An integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration.

Abstract

This invention discloses an integrated system for heterogeneous robot allocation, networking, and planning in underground space exploration, belonging to the field of robot system technology. It includes a perception layer for acquiring environmental information through lidar, visual sensors, and inertial navigation systems; a decision-making layer comprising a task allocation module, a path planning module, and a collision self-recovery module; and a communication layer for forming a green and energy-efficient self-organizing network communication using an adaptive event-triggered mechanism. The adaptive event-triggered mechanism evaluates the necessity of communication based on a meta-relational attention network. The task allocation module employs top-level distributed negotiation and bottom-level global allocation. The path planning module uses a brain-like learning strategy that integrates meta-relational attention networks and reinforcement learning. The collision self-recovery module performs collision detection based on the meta-relational attention network and executes the negotiation and recovery mechanism. This invention enables efficient collaborative exploration of heterogeneous robots in uncertain underground environments, effectively improving exploration efficiency and system robustness in extremely uncertain environments.

Description

Technical Field

[0001] This invention belongs to the field of robot system technology, specifically relating to an integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration. Background Technology

[0002] With the advancement of robotics technology, multi-robot systems have shown broad application prospects in fields such as underground space exploration, disaster relief, and military reconnaissance. Current technologies typically employ centralized or fully distributed task allocation strategies, relying on pre-defined communication topologies for formation, and combining traditional global or local path planning algorithms (such as the A* algorithm and dynamic window method) to complete exploration tasks.

[0003] However, these existing technical solutions have significant shortcomings when facing highly uncertain and unstructured underground environments. First, task allocation methods are mostly based on deterministic assumptions of complete environmental information, lacking online learning and adaptive optimization capabilities. This makes it difficult to effectively match heterogeneous robot capabilities and dynamically adjust tasks under dynamic uncertainties such as communication interruptions and missing prior maps. Second, existing topology formation and networking strategies are usually static or only consider connectivity, ignoring the differences in energy state, communication energy consumption, and mobility among heterogeneous robot nodes. This makes it difficult to maintain reliable and energy-efficient network connections under resource-constrained conditions. Finally, traditional path planning algorithms lack real-time obstacle avoidance and replanning capabilities when facing sudden obstacles and complex terrain. Furthermore, a single robot failure can easily lead to the interruption of the entire exploration mission, resulting in poor system robustness.

[0004] Therefore, there is an urgent need to develop an integrated system for heterogeneous robot allocation, networking, and planning for underground space exploration that can comprehensively solve the above problems, so as to achieve integrated functions of efficient collaborative exploration of heterogeneous robots, intelligent task allocation, dynamic communication networking, and adaptive path planning. Summary of the Invention

[0005] This invention addresses the aforementioned problems and overcomes the shortcomings of existing technologies by providing an integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration. This invention enables efficient collaborative exploration of heterogeneous robots in uncertain underground environments. It is applicable to exploration scenarios with limited communication and complex terrain, such as post-disaster ruins and military bunkers. The aim is to effectively improve exploration efficiency and system robustness in extremely uncertain environments through the complementary capabilities and intelligent collaboration of heterogeneous robots.

[0006] To achieve the above objectives, the present invention adopts the following technical solution.

[0007] This invention provides an integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration, comprising: The perception layer is used to acquire environmental information through lidar, visual sensors, and inertial navigation systems. The decision-making layer, connected to the perception layer, includes a task allocation module, a path planning module, and a collision self-recovery module; The communication layer connects the perception layer and the decision layer, and is used to form a green and energy-saving self-organizing network communication with the help of an adaptive event triggering mechanism. The adaptive event triggering mechanism evaluates the necessity of communication based on a meta-relational attention network, which is a multi-layered network comprising an input layer, a capability encoding layer, a relational attention layer, a role reasoning layer, and an output layer. The task allocation module adopts a two-layer architecture of top-level distributed negotiation and bottom-level global allocation. The path planning module adopts a brain-like learning strategy that integrates the meta-relational attention network and reinforcement learning. The collision self-recovery module performs collision detection and executes a negotiation recovery mechanism based on the meta-relational attention network.

