Foot type bionic robot cluster collaborative hunting system based on wolf pack visual system characteristics

Abstract

A foot-type bionic robot cluster collaborative hunting system based on wolf pack visual system features constructs a bionic visual sensor array by simulating wide-angle vision and a multi-mode perception mechanism of a wolf pack, carries out perception data collaborative fusion, and combines reinforcement learning and an Apollonius circle theory to design a hunting strategy. According to the method, efficient tracking and autonomous surrounding of a moving target by a foot type robot cluster in a complex dynamic environment are achieved, meanwhile, dependence on a single sensor is reduced, the formation cooperation response speed and the obstacle avoidance capacity are effectively improved, and technical support is provided for multi-robot cooperation in the scenes of military reconnaissance, disaster rescue and the like.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of bionic robot cluster cooperative hunting system based on the characteristics of wolf vision system. BACKGROUND

[0002] The current vision system of the legged robot mainly uses fixed field of view camera or narrow angle perception module, which lacks effective capture ability for wide area dynamic target, especially in low light or night environment, the performance of traditional optical sensor is significantly reduced, while in nature, the wolf pack can efficiently track the movement trajectory of the prey through wide angle vision and group cooperation strategy, but this biological feature has not been fully used in the field of legged robots. And the cluster cooperative hunting technology mainly relies on preset path planning to achieve target tracking, which has not been deeply combined with the motion control of legged robot and the design of bionic sensor, resulting in a significant gap between algorithm theory and engineering practice.

[0003] In summary, the existing technology has three major problems: insufficient adaptability to dynamic environment, limited visual perception and low cooperative efficiency, which seriously restricts the autonomous hunting ability of the legged robot cluster in complex scenes. SUMMARY

[0004] The present application provides a bionic robot cluster cooperative hunting system based on the characteristics of wolf vision system, which constructs a multi-camera cooperative system through bionic wolf vision characteristics, realizes wide area environmental perception, low light target tracking and multi-robot cooperation ability, and enhances the dynamic interaction and cluster hunting ability of the robot.

[0005] The bionic robot cluster cooperative hunting system based on the characteristics of wolf vision system provided by the present application mainly includes:

[0006] The perception module of the legged bionic robot includes:

[0007] The bionic binocular stereo vision sub-module installed at the front end of the robot head is used to simulate the forward stereo vision characteristics of the wolf pack, as well as the color blindness characteristics of the wolf and the high sensitivity to moving objects;

[0008] The wide area lateral vision enhancement sub-module installed on both sides of the robot trunk simulates the peripheral vision ability and detail recognition characteristics of the wolf, which is used to obtain lateral environmental information and broadcast it to other members of the cluster;

[0009] The infrared night vision perception sub-module installed at the top center of the robot simulates the night vision ability of the wolf, which is used to identify the body temperature target through the heat radiation characteristics; and the integrated near infrared fill light is used to enhance the texture detail capture of the binocular stereo vision sub-module in dark environment;

[0010] The rear auxiliary perception submodule, installed on the back of the robot, is used to supplement the blind spot in the rear view and provide obstacle warnings for the rear members of the cluster.

[0011] The multimodal sensing data collaborative fusion module adopts a fusion framework based on a multi-scale feature pyramid network to fuse the data features acquired by each sensing sub-module, capture irregular moving targets, and realize wide-area environmental perception.

[0012] The cluster collaborative capture control module adopts a multi-legged robot capture strategy based on the Apollonius circle and reinforcement learning algorithm to divide the target's advantageous area and control the legged bionic robots to cooperate in capturing the target.

[0013] Furthermore, the bionic binocular stereo vision submodule includes:

[0014] The front-facing binocular camera is installed at the front of the robot's head. The horizontal distance between the binocular cameras is 8-10cm, and the downward tilt angle with the robot's horizontal plane is 15° to 20°. The horizontal field of view of each camera is 120°, the resolution is not less than 1080P, and the frame rate is not less than 60 frames per second.

