

[0040] in, is a set of 16-dimensional state vectors, including the minimum values ​​of the 16 sectors of the lidar; S tar_flat It is a set of 2D state vectors, including the Euclidean distance Δd and the relative deviation angle Δθ between the robot’s current position and the local target point coordinates; S vehicle is a set of 2-dimensional state vectors, including the robot's current linear velocity v l and angular velocity ω;

[0041] The state space of the land leap skill is represented as:

[0042]

[0043] in, It is a set of 20-dimensional state vectors, which includes dividing the depth image into regions of equal area and taking the minimum value of the depth information of each region as the input of the vector; S tar_low It is a set of 3D state vectors, including the relative distances Δx, Δy, Δz of the robot relative to the local target point in the three coordinate axes of the world coordinate system; S body_low It is a set of 6-dimensional vectors containing the robot's own state information, including the robot's velocity v in the world coordinate system x ,v y ,v z , pitch angular velocity ω pitch , rolling angular velocity ω roll And the throttle value T throttle ;

[0044] The state space of the aerial flight skill is represented as:

[0045]

[0046] in, is a set of 16-dimensional state vectors from the lidar, S tar_flight It is a set of 2D state vectors, including the relative distances Δx and Δy of the robot relative to the local target point in the two coordinate axis directions in the world coordinate system; S body_flightIt is the status information of the robot itself, including the speed v of the robot in the world coordinate system x ,v y ,v z , pitch angular velocity ω pitch , yaw angular velocity ω yaw And the throttle value T throttle .

[0047] Preferably, the motion space of the land forward movement skill is represented as:

[0048] A1=[a v ,a ω ]

[0049] Among them, the motion control variables are: linear speed a v ∈[0,2]m / s and angular velocity a ω ∈[-π,π]rad / s;

[0050] The motion space of the land leap skill is expressed as: A2=[a p ,a r ,a T ]

[0051] Among them, the action control variables are: pitch angle rate a p ∈[-π,π]rad / s, roll angular rate a r ∈[-π,π]rad / s and throttle value a T ∈[0,1];

[0052] The action space of the aerial flying skill is expressed as:

[0053] A3=[a ap ,a v ,a y ]

[0054] Among them, the action control variables are: pitch angle a ap ∈[-π / 2,π / 2]rad, corresponding to the linear velocity a in the land forward movement skill v , heading angular velocity a y ∈[-π,π]rad / s.

[0055] Preferably, the reward function of the land forward movement skill is expressed as:

[0056]

[0057] in, For line speed bonus, is the angular velocity reward, For obstacle distance rewards, For collision rewards, For target rewards,

[0058] The reward function of the land leap skill is expressed as:

[0059]

[0060] in, is the X-axis distance reward, D is the distance between the robot and the local target point in the X-axis direction at the initial moment, and Δx is the distance between the robot and the local target point in the X-axis direction; is the Y-axis distance reward, Δy is the distance between the robot and the local target point in the Y-axis direction; For border rewards, For collision rewards, For obstacle distance rewards, For target rewards,

[0061] The reward function of the aerial flying skill is expressed as:

[0062]

[0063] in, For the pitch angle bonus, is the heading angular rate reward, For collision rewards, For obstacle distance rewards, For target rewards,

[0064] The control system described in the present invention, which is applied to the above-mentioned land-air amphibious robot autonomous navigation method, includes an onboard computer, a flight controller, a motor driver, an electronic speed controller (ESC), a battery, a power module, a depth camera, a UWB positioning module, and a laser radar. The battery directly powers the power module and the ESC, which converts the voltage into 12V to power the onboard computer and the motor driver, and the ESC powers the flight controller. The flight controller, motor driver, depth camera, UWB positioning, and laser radar are connected to the onboard computer via a USB hardware interface for data transmission.

[0065] For land forward movement skills, the onboard computer perceives the environment through lidar and UWB modules, and outputs action instructions to the motor driver to make the robot perform corresponding actions; for land flying skills, the onboard computer perceives the environment through depth cameras and UWB modules, and outputs action instructions to the flight controller, which controls the electronic speed controller to perform corresponding actions; for flight control skills, the onboard computer perceives the environment through lidar and UWB modules, and outputs action instructions to the flight controller, which controls the electronic speed controller to perform corresponding actions.

