A multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication

Abstract

The application discloses a kind of multi-quadruped robot cooperative search and rescue methods based on task perception semantic communication, comprising: obtaining the external environment perception information of quadruped robot, own state information and search and rescue task context information;Obtain high-value semantic token set;Carry out feature aggregation, map semantic feature vector to complex continuous channel symbol, channel decoding and semantic recovery are carried out to received signal, obtain task-related environmental semantic information;Build embodied decision input, output joint space residual action, superimpose it with reference joint position to generate target joint instruction, drive quadruped robot to execute corresponding movement.The application combines the dynamic semantic token screening of task perception with joint source channel coding, maps semantic features to continuous channel symbols for wireless transmission, and improves the transmission efficiency and recovery availability of mission-critical semantic information under limited bandwidth, low signal-to-noise ratio and wireless channel fading.

Description

Technical Field

[0001] This invention belongs to the field of multi-robot cooperative control technology, specifically relating to a multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication. Background Technology

[0002] Quadruped robots possess strong terrain adaptability, enabling them to traverse rubble, slopes, steps, and irregular obstacles. They are suitable for target search, environmental exploration, and collaborative rescue in unstructured post-disaster environments such as earthquakes, collapses, and fires. Multiple quadruped robots can expand the search range and improve task completion efficiency through division of labor and cooperation, but their collaborative capabilities are highly dependent on information exchange between robots.

[0003] Wireless links in post-disaster environments are typically affected by factors such as limited available bandwidth, obstruction, interference, fading, and intermittent connectivity. If robots directly exchange raw depth maps, point clouds, or dense visual features, it not only consumes significant communication resources and increases transmission latency, but also introduces a large amount of background information in the raw perception data that is weakly correlated with current navigation, obstacle avoidance, and target search. When communication resources are limited, transmitting all observations evenly is insufficient to meet the demands of real-time embodied decision-making.

[0004] Existing semantic communication and deep joint source-channel coding methods can compress high-dimensional sensory data and transmit it end-to-end in noisy channels. However, many methods prioritize sensory reconstruction itself, failing to explicitly distinguish the contribution of different spatial semantics to the search and rescue mission. Some dynamic pruning methods rely solely on visual redundancy for token selection, neglecting to fully utilize task context such as target orientation and relative teammate positions. Consequently, limited bandwidth may be consumed by low-task-value semantics, and even if the receiver achieves good overall reconstruction quality, it may not provide the most valuable information for robot action decisions.

[0005] On the other hand, existing multi-robot reinforcement learning methods typically focus on the motion strategy itself or assume ideal communication between robots. When semantic extraction, wireless transmission, and multi-robot control are designed independently, communication errors and semantic gaps are difficult to effectively consider in control strategy training, resulting in limited collaborative decision-making performance under conditions of complex obstacles, local observations, and unstable wireless links.

[0006] Therefore, there is a need for a method that can proactively filter high-value environmental semantics for search and rescue missions, adapt to restricted and unstable wireless links, and directly apply recovery semantics to distributed embodied decision-making of quadruped robots, thereby forming a unified closed loop between communication efficiency, semantic robustness and collaborative search and rescue performance. Summary of the Invention

[0007] To address the shortcomings of the existing technologies, the present invention aims to provide a multi-quadruped robot collaborative search and rescue method based on task-aware semantic communication, in order to solve the problems of high overhead of original perception data transmission in post-disaster environments with limited communication, difficulty in effectively distinguishing visual information redundancy from task value, separation of semantic communication and embodied motion control, and insufficient stability of multi-robot collaborative strategy training under local observation conditions.

[0008] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0009] The present invention discloses a multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication, comprising the following steps:

[0010] 1) Obtain the external environment perception information, self-state information and search and rescue mission context information of each quadruped robot;

[0011] 2) Convert the deep observation information in the collected external environment perception information into spatial semantic tokens, and combine them with the current target search, path navigation and obstacle avoidance tasks to dynamically evaluate and filter the importance of each spatial semantic token, remove redundant semantic tokens that are not highly relevant to the current search and rescue task, and obtain a set of high-value semantic tokens.