[0008] As a preferred embodiment of the present invention, the perception layer includes a LiDAR, a visual sensor, and an inertial navigation system that are configured and fused together; The lidar is used to acquire the three-dimensional geometric structure and depth information of the environment; the vision sensor is used to acquire the texture, color and semantic information of the environment; and the inertial navigation system is used to acquire the robot's body motion acceleration, angular velocity and attitude information. The lidar point cloud data and visual sensor image data are spatiotemporally registered and fused to form an environmental perception result that combines geometric and semantic information. The inertial navigation system provides time synchronization and motion compensation for the data fusion.

[0009] As another preferred embodiment of the present invention, the task allocation module is used to perform the initial allocation at the lower level through the meta-relational attention network, and to complete the final task allocation by performing a hybrid allocation method according to the top-level distributed negotiation strategy.

[0010] As another preferred embodiment of the present invention, the adaptive event triggering mechanism of the communication layer specifically includes: analyzing the robot's state, environmental information and network quality data through the meta-relational attention network, assessing the urgency of communication and making triggering decisions to dynamically adjust the network topology.

[0011] As another preferred embodiment of the present invention, the underlying initial allocation strategy is specifically as follows: using the meta-relational attention network, the robot is allocated as a central command robot, a reconnaissance robot, or a relay robot according to the robot's mobility, sensor range, and load capacity.

[0012] As another preferred embodiment of the present invention, the hybrid allocation method adopted by the top-level distributed negotiation strategy is a distributed negotiation method based on dynamic alliance and reputation game, including a dynamic alliance formation unit, a reputation game negotiation unit, an intelligent concession strategy unit, and a reputation evolution unit.

[0013] As another preferred embodiment of the present invention, the collision self-recovery module specifically includes: The collision detection unit uses the meta-relational attention network to classify collision states into mild jamming, moderate entrapment, or severe malfunction based on IMU abnormal acceleration, current sensor load abrupt change, and depth camera visual collision recognition information. The negotiation recovery unit is used to call the recovery strategy library according to the collision status, and to form a temporary rescue team with the health robot to carry out dynamic rescue based on the force feedback collaborative drag strategy.

[0014] As another preferred embodiment of the present invention, the meta-relational attention network includes: The input layer is used to receive sensor configurations, robot body parameters, and environmental characteristics. Capability coding layer, used to encode the input to obtain capability coding vector; The relational attention layer is used to calculate the relational strength between the ability vector and the character prototype through a multi-head attention mechanism, and obtain the relational attention matrix. The role inference layer is used for team-level feature propagation and role probability calculation based on the relational attention matrix and graph convolutional network. The output layer is used to output the decision results.

[0015] As another preferred embodiment of the present invention, the dynamic alliance forming unit forms a temporary alliance based on task requirements and robot capability complementarity, and the alliance value is obtained by weighted calculation of capability matching degree, execution cost and synergistic gain. The reputation game negotiation unit conducts multiple rounds of negotiation based on the reputation-ability matrix and historical performance. The intelligent concession strategy unit performs multiple rounds of proposal-feedback iterative adjustments based on the consensus level. The reputation evolution unit dynamically updates the robot's reputation value based on the negotiation results and task performance.

[0016] As another preferred embodiment of the present invention, the capability coding layer is specifically implemented through the following steps: S1. Encode the sensor configuration using a two-layer fully connected network and perform layer normalization. S2. The encoded sensor features, robot body parameters and environmental features are spliced ​​together; S3. A multidimensional capability encoding vector is obtained by mapping the fully connected network to the activation function.