[0066] Beneficial effects: The present invention has the following advantages: 1. The present invention uses fusion perception to perform environmental scenario discrimination, can understand complex and changeable environmental information in real time, and improves the accuracy and real-time performance of environmental perception of land and air amphibious robots; 2. Based on the environmental scenario discrimination results, the present invention dynamically adjusts the position generation of local target points and the calling strategy of motion skills, which not only improves the efficiency of path planning, reduces the ineffective movement of land and air amphibious robots, and achieves more reasonable energy consumption management, but also effectively avoids control conflicts caused by unintelligent skill switching in scenarios with diversified task spaces, thereby improving the navigation efficiency of land and air amphibious robots while ensuring the stability of control; 3. It is suitable for operation scenarios with dense obstacles, irregular terrain, and diversified task spaces. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 Schematic diagram of the process of this method;

[0068] Figure 2 This is the network structure diagram of EfficientNet-B0 with channel-level fusion;

[0069] Figure 3 A flowchart for classifying environmental scenarios from the robot's perspective;

[0070] Figure 4 Schematic diagram of the robot's local target point estimation in three environmental scenarios;

[0071] Figure 5 A map showing the correspondence between robot skills and environmental scenarios;

[0072] Figure 6 This is the network structure diagram of the robot skill algorithm;

[0073] Figure 7 (a) and (b) are schematic diagrams of the robot's lidar sector division and the relative position of the target point;

[0074] Figure 8 (a) and (b) are schematic diagrams of the robot's depth map area division and the relative position of the target point;

[0075] Figure 9(a) to (c) are simulation training environment diagrams for the three skills of the land and air amphibious robot;

[0076] Figure 10 Electrical connection diagram for the robot control execution system;

[0077] Figure 11 This is a diagram of the autonomous navigation system of the land and air amphibious robot;

[0078] Figure 12 This is the main control flow chart of the autonomous navigation system. DETAILED DESCRIPTION

[0079] The technical solution of the present invention is described in detail below with reference to the embodiments and drawings.

[0080] like Figure 1 As shown, the autonomous navigation method of the land-air amphibious robot of the present invention includes the following contents:

[0081] Step 1: Collect images, including RGB images and depth images of the same scene.

[0082] Step 2: Generate a flat terrain mask M of the RGB image.

[0083] The flat terrain mask is generated by using the improved lightweight terrain semantic segmentation network Effi-UNet to perform pixel-level semantic segmentation on the collected RGB image, extract the flat terrain area in the image, and generate a flat terrain mask M∈{0,1} H×W , H×W is the size of the image.

[0084] The improved lightweight terrain semantic segmentation network, Effi-UNet, is derived from the existing U-Net network by pruning, introducing a Ghost module, and adding a bottleneck module. Effi-UNet achieves excellent inference efficiency without significantly sacrificing accuracy, making it particularly suitable for deployment in scenarios where inference speed is critical.

[0085] Step 3: Fuse the flat terrain mask M with the depth image and perform environmental scene discrimination on the fused image.

[0086] Combine the flat terrain mask M with the depth image Fusion in the channel dimension forms a dual-channel image The dual-channel image I f The image is input into the improved EfficientNet-B0 classification network to obtain the environmental scene label c∈{1,2,3} of the image, which respectively represents the three scenes of "flat land area", "low obstacle area" and "air crossing area".

[0087] like Figure 2As shown, in order to be compatible with channel expansion, the present invention performs channel adaptation reconstruction on the first layer of EfficientNet-B0 and retains its efficient MBConv structure, and finally outputs the environment scene label, as shown in Figure 3 shown.

[0088] The classification function form of the EfficientNet-B0 classification network is defined as follows:

[0089]

[0090] Among them, f EffB0 represents the adapted EfficientNet-B0 classifier, [M,D] represents the fused image after channel-level splicing, and P(*) represents the predicted probability.

[0091] Combine the flat terrain mask M with the depth image Fusion in the channel dimension preserves the spatial geometric features of the depth image and the semantic guidance information of the mask map, achieves information complementarity, and improves the ability to recognize complex environmental scenarios.

[0092] Step 4: Based on the environmental scenario discrimination results and the global target position in the flat terrain mask M, determine the local target point for short-term advancement to the global target.