[0012] 3) The high-value semantic token set is feature-aggregated, and the semantic feature vector is mapped to complex-valued continuous channel symbols through a source channel coding network and sent to the cooperative quadruped robot; the cooperative quadruped robot performs channel decoding and semantic recovery on the received signal to obtain task-related environmental semantic information;

[0013] 4) Integrate the current state information of the quadruped robot, the context information of the search and rescue mission, and the semantic information of the mission-related environment to construct embodied decision input, input the shared embodied motion strategy network, output the joint space residual motion, superimpose it with the reference joint position to generate the target joint command, and drive the quadruped robot to perform the corresponding motion.

[0014] Further, step 1) specifically includes:

[0015] Collect information about the external environment and the quadruped robot's own state.

[0016] The external environment perception information is a local depth map collected by a depth camera, which describes the terrain geometry, obstacle distribution and passable area information within the current field of view of the quadruped robot.

[0017] The quadruped robot's self-state information describes the quadruped robot's own motion state, joint state, and foot contact state;

[0018] Construct search and rescue mission context information; A quadruped robot at any time Search and rescue mission context Represented as:

[0019] ;

[0020] in, This refers to the search and rescue mission guidance information obtained based on current local observation information and existing environmental semantics, used to describe the spatial relationship between the candidate target area, the search front area, or the high-value exploration area and the current quadruped robot; Indicating a collaborative quadruped robot Compared to current quadruped robots Status information;

[0021] ;

[0022] ;

[0023] In the formula, and These represent the projections of the task guidance area onto the horizontal and vertical axes of the quadruped robot's body coordinate system, respectively. This indicates the heading deviation between the current orientation of the quadruped robot and the mission guidance direction; and These represent collaborative quadruped robots. Compared to current quadruped robots Relative displacement in the horizontal and vertical directions of the body coordinate system This indicates the relative heading angle between the two.

[0024] Further, step 2) specifically includes:

[0025] 21) Divide the depth observation information in the external environment perception information into multiple image blocks, perform block embedding on each image block, and add category tokens and position codes to obtain an initial visual semantic token sequence;

[0026] 22) Determine the global dependency relationship of the visual semantic tokens in the initial visual semantic token sequence; extract the local semantic features of each visual semantic token, and aggregate the local semantic features of the currently retained visual semantic tokens to obtain global context features;

[0027] 23) By integrating local semantic features, global context features, and current search and rescue mission context information, the correlation between each visual semantic token and the current search and rescue mission is evaluated to obtain the corresponding retention probability or importance score. After screening, a set of high-value semantic tokens is obtained.

[0028] Furthermore, step 3) specifically includes:

[0029] 31) Pool and aggregate the obtained high-value semantic token set to obtain the semantic feature vector of the cooperative quadruped robot;

[0030] 32) The semantic feature vector is mapped to a complex-valued continuous channel symbol through a channel coding network, and the channel symbol is power normalized according to the average transmit power constraint;

[0031] 33) The normalized continuous channel symbols are transmitted to the cooperative quadruped robot through the wireless channel between the quadruped robots;

[0032] 34) The cooperative quadruped robot sequentially decodes and recovers the received continuous channel symbols through the channel decoder and semantic decoder to obtain task semantic information related to target search, path navigation, obstacle avoidance and environmental occupancy status.

[0033] Further, step 4) specifically includes:

[0034] 41) Integrate the current state information of the quadruped robot, the context information of the search and rescue mission, and the semantic information of the mission-related environment to construct embodied decision input;

[0035] 42) The embodied decision input is input into the embodied motion strategy network, the probability distribution corresponding to the joint space residual action is output, and the joint space residual action is obtained by sampling according to the probability distribution;

[0036] 43) The joint space residual motion is superimposed with the reference joint position to generate the target joint command, and the corresponding joint control quantity is generated by the proportional differential controller according to the target joint command and the current joint state, and the quadruped robot is driven to perform the corresponding motion.

[0037] The beneficial effects of this invention are:

[0038] 1. This invention does not directly transmit complete depth observations or all visual features. Instead, it dynamically assesses the importance of spatial semantic tokens in conjunction with the current search and rescue mission. It prioritizes retaining task semantic information related to target search, obstacle distribution, passable areas, and high-value exploration areas. This reduces data redundancy with low relevance to the current search and rescue decision from the source of perception information and reduces wireless communication overhead between quadruped robots.