[0017] Beneficial effects of this invention: This invention provides an integrated system for heterogeneous robot allocation and networking planning in underground space exploration. By constructing a three-layer architecture of perception, decision-making, and communication, and integrating a two-layer task allocation module based on a meta-relational attention network, a brain-like learning path planning module, a collision self-recovery module, and an adaptive event-triggered communication mechanism, it achieves systematic optimization of heterogeneous robot collaborative exploration in uncertain underground environments. The system can dynamically and rationally allocate command, reconnaissance, and relay roles based on real-time environment and robot status, and maintain task execution through reputation-based dynamic alliance negotiation when communication is interrupted. Its brain-like path planning strategy improves navigation efficiency and adaptability in unknown unstructured environments; the adaptive event-triggered mechanism significantly reduces network communication energy consumption and extends system operating time; and the collision self-recovery mechanism ensures continuous system operation under single-point failures. In summary, this invention effectively improves the overall detection efficiency, environmental adaptability, communication reliability, and task robustness of multi-heterogeneous robot systems. Attached Figure Description

[0018] Figure 1 This is a schematic block diagram of the overall structure of an integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration, according to the present invention.

[0019] Figure 2 This is a flowchart illustrating the capability encoding layer of a heterogeneous robot allocation and networking integrated system for underground space exploration according to the present invention. Detailed Implementation

[0020] To make the technical problems solved, the technical solutions, and the beneficial effects of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0021] Combination Figure 1 and Figure 2As shown in the figure, an embodiment of the present invention provides an integrated system for heterogeneous robot allocation and networking planning for underground space exploration, comprising: a perception layer for acquiring environmental information through lidar, visual sensors, and an inertial navigation system; a decision layer connected to the perception layer, including a task allocation module, a path planning module, and a collision self-recovery module; and a communication layer connecting the perception layer and the decision layer, used to form a green and energy-saving self-organizing network communication using an adaptive event triggering mechanism. The adaptive event triggering mechanism evaluates the necessity of communication based on a meta-relational attention network, which is a multi-layered network including an input layer, a capability encoding layer, a relational attention layer, a role reasoning layer, and an output layer. The task allocation module adopts a two-layer architecture of top-level distributed negotiation and bottom-level global allocation to realize task allocation for heterogeneous robots in underground space exploration under unknown environments. The path planning module adopts a brain-like learning strategy that integrates the meta-relational attention network and reinforcement learning to realize path planning for heterogeneous robots in underground space exploration under unknown environments, enabling end-to-end path planning. The collision self-recovery module performs collision detection based on the meta-relational attention network and executes a negotiation recovery mechanism to achieve rapid self-recovery after robot collisions, improving the robot's task completion efficiency. Through the cooperation between the perception layer, decision-making layer, and communication layer, the rationality of multi-heterogeneous robot allocation, networking reliability, and planning efficiency can be effectively improved.

[0022] Specifically, the perception layer includes a collaboratively configured and data-fused LiDAR, a visual sensor, and an inertial navigation system; wherein, the LiDAR is used to acquire the three-dimensional geometric structure and depth information of the environment; the visual sensor is used to acquire the texture, color, and semantic information of the environment; the inertial navigation system is used to acquire the robot's body motion acceleration, angular velocity, and attitude information; the LiDAR point cloud data and the visual sensor image data are spatiotemporally registered and fused to form an environmental perception result that combines geometric and semantic information, and the inertial navigation system provides time synchronization and motion compensation for the data fusion.

[0023] Specifically, the task allocation module is used to perform the initial allocation at the lower level through the meta-relational attention network, and to complete the final task allocation by performing a hybrid allocation method according to the top-level distributed negotiation strategy.

[0024] Specifically, the adaptive event triggering mechanism of the communication layer includes: analyzing robot state, environmental information, and network quality data through the meta-relational attention network to assess the urgency of communication and make triggering decisions to dynamically adjust the network topology. The adaptive event triggering mechanism utilizes the meta-relational attention network to evaluate whether to communicate, achieving dynamic topology management and a green, energy-efficient self-organizing network. The specific process is as follows: First, the perception layer collects multi-dimensional data such as robot state, environmental information, and network quality; second, it uses the meta-relational attention network to analyze the complex dependencies between these data; then, based on the analysis results, it assesses the necessity and urgency of current communication; next, it makes a triggering decision: communication (when and how) or hibernation (energy saving); finally, it dynamically adjusts the network topology based on the decision results to achieve a green, energy-efficient self-organizing network.