[0015] Furthermore, the wide-area lateral vision enhancement submodule includes:

[0016] Wide-angle cameras are symmetrically deployed on both sides of the robot's torso, with their optical axes kept horizontal and their height aligned with the forward-facing binocular module. The horizontal field of view of each camera is ≥90°, forming a 280° horizontal field of view coverage with the forward-facing binocular camera.

[0017] The camera generates an environmental map using a real-time image stitching algorithm and broadcasts the lateral environmental information to other members of the group.

[0018] Furthermore, the infrared night vision sensing submodule includes: a long-wave infrared camera installed at the center of the top of the robot, with a wavelength range of 8-14μm, a field of view of 60° to 90°, and an optical axis tilted forward at 15° to cover the area directly in front and to the upper side.

[0019] Furthermore, the rearward auxiliary perception submodule includes: a 120° wide-angle camera mounted on the back of the robot, with its optical axis tilted 15° downwards and backwards, and its vertical field of view covering an area from the ground to 1.2 times the height of the robot.

[0020] Furthermore, the rearward auxiliary perception submodule extracts rearward terrain features in real time using a lightweight semantic segmentation model, and combines this with positioning data to provide obstacle warnings for rearward members of the group; wherein:

[0021] The lightweight semantic segmentation model uses weighted cross-entropy loss as the semantic segmentation loss function to enhance obstacle region recognition.

[0022]

[0023] Among them, w c For class weights, y c For real labels, p c To predict the probability, C is the number of channels;

[0024] The semantic segmentation result is converted from pixel coordinates to a global coordinate system, realizing the mapping from image pixels to geographic coordinates:

[0025]

[0026] Where R(θ) is the robot's heading angle rotation matrix, (u,v) are the obstacle pixel coordinates, Z is the obstacle depth, f is the camera focal length, (x UWB ,y UWB ) are the coordinates for UWB positioning.

[0027] Furthermore, the multimodal perception data collaborative fusion module extracts low-level features in the shallow network; adaptively focuses on the moving target region through deformable convolution to capture irregular moving targets; and extracts high-level semantic features in the deep network to improve the robustness of target recognition, wherein:

[0028] (1) The deformable convolution is achieved through an offset Δp k Adaptively adjust the sampling position to capture irregularly moving targets:

[0029] Δp k =W offset *F in

[0030] Among them, W offset F represents the convolution kernel parameters of the offset prediction layer. in The input is a shallow feature map, where K is the kernel size, and 2K offsets are learned at each position.

[0031] The deformable convolution output is:

[0032]

[0033] Where p0 is the center position of the convolution kernel, p k Δp represents the relative position of the k-th sampling point within the convolution kernel. k For learnable offsets, w k Here, X represents the convolution kernel weights, and X represents the input feature map.

[0034] (2) The deep network uses a deep convolutional network to extract semantic features, as shown below:

[0035] F RGB =φ CNN (IRGB W CNN )

[0036] Among them, I RGB For the input RGB image, W CNN φ represents the weight parameters of the convolutional network. CNN This is the feature extraction function for convolutional networks.

[0037] Furthermore, the cluster-coordinated capture control module divides the target's advantageous region based on the concept of the Apollonius circle, specifically including the following methods:

[0038] Assume the speed of the i-th legged bionic robot Greater than the speed V of the moving target T Then the speed ratio satisfy The trajectory is the Apollonius circle, with the i-th legged bionic robot acting as the pursuer located outside the circle and the target located inside the circle, where O is the center of the circle, (x O ,y O () represents the coordinates of the center of the circle, and r represents the radius. Let x be the coordinates of the pursuer, (x T ,y T () represents the coordinates of the escapee;

[0039] The i-th legged bionic robot, Apollonius, circle S i It is expressed as follows:

[0040] (xx O ) 2 +(yy O ) 2 =r 2

[0041] in,

[0042]

[0043] According to the definition of the Apollonian circle, when At that time, the area within the Apollonius circle is the advantageous zone for the moving target.