[0039] 2. This invention combines a task-aware dynamic semantic token filtering mechanism with joint source channel coding, directly mapping the filtered compact semantic features into continuous channel symbols for wireless transmission. Combined with transmit power normalization and end-to-end channel decoding and semantic recovery mechanisms, it improves the transmission efficiency and recovery availability of key semantic information of the task under non-ideal communication conditions such as bandwidth limitation, low signal-to-noise ratio and wireless channel fading.

[0040] 3. This invention directly uses the recovered collaborative task-related environmental semantic information as the decision input of the embodied motion strategy network, which couples semantic information extraction, wireless transmission, collaborative decision-making and quadruped robot motion execution to form a closed loop of perception, communication, decision-making and execution, reducing the inconsistency between the traditional semantic reconstruction target and the actual collaborative search and rescue task requirements.

[0041] 4. This invention adopts a collaborative learning mechanism of centralized training and distributed execution. During the training phase, it utilizes global terrain projection, the decision states of each robot, and global lossless semantic context to assist in global value assessment, thereby improving the stability of policy learning under the condition of local observation of multiple robots. During the actual deployment phase, it does not need to rely on global privileged information, but can complete distributed collaborative decision-making by only using local perception, task context, and received teammate semantic information, thus adapting to the actual operating conditions of local observability and unstable wireless links in the post-disaster environment.

[0042] 5. This invention couples high-level collaborative search and rescue decision-making with low-level quadruped robot motion control through joint space residual motion, low-level proportional differential control, and a task reward mechanism that includes target approach, heading consistency, speed tracking, search and rescue progress, collision and termination status. While ensuring motion stability and obstacle avoidance safety, it improves the target search, path planning and collaborative environment exploration capabilities of multi-quadruped robots, thereby enhancing the overall search and rescue efficiency and collaborative execution capabilities in complex post-disaster environments. Attached Figure Description

[0043] Figure 1 This is a schematic diagram of the method of the present invention. Detailed Implementation

[0044] To facilitate understanding by those skilled in the art, the present invention will be further described below with reference to embodiments and accompanying drawings. The content mentioned in the embodiments is not intended to limit the present invention.

[0045] Reference Figure 1 As shown, the present invention provides a multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication, comprising the following steps:

[0046] 1) Obtain the external environment perception information, self-state information and search and rescue mission context information of each quadruped robot;

[0047] Specifically, step 1) includes:

[0048] A group of quadruped robots was set up to participate in the collaborative search and rescue. for:

[0049] ;

[0050] in, This indicates the number of quadruped robots participating in the coordinated search and rescue operation. Indicates the first A quadruped robot;

[0051] The first A quadruped robot Abstract representation:

[0052] ;

[0053] in, This represents the set of sensors of a robot. This represents the robot's computational and decision-making unit. This refers to the actuator system of a robot. This represents the mechanical dynamics and kinematic constraints of the robot. This refers to the robot's wireless communication module. ;

[0054] Collect information about the external environment and the quadruped robot's own state.

[0055] The external environment perception information is a local depth map acquired by a depth camera, describing the terrain geometry, obstacle distribution, and passable area information within the current field of view of the quadruped robot; A quadruped robot at any time Acquired depth observations Represented as:

[0056] ;

[0057] Where H, W, and C represent the height, width, and number of channels of the depth observation, respectively. It is a real number;

[0058] The quadruped robot's self-state information describes its own motion state, joint state, and foot contact state; the... A quadruped robot at any time its own state Represented as:

[0059] ;

[0060] in, and These represent the linear velocity and angular velocity of the quadruped robot's body, respectively. and These represent the joint position and joint velocity of the quadruped robot, respectively. This indicates the movement of the quadruped robot at the previous time step. Indicates the contact state of the quadruped robot's foot tips;

[0061] To establish a connection between the external environment perception information acquired by the quadruped robot and the current search and rescue mission, search and rescue mission context information is constructed; A quadruped robot at any time Search and rescue mission context Represented as:

[0062] ;

[0063] in, This refers to the task guidance information obtained based on current local observation information and existing environmental semantics, used to describe the spatial relationship between the candidate target region, the search front region, or the high-value exploration region and the current quadruped robot. Indicating a collaborative quadruped robot Compared to current quadruped robots Status information;

[0064] ;

[0065] ;

[0066] In the formula, and These represent the projections of the task guidance area onto the horizontal and vertical axes of the quadruped robot's body coordinate system, respectively. This indicates the heading deviation between the current orientation of the quadruped robot and the mission guidance direction; and These represent collaborative quadruped robots. Compared to current quadruped robots Relative displacement in the horizontal and vertical directions of the body coordinate system This indicates the relative heading angle between the two.