[0025] Specifically, the underlying initial allocation strategy is as follows: using the meta-relational attention network, the robot is allocated as a central command robot, a reconnaissance robot, or a relay robot based on the robot's mobility, sensor range, and load capacity; the central command robot, reconnaissance robot, and relay robot are respectively responsible for command, reconnaissance, and communication tasks, thereby completing the underlying initial allocation task.

[0026] Specifically, the top-level distributed negotiation strategy employs a hybrid allocation method based on dynamic alliances and reputation games, including a dynamic alliance formation unit, a reputation game negotiation unit, an intelligent concession strategy unit, and a reputation evolution unit. When the reconnaissance robot enters deep exploration and loses contact with the central command robot, the top-level distributed negotiation mechanism is activated, and the hybrid allocation method is used to complete the final task allocation.

[0027] Specifically, the dynamic alliance forming unit establishes temporary alliances based on task requirements and the complementarity of robot capabilities; for the task... t j The Robots Alliance is needed. The value of an alliance is defined as: , where α, β, and γ are weighting coefficients; The capability matching degree is defined as follows: ,in, It is a task t j Capability requirement vector; The execution cost is defined as follows: ,in, d ij It's distance. It's speed. e i move and e iwork It is the energy consumption coefficient; The cooperative gain is defined as follows: ,in, w ik It is a collaborative weight matrix.

[0028] Specifically, the reputation game negotiation unit conducts multiple rounds of negotiation based on the robot's reputation value and historical performance; (1) establishes a reputation-capability matrix for the alliance. C Robots in i Define the reputation-capability vector: p i =[ r i , c i , h i , t i ] T ,in, r i ∈ [0,1] represents the reputation value. For capability vectors, h i To ensure the success rate of historical cooperation, t i For response time reliability; (2) Calculate the initial negotiation weights for each robot: , where α+β+γ=1.

[0029] Specifically, the intelligent concession strategy unit performs multiple rounds of proposal-feedback iterations, adjusting the strategy according to the consensus level in each round; (1) the first Round proposal generation: Each robot Propose an allocation plan ,in a ij ∈ [0,1] represents a robot (2) Consensus calculation: (3) Feedback adjustment: If Then, a concession adjustment will be made: ,in, i consensus δ represents the consensus threshold. i （t） It is the concession amount, ∆ a i （t） It means adjusting the direction.

[0030] Specifically, the reputation evolution unit dynamically updates the reputation value based on the negotiation results and task performance; (1) Reputation update after an interaction: ,in, It's the learning rate. R actual It is the actual effect of cooperation. R expected This is the expected result; (2) R actual The calculation is as follows: ,in, , , These are the weighting coefficients.

[0031] Specifically, the collision self-recovery module includes: a collision detection unit, which uses the meta-relational attention network to classify the collision state into mild jamming, moderate entrapment, or severe malfunction based on IMU abnormal acceleration, current sensor load mutation, and depth camera visual collision recognition information; and a negotiation recovery unit, which calls the recovery strategy library according to the collision state and forms a temporary rescue team through a health robot to carry out dynamic rescue based on a force feedback collaborative drag strategy.

[0032] The aforementioned negotiation-based recovery mechanism is as follows: First, establish a recovery strategy library: Recovery Strategy = {Mild stuck: Move in the opposite direction and try different angles; Moderate stuck: Call for assistance from the collaborative robot; Severe fault: Enter safe mode and upload fault data}; Then, a collaborative recovery mechanism is established: health robots form temporary rescue teams, and their rescue roles are dynamically assigned, including rescuers, observers, and communication relays, and a collaborative dragging strategy based on force feedback is established. Finally, the post-recovery status update includes: updating the local map and marking dangerous areas; adjusting task assignments to avoid similar situations; and updating the case library to accumulate recovery experience.