[0044] Furthermore, the cluster-cooperative encirclement control module employs a multi-legged robot encirclement strategy based on reinforcement learning algorithms to determine the state space, action space, and reward function of the legged bionic robot, wherein:

[0045] (1) State space

[0046] Define global state The states of all legged bionic robots and the target are included, as follows:

[0047]

[0048] in Let be the position of the i-th legged bionic robot; Let x be the velocity of the i-th legged bionic robot; T ,y T ),V T These represent the position and velocity of the moving target, respectively; n is the number of agents.

[0049] (2) Action space

[0050] The model at a fixed speed, the movements of each legged bionic robot The direction angle is as follows:

[0051] a i =[θ i ];

[0052] (3) Reward function

[0053] For legged bionic robots, the goal is to minimize the dominant region of the moving target. When the dominant region is sufficiently small, the encirclement mission is considered successful, and the dominant region is represented by S. a express:

[0054]

[0055] The reward function for the legged bionic robot is then set as follows:

[0056] Compared with the prior art, the beneficial effects of this disclosure are:

[0057] (1) Construction of bionic vision sensor array: Based on the five characteristics of the wolf pack vision system (peripheral vision ability, motion sensitivity ability, night vision ability, color blindness and poor detail discrimination), a composite sensor module including binocular camera, infrared thermal imaging, etc. is constructed to realize wide-angle environment modeling.

[0058] (2) Multimodal perception data collaborative fusion: Reduce the dependence on a single sensor (such as GPS) and realize the efficient detection capability of legged robot swarms of moving targets in complex dynamic environments;

[0059] (3) The rear auxiliary perception submodule on the back not only fills the blind spot in the rear field of vision, but also provides obstacle warnings for the rear members of the cluster, improving the cluster's collaborative perception capability.

[0060] (4) Encirclement strategy integrating reinforcement learning and Apollonius circle: An encirclement domain generation algorithm based on dynamic Apollonius circle is proposed. A decision model is constructed by combining a deep reinforcement learning framework. The motion trajectory of the robot group is optimized in real time through reinforcement learning algorithm, so as to realize the adaptive shrinkage of the encirclement radius with the target speed.

[0061] (5) Effectively improve the collaborative response speed and obstacle avoidance capability of robot clusters, and enhance autonomous detection and capture performance in complex dynamic scenarios;

[0062] (6) It provides technical support for multi-robot collaborative tasks in fields such as military reconnaissance and disaster relief. Attached Figure Description

[0063] The above and other objects, features and advantages of this disclosure will become more apparent from the more detailed description of exemplary embodiments of this disclosure taken in conjunction with the accompanying drawings, in which the same reference numerals generally represent the same components.

[0064] Figure 1 This is a general block diagram of a swarm of legged bionic robots for cooperative trapping, according to the present disclosure.

[0065] Figure 2 This is a diagram showing the sensor layout structure.

[0066] Figure 3 For Apollonius's circle;

[0067] Figure 4 This is the area of ​​advantage for the target movement. Detailed Implementation

[0068] Preferred embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While preferred embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that the present disclosure will be thorough and complete, and will fully convey the scope of the present disclosure to those skilled in the art.

[0069] To address the problems of poor dynamic environment adaptability, limited visual perception range, and low collaborative efficiency faced by existing legged robot swarms in collaborative hunting, this disclosure provides a collaborative hunting system for legged biomimetic robot swarms based on the characteristics of the wolf pack's visual system.

[0070] In one exemplary implementation, the overall system structure is as shown in the attached figure. Figure 1 As shown, it mainly includes: a bionic vision sensor array module, a multimodal perception data collaborative fusion module, and a legged robot encirclement control strategy module based on reinforcement learning and Apollonius circle design.