[0093] like Figure 3 As shown, in the flat land area scenario, the traversable area A is extracted from the flat terrain mask map M. flat , the robot is in the traversable area A flat When driving in the system, the local target point is expressed as:

[0094]

[0095] Among them, d(p,x r ) represents the candidate local target point p and the robot's current position x r The Euclidean distance between the two, function Reachable sub (p) Determine whether point p is reachable within the current perceptual map and motor capabilities.

[0096] In the low obstacle area scenario, the robot needs to use the land leap strategy to cross the low obstacles. Therefore, the boundaries of the low obstacles in the flat terrain mask M are first identified by semantic segmentation and edge detection technology. At the edge of the robot's ability to fly over low obstacles Search for a flat and passable area A on one side fall , in the flat and accessible area A fall The best landing point after the robot leaps is selected as the local target point, which is expressed as:

[0097]

[0098] in, Represents the distance between the candidate local target point p and the obstacle boundary, which is used to control the land leap amplitude; function Reachable PostJump (·) Determine whether the local target point p is within the area where it can safely land and continue moving forward (e.g. Figure 4 As shown in the low obstacle area, the low obstacle divides the mask image into two parts, with the obstacle as the boundary. The upper part is the area for continued movement. Due to the physical space size of the robot itself, it is safe to leave a certain distance behind the obstacle.

[0099] In the aerial crossing area scenario, the robot continues to move towards the global target x after taking off. g Flight. As the robot moves towards the global target point, it continuously determines whether there is a safe landing point P in the visible area. land , as a local target point, if there is no safe landing point P land Then continue to fly, and the local target point is expressed as:

[0100]

[0101] Among them, d(p,x g ) represents the candidate landing point p to the global target position x g The distance between the candidate landing point and the global target is given priority; the function VisibleFromAir(·) determines whether the candidate landing point p is within the field of view of the flight trajectory and meets the height and terrain safety requirements.

[0102] The landing point search process is as follows: During flight, the robot continuously collects RGB images and converts them into flat terrain mask images through step 2, and determines in real time whether there is a passable area A on the land in the image. flat , and in A flat Find a landing spot.

[0103] Step 5: Determine the robot's motion skill switching strategy to reach the global target point in various environmental scenarios.

[0104] Step 501: Initialize the robot's motion skills, such as Figure 5 As shown, if the initial judgment environment scenario is a flat land area, the land forward strategy is called; if it is a low obstacle area, the land leap strategy is called; if it is an air crossing area, the air flight strategy is called.

[0105] Step 502: While the robot is moving toward the global target point, it continuously determines whether it has reached the local target point and the global target point. If it has not reached the local target point, it continues to move toward the local target point under the current motor skill. If it has reached the local target point but has not reached the global target point, it continues to determine the current environmental scenario and determine the next local target point and move toward it. If it has reached the global target point, the robot's autonomous navigation is completed.

[0106] Step 503: When traveling to the local target point, the robot determines whether it can reach the local target point under the current movement skill, or whether there is an obstacle to the current movement skill. The judgment function is expressed as:

[0107]

[0108] Among them, P LTP is the local target point. When the movement skill is land forward, P LTP =P sub ; When the sports skill is land leap, P LTP =P jump ; When the movement skill is air crossing, P LTP =P land , if there is no P land , then Switch=0; Reachable(·) represents the target point reachability function, Skill current Indicates motor skills.

[0109] When Switch = 1, it indicates that the robot cannot reach the local target point under the current motion skill, or the current motion skill is blocked and cannot reach the local target point, then the current local target point is canceled, the posture is readjusted to face the global target, and steps 1 to 5 are re-executed.

[0110] When Switch = 0, it indicates that the robot can reach the local target point under the current motion skill, or the current motion skill is not blocked and can reach the local target point without switching the motion skill.

[0111] Step 504: After the motor skill is switched, step 503 is repeated until Switch = 0, and then step 502 is repeated until the robot's autonomous navigation is completed.

[0112] In order to enable the robot to successfully complete the autonomous navigation task, the present invention uses the TD3 algorithm to select the optimal actions for the three sports skills of land forward, land leap, and air flight. The specific contents are as follows:

[0113] 1. Design the state space of motor skills.

[0114] (1) The state space of land forward movement skill is expressed as:

[0115]

[0116] in, is a set of 16-dimensional vectors, such as Figure 7 (a) shows the minimum value of 16 sectors of the laser radar; S tar_flat is a set of 2-dimensional vectors, such as Figure 7 (b) shows the Euclidean distance Δd and the relative deviation angle Δθ between the robot’s current position and the local target point coordinates; S vehicle is a set of 2-dimensional vectors, including the robot's current linear velocity v l and angular velocity ω.