[0033] Specifically, the meta-relational attention network consists of an input layer, a capability encoding layer, a relational attention layer, a role reasoning layer, and an output layer. The input layer receives sensor configurations, robot body parameters, and environmental features. The capability encoding layer then obtains a capability encoding vector. The relational attention layer then learns the capability-role relational attention mechanism, studying the importance weights of different capability dimensions for different roles. Finally, the role reasoning layer performs reasoning decisions based on capability-role relationships. Furthermore, the meta-relational attention network learns network weight parameters through meta-learning, thereby completing the top-level task allocation.

[0034] Specifically, the capability encoding layer of the meta-relation attention network is used to encode the input to obtain a capability encoding vector; the capability encoding layer is specifically as follows: First, sensor configuration is achieved through sensor feature encoding. After two layers of fully connected network coding: , in, s i For the first i Sensor configuration vectors for each robot; d s Configure the dimensions of the vectors for the sensors; h i sensor For the first i Sensor feature encoding results for a robot; W 1 sensor , W 2 sensor Here are the weight matrices for the first and second layers of the sensor coding network; b 1 sensor , b 2 senso Here, σ represents the bias vectors of the first and second layers of the sensor coding network; σ is the activation function, and the layer normalization operation is defined as: , in, For small constants in layer normalization, For input vectors x variance For input vectors x The mean, c 1. β 1 represents the scaling and offset parameters for layer normalization; Secondly, multimodal feature fusion is performed, as follows: , , in, h i This is the joint feature vector obtained by concatenating the five features; These correspond to the feature codes for mobility, communication, load capacity, and environmental awareness, respectively. W 1 fusion , W 2 fusion This is the weight matrix for merging the first and second layers of the network; b 1 fusion , b 2 fusion This is the bias vector of the first and second layers of the fusion network; ultimately, we obtain... d v dimensional capability vector .

[0035] Specifically, the relational attention layer of the meta-relational attention network is used to calculate the relational strength between the capability vector and the character prototype through a multi-head attention mechanism to obtain the relational attention matrix; the relational attention layer is specifically as follows: First, we learn the character archetype and define it. Learnable prototype matrix for each character: , in, K The total number of characters is the [number]th. The prototype vector of each character; Secondly, establish a multi-head attention mechanism for each role. Multi-head attention is used to calculate attention weights, as follows: Query Matrix , Key matrix,

[0036] Value matrix , Attention score calculation: ; in, W k Q This is the weight matrix for the query matrix; W k K The weight matrix is ​​the key matrix; v i This is a capability vector; W k V The weight matrix is ​​the value matrix; p k The prototype of the character; softmax is the normalized exponential function; d q , d v This specifies the dimensions of the query vector and key vector.

[0037] Then, relationship strength prediction is performed, and the capability vector is calculated. v i With the prototype of the character p k The strength of the relationship is as follows: splicing characteristics: , Predicting relationship strength using a three-layer neural network: , in, r ik ∈R M Represents robots With the character The relationship strength vector; W 1, W 2, W 3 represents the weight matrix of a three-layer neural network; b 1, b 2, b 3 represents the bias vector of a three-layer neural network; M is the dimension of the relation strength vector.

[0038] The relational attention matrix is ​​obtained as follows: , in, α ik It is attention weight. r ik It is the relation strength vector.

[0039] Specifically, the role inference layer of the meta-relational attention network is used for team-level feature propagation and role probability calculation based on the relational attention matrix and graph convolutional network; the role inference layer is specifically as follows: (1) Construct a robot relationship graph and define a robot capability similarity matrix. : ; in, n The total number of robots, The bandwidth parameter of the Gaussian kernel; (2) Construct the adjacency matrix A : ; in, S ij For robots i With robots j Similarity of abilities; t The similarity threshold; (3) Perform graph convolution operation, as follows: (3.1) Using graph convolutional networks to propagate information, let the first... l Layer node characteristics are The adjacency matrix is A Normalized adjacency matrix: ; in, D It is a degree matrix. I It is the identity matrix; (3.2) Graph convolution formula: ; Among them, afterL Team-level features are obtained after layer propagation. H （L） , W （l） For the first l The weight matrix of layer graph convolution; (3.3) Perform role probability calculations for robots. Combining individual relationship characteristics and team characteristics: ; in, A i [:, m ] is the relational attention matrix A i The m List; h i （L） For robots i The corresponding row vector in the team-level feature matrix; M is the dimension of the relation strength vector.