[0071] The construction method for each module is described in detail below:

[0072] 1. A biomimetic visual sensor array based on the visual characteristics of wolf packs

[0073] Wolf pack vision systems possess five key characteristics: peripheral vision, motion sensitivity, night vision, color blindness, and poor detail discrimination. In this embodiment, based on these wolf pack visual characteristics, a legged bionic robot perception system is constructed, and the layout of the bionic vision sensors is improved, as shown in the attached figure. Figure 2 As shown.

[0074] (1) Bionic binocular stereo vision module

[0075] A binocular camera module is installed at the front of the robot's head to simulate the forward stereoscopic vision of a wolf pack. The horizontal spacing between the binocular cameras is set to 8-10cm to simulate the eye spacing of a wolf, and they are tilted downwards at a 15°–20° angle to the robot's horizontal plane to optimize ground target capture. Each camera has a horizontal field of view of 120°, a resolution of no less than 1080P, and a frame rate of no less than 60 frames per second.

[0076] Target depth information is calculated using a binocular parallax algorithm, and combined with an improved YOLOv8 model, dynamic target detection and locking are achieved. The channel attention mechanism formula is as follows:

[0077]

[0078] Among them, w c For channel attention weights, F c Let C be the Cth channel of the input feature map, and σ be the Sigmoid activation function. In grayscale mode, the number of channels C = 1. The model automatically enhances the edge gradient features of moving targets and suppresses background texture details, which aligns with the colorblindness of wolves.

[0079] This module enhances the stereo matching accuracy of ground targets through binocular spacing and tilt angle design, while simulating the wolf's high sensitivity to moving objects. The optimized model for stereo matching accuracy of ground targets is shown below:

[0080]

[0081] Where ∈ represents the stereo matching error, H represents the camera's height above the ground, θ represents the tilt angle, B represents the horizontal distance between the two eyes, and k represents the calibration coefficient, which is determined by the lens distortion rate.

[0082] The camera uses a black-and-white / grayscale mode to simulate the colorblindness of wolves, retaining only the outline and depth information of moving targets while weakening detail recognition to reduce computational load. as follows:

[0083]

[0084] Where α is the algorithm complexity coefficient, and N pixel Where C is the number of image pixels and C is the number of channels. Grayscale conversion reduces C from 3 to 1, decreasing the computational load by 66.7%.

[0085] This module transforms the biological characteristics of wolves into quantifiable machine vision constraints, such as grayscale and attention weights, and ultimately achieves biomimetic optimization through mathematical modeling coupled with hardware parameters.

[0086] (2) Wide-area lateral vision enhancement module

[0087] Wide-angle cameras are symmetrically deployed on both sides of the robot's torso, with a single-lens horizontal field of view ≥90°, forming a 280° horizontal field of view coverage with the front binocular camera. The horizontal field of view superposition model is shown below:

[0088] θ total =θ f +2·(θ s -θ overlap )

[0089] Where, θ f θ represents the horizontal field of view of the front-facing binocular module. s For the horizontal field of view of the monocular wide-angle module, θ overlap θ represents the overlap angle between the lateral and forward visual fields. f =120°, θ s ≥90°, then θ overlap When the angle is less than 10°, it can achieve 280° horizontal coverage.

[0090] The camera is mounted close to the robot body, at the same height as the forward-facing binocular module, with its optical axis kept horizontal to avoid overlapping and interference in the field of view. An environmental map is generated using a real-time image stitching algorithm, and lateral environmental information is broadcast to other members of the group. This module simulates the peripheral vision capabilities of a wolf through its side-view layout, and employs motion detection preprocessing to highlight dynamic targets while downplaying static details, echoing the wolf's weak detail recognition characteristics. Inter-frame differencing is used to highlight dynamic targets.

[0091]

[0092] Among them, I t For the current frame image, I t-1 The previous frame is represented by T, which is a motion threshold dynamically adjusted based on lighting conditions. This is an indicator function that only retains the motion region.