[0117] (2) The state space of the land leap skill is represented as:

[0118]

[0119] in, is a set of 20-dimensional vectors, such as Figure 8 As shown in (a), the depth image is divided into regions of equal area, and the minimum value of the depth information of each region is taken as the input of the vector; S tar_low is a set of 3-dimensional vectors, such as Figure 8 (b) shows the relative distances Δx, Δy, and Δz of the robot relative to the local target point in the three coordinate axes of the world coordinate system; S body_low It is a set of 6-dimensional vectors containing the robot's own state information, including the robot's velocity v in the world coordinate system x ,v y ,v z , pitch angular velocity ω pitch , rolling angular velocity ω roll And the throttle value T throttle .

[0120] (3) The state space of aerial flight skills is represented as:

[0121]

[0122] in, is a set of 16-dimensional vectors from the laser radar, S tar_flight is a set of 2-dimensional vectors. S body_flight It is the status information of the robot itself, including the speed v of the robot in the world coordinate system x ,v y ,v z , pitch angular velocity ω pitch , yaw angular velocity ω yaw And the throttle value T throttle .

[0123] 2. Design the action space for sports skills

[0124] (1) The robot uses wheels when moving on land, so the action space contains two variables, namely linear velocity a and v ∈[0,2]m / s and angular velocity a ω ∈[-π,π]rad / s. The motion space of the land forward movement skill is expressed as:

[0125] A1=[a v ,a ω ]

[0126] The target point is in front of the robot's current position. The linear velocity is designed to make the robot move forward and not backward, so the linear velocity is not negative. In order to reduce the deviation from the target point, the angular velocity is designed to make the robot deviate from the obstacle and approach the target point.

[0127] (2) The purpose of the robot's land leap is to fly over the obstacle in front of it and finally land on the flat area behind the obstacle. In order to make the robot's land leap trajectory more stable, the action space contains three control variables, namely, the pitch angle rate a p ∈[-π,π]rad / s, roll angular rate a r ∈[-π,π]rad / s and throttle value a T ∈[0,1]. The motion space of the land leap skill is expressed as:

[0128] A2=[a p ,a r ,a T ]

[0129] The pitch rate is designed to adjust the robot's pitch angle, thereby moving it forward. The roll rate is designed to offset the robot's left-right deviation, ensuring accurate Y-axis position in the world coordinate system. The throttle control controls the robot's upward motion to avoid collisions with obstacles. To move downward, the throttle is reduced, allowing gravity to control the movement.

[0130] (3) The action space of aerial flight is similar to the action space of land forward movement. The action space of aerial flight includes two control variables: the pitch angle a of the control body and the pitch angle α of the control body. ap ∈[-π / 2,π / 2]rad, which makes the robot move forward, corresponding to the linear velocity a in the land forward skill v ; By adjusting the robot's heading angular velocity a y ∈[-π,π]rad / s control, so that it can avoid obstacles in front, corresponding to the angular velocity a in the land forward skill ωThe action space of the aerial flying skill is represented as:

[0131] A3=[a ap ,a v ,a y ]

[0132] 3. Designing reward functions for motor skills

[0133] (1) The simulation training environment for land forward skills is as follows Figure 9 As shown in (a), the design of its reward function is mainly based on the characteristics of the robot's behavioral requirements and the interaction with the environment when performing the land forward task. The reward function of the land forward skill is expressed as:

[0134]

[0135] in, For line speed bonus, The reward for the robot moving towards the goal at a higher speed helps to improve the robot's forward speed; is the angular velocity reward, Excessive angular velocity may cause the robot to deviate from the target path, affecting the performance of the task. Therefore, penalizing unnecessary angular velocity can help the robot maintain a stable steering action. For obstacle distance rewards, When the robot is too close to an obstacle, a negative reward is given to encourage the robot to maintain a safe distance and avoid collision; For collision rewards, Its purpose is to encourage the robot to avoid obstacles through strong punishment; For target rewards, Represents the reward for the robot completing the task.