[0067] During the coordinated search and rescue operation, the first The quadruped robot can also receive and recover task-related environmental semantic information sent by the cooperative quadruped robot; therefore, the first... A quadruped robot Local collaborative observation Represented as:

[0068] ;

[0069] in, Indicated by collaborative quadruped robots Send and in quadruped robot The task-related environmental semantic information is recovered.

[0070] 2) Convert the deep observation information in the collected external environment perception information into spatial semantic tokens, and combine them with the current target search, path navigation and obstacle avoidance tasks to dynamically evaluate and filter the importance of each spatial semantic token, remove redundant semantic tokens that are not highly relevant to the current search and rescue task, and obtain a set of high-value semantic tokens.

[0071] Specifically, step 2) includes:

[0072] Deep observation in external environment perception information The image is divided into multiple image blocks according to a preset image block size, and each image block is mapped to a spatial token using Patch Embedding to obtain a spatial token sequence. as follows:

[0073] ;

[0074] in, This represents the number of spatial tokens obtained from the depth map partitioning. This indicates the embedding dimension of the spatial token. Indicates the first A space token, express 3D real vector space;

[0075] An initial category token for aggregating global visual semantic information is added before the spatial token sequence, and a positional code is superimposed. The initial visual semantic token sequence is represented as follows:

[0076] ;

[0077] in, This represents the initial category token used to aggregate global visual semantic information; This represents the position code corresponding to each spatial token position;

[0078] Setting the first task-aware dynamic semantic coding network The input of each Transformer layer is When the first The Transformer layer does not belong to the preset dynamic pruning layer set. When the feature is updated, the standard Transformer module is used; when the first... When a Transformer layer belongs to a preset dynamic pruning layer set P, the lightweight token importance prediction module is activated to dynamically filter the current visual semantic tokens.

[0079] The local semantic features of the current j-th image patch token are obtained using a local multilayer perceptron. , represented as:

[0080] ;

[0081] Subsequently, the local semantic features of the currently retained visual semantic tokens are aggregated to obtain global contextual features. ,as follows:

[0082] ;

[0083] in, Use average pooling or other feature aggregation methods;

[0084] The local semantic features, search and rescue mission context information, and global context features are fused and input into the importance prediction module, which outputs the discard probability and retention probability of each image block token.

[0085] Among them, the search and rescue mission context This is used to impose task constraints on the assessment of token importance, so that importance prediction not only depends on the visual features of the image itself, but also on the current target search, path navigation and obstacle avoidance requirements.

[0086] Therefore, visual tokens associated with target direction, obstacle boundaries, passable areas, search frontiers, and high-value exploration areas receive higher retention priority, while tokens associated with background or redundant areas that are less relevant to the current search and rescue decision receive lower retention priority.

[0087] Since discrete token deletion operations cannot directly participate in end-to-end backpropagation, the Gumbel-Softmax method is used during the training phase to continuously approximate the token retention decision; for any token, a corresponding soft retention value is generated based on the importance prediction result. The soft mask vector is composed of the soft-reserved values ​​of all tokens in the current pruning layer.

[0088] To ensure that the tensor dimensions remain consistent within the same training batch, the training phase does not physically remove tokens to be pruned directly from the token sequence. Instead, attention masking is applied to the self-attention computation process based on a soft retention mask, so that suppressed tokens no longer significantly participate in the feature updates of other tokens, while ensuring that the network can perform parallel computation and gradient backpropagation.

[0089] Explicit token filtering is performed using the Top-k method; the first token is set. The target token retention ratio for each dynamic pruning layer is: The current number of image block tokens in this dynamic pruning layer is Then retain the number of tokens. for:

[0090] ;

[0091] Based on the token importance score output by the importance prediction module, all image patch tokens are sorted from highest to lowest score, and the top-ranked tokens are selected. The tokens form a set of retained indices, and the pruned token sequence is obtained.