[0093] (3) Integration of infrared night vision sensing module

[0094] A long-wave infrared camera with a wavelength range of 8-14 μm and a field of view of 60°–90° is mounted at the center of the top of the robot. Its optical axis is tilted forward by 15° to cover the area directly in front and to the upper side. This module performs pixel-level fusion with a visible light camera, enabling the identification of targets with body temperature (such as humans or animals) through thermal radiation characteristics under low-light conditions, simulating the night vision capability of a wolf's rod cells. The grayscale value I of the long-wave infrared camera... IR The relationship with the target temperature T can be expressed by the following formula:

[0095]

[0096] Where ∈ represents the target emissivity, and σ = 5.67 × 10 -8 W / (m 2 K 4 A represents the pixel area of ​​the target in the image, and d represents the distance between the robot and the target obtained through binocular ranging or depth estimation. This formula converts infrared grayscale values ​​into a temperature distribution map, highlighting high-temperature objects.

[0097] An integrated 850nm near-infrared illuminator works in conjunction with the night vision module to enhance texture detail capture in low-light environments while retaining grayscale output for thermal imaging, simulating the colorblindness of wolves. The relationship between the illumination intensity L of the near-infrared illuminator and the distance d is as follows:

[0098]

[0099] Where P is the power of the fill light and η is the luminous efficacy coefficient. The fill light enhances the texture details of the visible light camera, but the thermal imaging of the infrared camera relies solely on the target's own radiation. Therefore, the grayscale output is not affected by the fill light, which is consistent with the colorblindness of wolves.

[0100] Infrared and visible light data are weighted and fused to prioritize the preservation of the thermal radiation profile of moving targets. Visible light data I S With infrared data I IR The fusion formula is:

[0101] I fused =α·I S +(1-α)·I IR

[0102] in, α represents the dynamic weighting coefficient. M(x,y) is the motion detection mask, and k is the sensitivity adjustment. For moving target regions (M(x,y)>T), α approaches 0, prioritizing the preservation of infrared thermal radiation contours; for static regions, α approaches 1, preserving visible light texture details.

[0103] (4) Backward auxiliary sensing module

[0104] A 120° wide-angle camera is mounted on the robot's back, with its optical axis tilted 15° downwards and backwards. The vertical field of view covers an area from the ground to 1.2 times the robot's height. Adaptive exposure control technology, based on ambient light intensity feedback, addresses image overexposure issues in backlit scenes. Ambient light intensity L env With camera exposure time t exp The relationship is:

[0105]

[0106] Among them, t ref L is the baseline exposure time. th Where L is the exposure threshold and k is the adjustment coefficient. When the ambient light intensity L... env Exceeding threshold L tj In high light, the exposure time decreases rapidly to avoid overexposure; in low light, maintain a long exposure time.

[0107] A lightweight semantic segmentation model is used to extract rear terrain features in real time, and weighted cross-entropy loss is used as the semantic segmentation loss function to enhance obstacle region recognition.

[0108]

[0109] Among them, w c For class weights, y c For real labels, p c To predict probabilities.

[0110] By combining location data, obstacle warnings are provided to members behind the group, achieving a mapping from image pixels to geographic coordinates. The semantic segmentation results (pixel coordinates) are converted to a global coordinate system.

[0111]

[0112] Where R(θ) is the robot's heading angle rotation matrix, (u,v) are the obstacle pixel coordinates, Z is the obstacle depth, f is the camera focal length, (x UWB ,y UWB ) are the coordinates for UWB positioning.

[0113] This module supplements the wolf's rear field of vision blind spots with a wide-angle design and uses low-resolution and motion noise reduction algorithms to reduce the reliance on detail recognition.

[0114] 2. Collaborative Fusion of Multimodal Sensing Data

[0115] In this embodiment, wide-area environmental perception is achieved through the collaborative fusion of biomimetic visual sensor layout and multimodal perception data. A fusion framework based on a multi-scale feature pyramid network (FPN) is constructed, taking feature maps from visible light, infrared, and depth data as input. Low-level features such as edges and textures are extracted in the shallow network, and deformable convolution adaptively focuses on the moving target region. The deformable convolution uses an offset Δp. k Adaptively adjust the sampling position to capture irregularly moving targets, as detailed below:

[0116] Δp k =W offset *F in

[0117] Among them, W offset F represents the convolution kernel parameters of the offset prediction layer. in The input is a shallow feature map, K is the kernel size, and 2K offsets are learned at each position.