[0040] (3.4) Calculation of character probability: ; in, q i The Each element represents a robot. Assigned role The probability, W role , b role The weight matrix and bias vector of the role probability prediction layer; f i For robots i The final feature vector is obtained by concatenating the column average of the relational attention matrix with the team features.

[0041] Specifically, the output layer of the meta-relational attention network is used to output the decision result.

[0042] Specifically, the dynamic alliance forming unit forms a temporary alliance based on task requirements and the complementarity of robot capabilities, and the alliance value is calculated by weighting capability matching degree, execution cost and synergy gain; the reputation game negotiation unit conducts multiple rounds of negotiation based on reputation-capability matrix and historical performance; the intelligent concession strategy unit conducts multiple rounds of proposal-feedback iterative adjustment based on consensus; and the reputation evolution unit dynamically updates the robot's reputation value based on negotiation results and task performance.

[0043] Specifically, the capability coding layer is implemented through the following steps: S1. Encode the sensor configuration using a two-layer fully connected network and perform layer normalization. S2. The encoded sensor features, robot body parameters and environmental features are spliced ​​together; S3. A multidimensional capability encoding vector is obtained by mapping the fully connected network to the activation function.

[0044] In summary, (1) This invention integrates task allocation technology, path planning technology, communication technology, automatic control technology, and artificial intelligence technology to propose an integrated system for heterogeneous robot allocation and networking planning in underground space exploration under uncertain environments. Through environmental perception, green self-organizing network, task allocation, path planning, and collision self-recovery, it effectively improves the rationality of multi-heterogeneous robot allocation, network reliability, and planning efficiency, and endows multi-heterogeneous robot operation with reliability, safety, and environmental adaptability. (2) This invention integrates task allocation technology and artificial intelligence technology. Through a hybrid task allocation framework, it can seamlessly switch between normal and interrupted communication modes to ensure the continuous operation of the system and effectively improve the adaptability of heterogeneous robot exploration. It adopts a two-layer architecture of "top-level distributed negotiation and bottom-level global allocation" to realize the task allocation of heterogeneous robots in underground space exploration under unknown environments. (3) This invention integrates path planning technology, artificial intelligence technology, and automatic control technology. The path planning adopts a brain-like learning strategy that integrates meta-relational attention network and reinforcement learning, which can autonomously learn environmental features and improve the planning efficiency of robots in unknown environments. (4) This invention integrates artificial intelligence technology and automatic control technology to design a collision self-recovery mechanism, ensuring that a single robot failure will not lead to system paralysis, thereby improving task completion rate and system robustness. (5) This invention integrates communication technology and artificial intelligence technology to establish an adaptive event triggering mechanism, dynamically adjusting the network topology based on decision results to achieve green and energy-saving self-organizing network communication.

[0045] It is understood that the above specific description of the present invention is only for illustrating the present invention and is not limited to the technical solutions described in the embodiments of the present invention. Those skilled in the art should understand that modifications or equivalent substitutions can still be made to the present invention to achieve the same technical effect; as long as the use needs are met, they are all within the protection scope of the present invention.