[0136] (2) The simulation training environment of land leaping skills is as follows Figure 9 As shown in (b), the reward function design mainly considers how the robot can complete the task efficiently and safely during the land leap. The reward function of the land leap skill is expressed as:

[0137]

[0138] in, is the X-axis distance reward, This is designed to encourage the robot to approach the target point. As the robot approaches the target point, the distance Δx in the X-axis direction gradually decreases, and the corresponding reward also increases. It is initially zero. is the Y-axis distance reward, Although the robot's Y-axis deviation is initially zero, it inevitably deviates during movement. To suppress this deviation, a negative reward is designed, and the value of the negative reward increases as the deviation increases. This design effectively reduces unnecessary lateral deviations during the robot's land leaps, maintaining mission accuracy. For border rewards, Flying out of the boundary means that the robot has left the mission area, which not only affects the completion of the mission but also may bring safety hazards. Therefore, a large negative reward is given to encourage the robot to always stay in the safe area; For collision rewards, When a collision occurs, a large negative reward is immediately generated to ensure that the robot avoids contact with obstacles as much as possible and ensures the smooth progress of the mission; For obstacle distance rewards, This design is intended to encourage the robot to maintain a safe distance from obstacles and reduce the risk of collision; For target rewards, This reward clarifies the mission goal and encourages the robot to reach the target point as quickly as possible.

[0139] (3) Simulation training environment for aerial flight skills such as Figure 9 As shown in (c), the design of its reward function is similar to that of the land forward skill. Five reward functions are set, including two positive rewards and three negative rewards. The reward function of the aerial flight skill is expressed as:

[0140]

[0141] in, For the pitch angle bonus, This is designed to encourage the robot to control its flight attitude by adjusting its pitch angle. A larger pitch angle allows the robot to approach the target quickly, so a positive reward is given. is the heading angular rate reward, In order to suppress excessive rotation, prevent the robot from producing unstable motion trajectory and ensure flight stability; For collision rewards, Its purpose is to encourage the robot to avoid obstacles through strong punishment; For obstacle distance rewards, When the robot is too close to an obstacle, a negative reward is given to encourage the robot to maintain a safe distance and avoid collision; For target rewards, Represents the reward for the robot completing the task.

[0142] 4. For the state spaces S1~S3, action spaces A1~A3 and reward functions R1~R3 ​​of three different sports skills, independent TD3 networks (TD3-1, TD3-2, TD3-3) are constructed and trained respectively.

[0143] (1) TD3 includes an Actor network and a dual Critic network, which are used to output the optimal action in each action space. Figure 6 As shown in Figure 1, the Actor Network in the TD3 algorithm uses a three-layer fully connected structure with 1000 and 800 neurons in the hidden layers, respectively, and an output layer dimension of 2. It is used to generate continuous actions to control the robot. The Actor Network outputs a deterministic action based on the current state and is a core component of the decision-making process.

[0144] The critic part consists of two Q networks with symmetrical structures and independent parameters, which are denoted as and Each critic network takes the concatenated vector of state and action as input, and passes through the transformed fully connected layer (TFC), combined fully connected layer (CFC) and output layer in sequence to output the corresponding action value estimates Q1 and Q2 respectively.

[0145] During training, the actor network proposes the optimal action, while the critic network evaluates it and outputs the expected reward for the current state. The TD3 algorithm employs a dual-critic architecture, minimizing the two Q values ​​in target value calculation to prevent overestimation of action value, thereby providing a more reliable and stable optimization signal for the actor. The actor and critic networks collaborate through alternating optimization to achieve continuous improvement and convergence of the strategy.

[0146] (2) For different motor skills training tasks, three state spaces S1, S2, and S3 are divided, corresponding to three action spaces A1, A2, and A3 respectively. To this end, three independent TD3 networks (TD3-1, TD3-2, and TD3-3) are designed and trained in parallel. Each network consists of an Actor network and two Critic networks. To learn the optimal strategy from the corresponding state space to the action space. During the training process, the strategy update process is carried out according to the following logic:

[0147] 1) Experience sampling stage

[0148] At each time step, the Actor network collects a sample quadruple (s, a, r, s') from the interaction with the environment, where the state s∈S i,i=1,2,3; execute action a∈A i , the continuous action output by the current policy network; r is the execution action according to the reward function R i , i=1,2,3 calculated reward value, s'∈S i For the next state.