[0118] The deformable convolution output is as follows:

[0119]

[0120] Where p0 is the center position of the convolution kernel, p k Δp represents the relative position of the k-th sampling point within the convolution kernel. k For learnable offsets, w k Let be the kernel weights and X be the input feature map. Deformable convolution uses an offset Δp. k It adaptively adjusts the sampling position to capture irregularly moving targets.

[0121] To improve the robustness of target recognition, high-level semantic features are extracted from deep networks. Specifically, deep convolutional networks are used to extract semantic features, as shown below:

[0122] F RGB =φ CNN (I RGB W CNN )

[0123] Among them, I RGB For the input RGB image, W CNN φ represents the weight parameters of the convolutional network. CNN This is the feature extraction function for convolutional networks.

[0124] 3. Cluster-based coordinated encirclement and capture

[0125] In this embodiment, a reinforcement learning-based multi-legged robot encirclement strategy is designed for a moving target. This strategy includes:

[0126] 3.1 By applying the concept of the Apollonius circle, the target's advantageous region is divided. Then, the legged bionic robot continuously reduces the advantageous region of the moving target, thereby ensuring the cooperative capture of the target within a limited time.

[0127] For details, see attached. Figure 3 As shown, assume the speed of the i-th legged bionic robot Greater than the speed V of the moving target T Then the speed ratio satisfy The trajectory of this is the Apollonius circle. For example... Figure 3 As shown, the i-th legged bionic robot, acting as the pursuer, is located outside the circle, while the moving target is located inside the circle. O is the center of the circle, (x O ,y O ) represents the coordinates of the center of the circle, and r represents the radius. The coordinates of the pursuer, (x T ,y T () represents the coordinates of the escapee.

[0128] The i-th legged bionic robot, Apollonius, circle S i It is expressed as follows:

[0129] (xx O ) 2 +(yy O ) 2 =r 2

[0130] in

[0131]

[0132] According to the definition of the Apollonian circle, when At that time, the area within the Apollonius circle is the advantageous region for the moving target, specifically as follows: Figure 4 The gray area is shown in the image.

[0133] 3.2 A trapping strategy is designed using reinforcement learning algorithms to determine the state space, action space, and reward function of the legged bionic robot.

[0134] (1) State space

[0135] Define global state The states of all legged bionic robots and the target are included, as follows:

[0136]

[0137] in Let be the position of the i-th legged bionic robot. Let x be the velocity of the i-th legged bionic robot. T,y T ),V T denoted by , where and represent the position and velocity of the moving target, respectively. represents the number of agents.

[0138] (2) Action space

[0139] The motion of the i-th legged bionic robot at a fixed speed. The direction angle θ i The details are as follows:

[0140] a i =[θ i ]

[0141] (3) Reward function

[0142] For legged bionic robots, the goal is to minimize the dominant region of the moving target. When the dominant region is sufficiently small, the capture mission is considered successful. The dominant region is represented by S... a Specifically, it can be represented as follows:

[0143]

[0144] Among them, S i S represents the dominant area of ​​the moving target under the influence of the i-th legged bionic robot. a Let represent the intersection of the advantageous areas of a moving target under the influence of n legged bionic robots. Then, the reward function for the legged bionic robots is set as follows:

[0145] Regarding the calculation of the intersection of multiple dominant regions, when the number of legged robots is less than 3, an analytical method can be used to solve the problem. The intersection points are represented by A(x1,y1), B(x2,y2), and C(x3,y3), as follows:

[0146]

[0147] Among them, A poly The area of ​​the polygon enclosed by the intersection points is calculated as follows:

[0148]

[0149] A seg,i The formula for the area of ​​an arc is as follows:

[0150]

[0151] Taking the area of ​​the arc between points A and B as an example,

[0152] When the number of legged robots is greater than 3, the Monte Carlo method is used to generate random points within a rectangular area that includes all the circles. The proportion of points falling within the intersection of all the circles is counted, and multiplied by the area of ​​the rectangle to obtain the approximate area.