Claims

1. An integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration, characterized in that: include: The perception layer is used to acquire environmental information through lidar, visual sensors, and inertial navigation systems. The decision-making layer, connected to the perception layer, includes a task allocation module, a path planning module, and a collision self-recovery module; The communication layer connects the perception layer and the decision layer, and is used to form a green and energy-saving self-organizing network communication with the help of an adaptive event triggering mechanism. The adaptive event triggering mechanism evaluates the necessity of communication based on a meta-relational attention network, which is a multi-layered network including an input layer, a capability encoding layer, a relational attention layer, a role reasoning layer, and an output layer. The task allocation module adopts a two-layer architecture of top-level distributed negotiation and bottom-level global allocation. The path planning module adopts a brain-like learning strategy that integrates the meta-relational attention network and reinforcement learning. The collision self-recovery module performs collision detection and executes a negotiation recovery mechanism based on the meta-relational attention network. The task allocation module is used to perform initial allocation at the lower level through the meta-relational attention network, and to complete the final task allocation by executing a hybrid allocation method according to the top-level distributed negotiation strategy. The initial allocation strategy at the lower level is specifically as follows: using the meta-relational attention network, the robot is allocated as a central command robot, a reconnaissance robot, or a relay robot based on the robot's mobility, sensor range, and load capacity. The hybrid allocation method adopted by the top-level distributed negotiation strategy is a distributed negotiation method based on dynamic alliance and reputation game, including a dynamic alliance formation unit, a reputation game negotiation unit, an intelligent concession strategy unit, and a reputation evolution unit. The meta-relational attention network includes: The input layer is used to receive sensor configurations, robot body parameters, and environmental characteristics. The capability encoding layer is used to encode the input to obtain a capability encoding vector. The capability encoding layer is implemented through the following steps: S1, the sensor configuration is encoded by a two-layer fully connected network and layer normalization is performed; S2, the encoded sensor features, robot body parameters and environmental features are concatenated; S3, a multi-dimensional capability encoding vector is obtained by mapping the fully connected network with an activation function. The relational attention layer is used to calculate the relational strength between the ability vector and the character prototype through a multi-head attention mechanism, and obtain the relational attention matrix. The role inference layer is used for team-level feature propagation and role probability calculation based on the relational attention matrix and graph convolutional network. The output layer is used to output the decision results.

2. The integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration according to claim 1, characterized in that: The perception layer includes a collaboratively configured and data-fused lidar, visual sensor, and inertial navigation system; The lidar is used to acquire the three-dimensional geometric structure and depth information of the environment; the vision sensor is used to acquire the texture, color and semantic information of the environment; and the inertial navigation system is used to acquire the robot's body motion acceleration, angular velocity and attitude information. The lidar point cloud data and visual sensor image data are spatiotemporally registered and fused to form an environmental perception result that combines geometric and semantic information. The inertial navigation system provides time synchronization and motion compensation for the data fusion.

3. The integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration according to claim 1, characterized in that: The adaptive event triggering mechanism of the communication layer specifically includes: analyzing robot state, environmental information and network quality data through the meta-relational attention network, assessing communication urgency and making triggering decisions to dynamically adjust the network topology.

4. The integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration according to claim 1, characterized in that: The collision self-recovery module specifically includes: The collision detection unit uses the meta-relational attention network to classify collision states into mild jamming, moderate entrapment, or severe malfunction based on IMU abnormal acceleration, current sensor load abrupt change, and depth camera visual collision recognition information. The negotiation recovery unit is used to call the recovery strategy library according to the collision status, and to form a temporary rescue team with the health robot to carry out dynamic rescue based on the force feedback collaborative drag strategy.

5. The integrated system for the allocation, networking, and planning of heterogeneous robots for underground space exploration according to claim 1, characterized in that: The dynamic alliance forming unit forms a temporary alliance based on task requirements and robot capability complementarity. The alliance value is calculated by weighting capability matching degree, execution cost and synergy gain. The reputation game negotiation unit conducts multiple rounds of negotiation based on the reputation-ability matrix and historical performance. The intelligent concession strategy unit performs multiple rounds of proposal-feedback iterative adjustments based on the consensus level. The reputation evolution unit dynamically updates the robot's reputation value based on the negotiation results and task performance.

Images