[0149] 2) Target Q value calculation

[0150] The target policy network generates the target action:

[0151]

[0152] Among them, μ φ' (s') The action given by the target policy network in the next state s';

[0153] The dual critic network calculates the target Q value and uses the minimum value strategy to suppress overestimation bias:

[0154]

[0155] r is the reward function output, which ensures that the learning objectives of different models are consistent but the content is differentiated; γ is the discount factor, which controls the importance of future rewards in the current decision, 0<γ<1;

[0156] 3) Critic Network Update

[0157] Each critic network optimizes its parameters independently, with the goal of making its Q value close to the target Q value mentioned above:

[0158]

[0159] (s t ,a t ) is the state space S from the corresponding motor skill i With action space A i The state and action of the network ensure the consistency and stability of value estimation.

[0160] 4) Delayed Updates to the Actor Network

[0161] To improve the stability of policy updates, TD3 uses a delayed update mechanism: after every d = 2 critic network updates, the actor network is updated once. The actor's goal is to maximize the Q value estimated by the first critic network:

[0162]

[0163] Here the state s∈S i,i=1,2,3, ensure that the policy update is consistent with the environment state corresponding to the task; represents the policy network parameters Find the gradient; Indicates finding the gradient of the action variable a.

[0164] 5) Target network soft update

[0165] All networks use a soft update mechanism to smoothly converge to the main network and improve training stability:

[0166] θ′ j ←τθ j +(1-τ)θ′ j ,φ′←τφ+(1-τ)φ′

[0167] θ j is the parameter of the jth Critic network; φ represents the parameter of the Actor network; θ' j , φ' represents the corresponding target network parameters; τ is the soft update factor, which controls the speed at which the target network parameters follow the main network.

[0168] The autonomous navigation system of the amphibious robot of the present invention is electrically connected as follows Figure 10 As shown in the figure, the battery directly powers the ESC, which is then converted to 12V by the power module to power the onboard computer and motor driver. The onboard computer receives data from the camera, UWB module, and LiDAR via a USB port to sense the surrounding environment. The onboard computer then sends control commands to the flight controller and motor controller, which in turn provide feedback on attitude and other information.

[0169] The autonomous navigation system for amphibious robots is designed for intelligent robot platforms with dual-mode motion capabilities on the ground and in the air. It is suitable for efficient path planning and motion control tasks in complex environments, such as Figure 11 shown.

[0170] The main program flow of the autonomous navigation task is as follows Figure 12As shown in the figure, after the system starts, it waits for input of the destination coordinates. Once input is received, the system begins to adjust its posture, turning itself toward the destination. Once the posture adjustment is complete, environmental perception begins. The camera acquires RGB images, and a semantic segmentation network outputs a flat terrain mask. This, combined with the depth image, is fed into the environmental classification network to determine the current environmental scenario category of the amphibious robot. If the scenario is identified as a flat land area or an area with low obstacles, the flat terrain mask is used to calculate the local target point. Once the local target point is determined, the corresponding skill planning strategy is activated. Sensor data is input to generate corresponding action commands and movement is initiated. During this process, each action is determined to determine whether the destination and local target point have been reached. If the destination is reached, the system returns to the beginning of the program and re-enters the destination coordinates. If the destination is not reached but the local target point is reached, the system restarts the current scenario assessment. If the local target point is not reached, the next action command of the skill planning strategy is executed. When the scenario is identified as an aerial crossing area, the robot takes off vertically and flies towards the destination after stabilizing. During the flight, it continuously determines whether it has reached the destination and observes the ground conditions to see if the conditions for landing are met. If not, it continues to maintain the flight mode. If so, it lands on the ground and restarts the scenario identification.

Claims

1. A method for autonomous navigation of an amphibious robot, characterized in that: The following steps are involved: Step 1: Collect images, including RGB images and depth images of the same scene; Step 2: Generate a flat terrain mask M of the RGB image; Step 3: Fuse the flat terrain mask M with the depth image and perform environmental scene discrimination on the fused image, including "flat land area", "low obstacle area" and "air crossing area"; Step 4: Based on the environmental scenario discrimination results and the global target position in the flat terrain mask M, determine the local target point for short-term advancement to the global target; Step 5: Under various environmental scenarios, determine the robot's motion skill control strategy for moving to the local target point and ultimately reaching the global target point.