[0153]

[0154] Where N is the total number of sampling points, M is the number of points falling within the intersection of all circles, and S area The area is the sampling region.

[0155] In this embodiment, by simulating the wide-angle vision and multimodal perception mechanism of wolf packs, a biomimetic vision sensor array is constructed and perception data is collaboratively fused. Combining reinforcement learning and Apollonius circle theory, a capture strategy is designed to enable robot swarms to efficiently track and autonomously capture moving targets in complex dynamic environments. At the same time, it reduces the dependence on a single sensor (such as GPS), improves the formation's collaborative response speed and obstacle avoidance capabilities, and provides multi-robot collaborative technology support for scenarios such as military reconnaissance and disaster relief.

[0156] The above technical solutions are merely exemplary embodiments of the present invention. For those skilled in the art, based on the application methods and principles disclosed in the present invention, it is easy to make various types of improvements or modifications, and not limited to the methods described in the specific embodiments of the present invention. Therefore, the methods described above are merely preferred and not restrictive.

Claims

1. A kind of based on the feature of wolf vision system Foot-type bionic robot cluster cooperative hunting system, it is characterized in, The method comprises the following steps: A bionic visual perception module, including a sensor array installed at multiple positions of a legged robot and facing multiple directions; A multi-modal perception data collaborative fusion module adopts a fusion framework based on a multi-scale feature pyramid network to fuse data features obtained by each sensor array in the perception module, capture irregular moving targets, and realize wide-area environment perception; A cluster collaborative hunting control module adopts a multi-legged robot hunting strategy based on Apollonius circles and reinforcement learning algorithm to divide the advantage area of the target and control the legged bionic robot to cooperatively hunt the target.

2. The system of claim 1, wherein, The bionic visual perception module comprises: A bionic binocular stereo vision sub-module installed at the front end of the head of the robot, which simulates the forward stereo vision characteristics of a wolf pack, while simulating the color blindness characteristics of a wolf and the high sensitivity to moving objects; A wide-area lateral vision enhancement sub-module installed on both sides of the torso of the robot, which simulates the peripheral vision ability and detail recognition characteristics of a wolf, is used to obtain lateral environmental information, and broadcasts it to other members of the cluster; An infrared night vision perception sub-module installed at the central top of the robot, which simulates the night vision ability of a wolf, is used to identify body temperature targets through thermal radiation characteristics; and an integrated near-infrared fill light is used to enhance the texture detail capture of the binocular stereo vision sub-module in dark environments; A rear auxiliary perception sub-module installed on the back of the robot is used to supplement the rear visual field blind area and provide obstacle warning for the rear members of the cluster.

3. The system of claim 2, wherein, The bionic binocular stereo vision sub-module comprises: A forward binocular camera installed at the front end of the head of the robot, with a horizontal distance of 8-10 cm between the two cameras and a downward inclination angle of 15°-20° with respect to the horizontal plane of the robot; a monocular horizontal field of view of 120°, a resolution of not less than 1080P, and a frame rate of not less than 60 frames per second.

4. The system of claim 2, wherein, The wide-area lateral vision enhancement sub-module comprises: Wide-angle cameras symmetrically deployed on both sides of the torso of the robot, with horizontal optical axes and a height level with the forward binocular module; a monocular horizontal field of view of ≥90°, forming a horizontal field of view coverage of 280° with the forward binocular camera; The cameras generate an environment map through real-time image stitching algorithm and broadcast lateral environmental information to other members of the group.

5. The system of claim 2, wherein, The infrared night vision perception sub-module comprises: a long-wave infrared camera installed at the central top of the robot, with a wavelength range of 8-14 μm and a field of view angle of 60°-90°, and the optical axis is inclined forward by 15° to cover the front and side areas.

6. The system of claim 2, wherein, The rear auxiliary perception sub-module comprises: a 120° wide-angle camera installed on the back of the robot, with the optical axis inclined downward by 15°, and the vertical field of view angle covers the area from the ground to 1.2 times the height of the robot.

7. The system of claim 2, wherein, The rear auxiliary perception sub-module extracts the rear terrain features in real time through a lightweight semantic segmentation model, and provides obstacle warning for the rear members of the group in combination with positioning data; wherein: The lightweight semantic segmentation model adopts a weighted cross-entropy loss as the semantic segmentation loss function to strengthen the identification of obstacle areas: where w c is the class weight, y c is the true label, p c is the predicted probability, and C is the number of classes. The semantic segmentation result is converted from pixel coordinates to a global coordinate system to realize the mapping from image pixels to geographic coordinates: Wherein, R(θ) is the robot heading angle rotation matrix, (u, v) is the obstacle pixel coordinates, Z is the obstacle depth, f is the camera focal length, (x UWB ,y UWB ) is the UWB positioning coordinates.

8. The system of any one of claims 1-7, wherein, The multi-modal perception data collaborative fusion module extracts low-level features in the shallow network; The adaptive focusing motion target area is captured by deformable convolution, and the high-level semantic features are extracted by a deep network to improve the robustness of target recognition. (1) the deformable convolution is through an offset Δp k Adaptive adjustment of sampling position to capture irregularly moving targets: Δp k = W offset * F in wherein W offset is a convolution kernel parameter of the offset prediction layer, F in is an input shallow feature map, K is a convolution kernel size, and 2K offsets are learned at each position. The deformable convolution output is: wherein p0 is the center position of the convolution kernel, p k is the relative position of the kth sampling point in the convolution kernel, Δp k is the learnable offset, w k is the convolution kernel weight, and X is the input feature map; The deep network adopts a deep convolution network to extract semantic features, as shown in the following formula: F RGB = φ CNN (I RGB ; W CNN ) where I RGB is the input RGB image, W CNN are the convolutional network weight parameters, and φ CNN is the convolutional network feature extraction function.

9. The system of claim 1, wherein, The cluster cooperative hunting control module divides the advantage area of the target based on the concept of Apollonian circle correlation, and the specific method includes: Assume the i th foot-type bionic robot velocity is greater than the moving target velocity V T , then the velocity ratio satisfies The trajectory that satisfies the above condition is Apollonian circle, the i th foot-type bionic robot as the pursuer is located outside the circle, the moving target is located inside the circle, O is the center of the circle, (x O ,y O ) is the center coordinate, r is the radius, is the coordinate of the pursuer, (x T ,y T ) is the coordinate of the escapee; Apollonian circle S of the ith legged robot i is represented as follows: (x - x O ) 2 (y - y O ) 2 = r 2 wherein According to the definition of Apollonius circle, when the inside of the Apollonius circle is the dominant region of the moving target.

10. The system of claim 1, wherein, The cluster cooperative hunting control module adopts a multi-legged robot hunting strategy based on a reinforcement learning algorithm to determine the state space, action space and reward function of the legged bionic robot, wherein: (1) State space Define global state Contains the state of all legged robots and targets, as follows: wherein is the position of the ith legged robot; is the velocity of the ith legged robot;(x T ,y T ), V T are the position and velocity of the moving target, respectively; n is the number of agents; (2) Action space Fixed speed die, each foot type bionic robot action For the direction angle, as follows: a i = [θ i ]; (3) Reward function For a legged robot, the goal is to minimize the advantage region of the moving target, and the pursuit task is considered successful when the advantage region is small enough. The advantage region is denoted by S a denotes the advantage region of the moving target. The reward function of the legged robot is set as:

Images