2. The autonomous navigation method according to claim 1, characterized in that: The method of fusing the flat terrain mask image M with the depth image and performing environmental scene discrimination on the fused image is as follows: fusing the flat terrain mask image M with the depth image Fusion in the channel dimension forms a dual-channel image The dual-channel image I f The image is input into the improved EfficientNet-B0 classification network to obtain the environmental scene labels of the image: "flat land area", "low obstacle area" and "air crossing area".

3. The autonomous navigation method according to claim 1, wherein: The method for determining the local target point for short-term advancement toward the global target based on the environmental scenario discrimination result and the global target position in the flat terrain mask M is as follows: In the flat land area scenario, the traversable area A is extracted from the flat terrain mask map M. flat , in the passable area A flat Select the local target point, expressed as: Among them, d(p,x r ) represents the candidate local target point p and the robot's current position x r The Euclidean distance between sub (p) is the target point reachability function; In the low obstacle area scenario, the low obstacle boundary B is extracted from the flat terrain mask M. obs , at the boundary B of the low obstacle obs Search for a flat and passable area A on one side fall , in the flat and accessible area A fall The best landing point after the robot leaps is selected as the local target point, which is expressed as: in, Represents the distance between the candidate local target point p and the obstacle boundary, Reachable PostJump (·) is the target point reachability function; In the aerial crossing area scenario, the robot takes off and heads towards the global target x g During the flight, the robot continuously determines whether there is a safe landing point P in the visible area. land , as a local target point, if there is no safe landing point P land Then continue to fly, and the local target point is expressed as: Among them, d(p,x g ) represents the candidate landing point p to the global target position x g The function VisibleFromAir(·) determines whether the candidate landing point p is within the field of view of the flight trajectory and meets the altitude and terrain safety requirements.

4. The autonomous navigation method according to claim 1, wherein: The robot's motion skill control strategy for moving to a local target point and ultimately reaching a global target point includes: Initialize the robot's movement skills, including walking on land, leaping on land, and flying in the air; When the robot is heading towards the global target, it continuously determines whether it has reached the local target and the global target. When heading towards the local target, the robot continuously determines whether it can reach the local target with the current movement skill, or whether there are any obstacles to the current movement skill, until it reaches the global target. If the robot cannot reach the local target point under the current motion skill, or the current motion skill is blocked and cannot reach the local target point, the current local target point is canceled, the posture is readjusted to face the global target, and steps 1 to 5 are repeated.

5. The autonomous navigation method according to claim 4, characterized in that: The implementation method is: Step 501: If the initial environment scenario is a flat land area, the land forward strategy is invoked; if it is a low obstacle area, the land leap strategy is invoked; if it is an air crossing area, the air flight strategy is invoked; Step 502: While the robot is moving toward the global target, it continuously determines whether it has reached both the local target and the global target. If it has not reached the local target, it continues moving toward the local target using the current motor skill. If it has reached the local target but not the global target, it continues to determine the current environment and the next local target to move toward. If it has reached the global target, the robot completes autonomous navigation. Step 503: When traveling to the local target point, the robot determines whether it can reach the local target point under the current movement skill, or whether there is an obstacle to the current movement skill. The judgment function is expressed as: Among them, P LTP is the local target point. When the movement skill is land forward, P LTP =P sub ; When the sports skill is land leap, P LTP =P jump ; When the movement skill is air crossing, P LTP =P land , if there is no P land , then Switch=0; Reachable(·) represents the target point reachability function, Skill current Indicates motor skills; When Switch = 1, it indicates that the robot needs to switch sports skills; when Switch = 0, it indicates that the robot does not need to switch sports skills; Step 504: After the motor skill is switched, step 503 is repeated until Switch = 0, and then step 502 is repeated until the robot's autonomous navigation is completed.

6. The autonomous navigation method according to claim 4, characterized in that: The TD3 algorithm is used to generate optimal actions for the three motor skills of land forward movement, land leap, and air flight in corresponding environmental scenarios. This includes designing the state space, action space, and reward function of the motor skills, training the TD3 neural network, and using the trained TD3 neural network to obtain the optimal action for the robot to perform in the current motion state based on the current environmental scenario. Among them, the TD3 neural network includes an Actor network and two Critic networks At each time step, the Actor network collects a sample quadruple (s, a, r, s) from the interaction with the environment. ' ), output the currently executed action a t , and calculate the execution action a t The reward value r t ; Continue to interact with the environment and accumulate rewards until the interaction ends and the total reward value is obtained; during several rounds of interaction, the TD3 neural network is continuously updated according to the previous training experience and the designed loss function until the training is completed.

7. The autonomous navigation method according to claim 6, characterized in that: The state space of the land forward movement skill is represented as: in, is a set of 16-dimensional state vectors, including the minimum values ​​of the 16 sectors of the lidar; S tar_flat It is a set of 2D state vectors, including the Euclidean distance Δd and the relative deviation angle Δθ between the robot’s current position and the local target point coordinates; S vehicle is a set of 2-dimensional state vectors, including the robot's current linear velocity v l and angular velocity ω; The state space of the land leap skill is represented as: Among them, S envdep It is a set of 20-dimensional state vectors, which includes dividing the depth image into regions of equal area and taking the minimum value of the depth information of each region as the input of the vector; S tar_low It is a set of 3D state vectors, including the relative distances Δx, Δy, Δz of the robot relative to the local target point in the three coordinate axes of the world coordinate system; S body_low It is a set of 6-dimensional vectors containing the robot's own state information, including the robot's velocity v in the world coordinate system x ,v y ,v z , pitch angular velocity ω pitch , rolling angular velocity ω roll And the throttle value T throttle ; The state space of the aerial flight skill is represented as: in, is a set of 16-dimensional state vectors from the lidar, S tar_flight It is a set of 2D state vectors, including the relative distances Δx and Δy of the robot relative to the local target point in the two coordinate axis directions in the world coordinate system; S body_flight It is the status information of the robot itself, including the speed v of the robot in the world coordinate system x ,v y ,v z , pitch angular velocity ω pitch , yaw angular velocity ω yaw And the throttle value T throttle .

8. The autonomous navigation method according to claim 6, characterized in that: The motion space of the land forward movement skill is expressed as: A1=[a v ,a ω ] Among them, the motion control variables are: linear speed a v ∈[0,2]m / s and angular velocity a ω ∈[-π,π]rad / s; The motion space of the land leap skill is expressed as: A2=[a p ,a r ,a T ] Among them, the action control variables are: pitch angle rate a p ∈[-π,π]rad / s, roll angular rate a r ∈[-π,π]rad / s and throttle value a T ∈[0,1]; The action space of the aerial flying skill is expressed as: A3=[a ap ,a v ,a y ] Among them, the action control variables are: pitch angle a ap ∈[-π / 2,π / 2]rad, corresponding to the linear velocity a in the land forward movement skill v , heading angular velocity a y ∈[-π,π]rad / s.

9. The autonomous navigation method according to claim 8, characterized in that: The reward function of the land forward skill is expressed as: in, For line speed bonus, is the angular velocity reward, For obstacle distance rewards, For collision rewards, For target rewards, The reward function of the land leap skill is expressed as: in, is the X-axis distance reward, D is the distance between the robot and the local target point in the X-axis direction at the initial moment, and Δx is the distance between the robot and the local target point in the X-axis direction; is the Y-axis distance reward, is the distance between the robot and the local target point in the Y-axis direction; For border rewards, For collision rewards, For obstacle distance rewards, For target rewards, The reward function of the aerial flying skill is expressed as: in, For the pitch angle bonus, is the heading angular rate reward, For collision rewards, For obstacle distance rewards, For target rewards, 10. A control system applied to the autonomous navigation method of the amphibious robot according to claim 1, characterized in that: It includes an onboard computer, flight controller, motor driver, ESC, battery, power module, depth camera, UWB positioning module, and lidar. The battery directly powers the power module and ESC, which converts the voltage into 12V to power the onboard computer and motor driver, and the ESC powers the flight controller. The flight controller, motor driver, depth camera, UWB positioning, and lidar are connected to the onboard computer via a USB hardware interface for data transmission. For land forward movement skills, the onboard computer perceives the environment through lidar and UWB modules, and outputs action instructions to the motor driver to make the robot perform corresponding actions; for land flying skills, the onboard computer perceives the environment through depth cameras and UWB modules, and outputs action instructions to the flight controller, which controls the electronic speed controller to perform corresponding actions; for flight control skills, the onboard computer perceives the environment through lidar and UWB modules, and outputs action instructions to the flight controller, which controls the electronic speed controller to perform corresponding actions.

Images