[0092] After multiple dynamic pruning layers for progressive filtering, a set of high-value semantic tokens is obtained. ,as follows:

[0093] ;

[0094] in, Represents a set of high-value semantic tokens. This represents a category token used for global semantic aggregation. Indicates the number of items retained after dynamic filtering. A high-value spatial semantic token This indicates the number of high-value spatial semantic tokens that are ultimately retained. This indicates the number of spatial semantic tokens before dynamic pruning. This indicates that the number of semantic tokens retained after dynamic filtering is much smaller than the number of semantic tokens before filtering.

[0095] Category tokens are retained throughout each dynamic pruning layer to maintain global semantic aggregation.

[0096] Through the aforementioned task-aware dynamic token filtering process, the transmitting end no longer equally retains all visual areas in the depth observation, but dynamically selects the information to be transmitted based on its actual contribution to the current search and rescue mission, thereby reducing semantic redundancy in wireless communication from the sensing source end.

[0097] 3) The high-value semantic token set is feature-aggregated, and the semantic feature vector is mapped to complex-valued continuous channel symbols through a source channel coding network and sent to the cooperative quadruped robot; the cooperative quadruped robot performs channel decoding and semantic recovery on the received signal to obtain task-related environmental semantic information;

[0098] Specifically, step 3) includes:

[0099] The set of high-value semantic tokens obtained in step 2) Pooling aggregation is performed to obtain the semantic feature vector of the cooperative quadruped robot. ,as follows:

[0100] ;

[0101] in, This can be achieved by reading category tokens, average pooling, or a combination of both.

[0102] The semantic feature vector is processed through a channel coding network consisting of several fully connected layers. Mapped to complex-valued continuous channel symbols, as follows:

[0103] ;

[0104] in, This represents a channel encoder with parameter α. Represents complex-valued continuous channel symbols. , For the number of channels, Indicates the number of times the channel has been used;

[0105] Based on the average transmit power constraint P of the quadruped robot's wireless communication module, the power normalization of the complex-valued continuous channel symbols is expressed as:

[0106] ;

[0107] in, The complex-valued continuous channel symbol obtained after power normalization; This indicates the average transmit power allowed by the wireless communication module; This represents the mathematical expectation operation, used to calculate the average power of transmitted channel symbols;

[0108] The power-normalized complex-valued channel symbols are transmitted via a wireless channel between quadruped robots; the wireless transmission process is uniformly represented as:

[0109] ;

[0110] in, This indicates the signals received by the collaborative quadruped robot. This indicates the wireless channel gain between a quadruped robot and a collaborative quadruped robot. Indicates channel noise;

[0111] Based on the actual wireless propagation conditions in the post-disaster environment, the wireless channel between the quadruped robots can be modeled using additive white Gaussian noise, Rayleigh fading, or Ricean fading channels to describe different levels of noise, non-line-of-sight propagation, multipath fading, and main path propagation characteristics.

[0112] The collaborative quadruped robot uses a channel decoder to decode the received signal and obtain a set of high-value semantic tokens. Then, the semantic decoder recovers the task semantic information related to target search, path navigation, obstacle avoidance, and environmental occupancy status. The task-aware dynamic semantic coding network, channel coding network, channel decoder, and semantic decoder together constitute the task-aware semantic communication network, which is used to complete the processing from task-related semantic feature extraction, wireless transmission to semantic recovery at the receiving end.

[0113] The semantic communication process in this embodiment does not aim to accurately recover the original depth map of the sending end pixel by pixel. Instead, it aims to ensure that the recovered environmental semantics can retain information useful for downstream collaborative search and rescue decisions, such as obstacle occupancy status, target relative direction, passable areas, and high-value search areas.

[0114] The training objective of the task-aware semantic communication network Including task-related semantic recovery loss and token sparsity constraints, it is represented as:

[0115] ;

[0116] in, Used to constrain the effectiveness of recovering semantic information for downstream search and rescue missions. Used to limit the number of semantic tokens actually retained and sent. This represents the trade-off coefficient between the two.

[0117] Furthermore, considering search and rescue mission rewards, mission-related semantic recovery losses, token sparsity constraints, and wireless channel resource overhead, the semantic selection and transmission process at the transmitting end is kept consistent with the mission decision-making and action execution process at the receiving end, thereby avoiding the problem of only improving visual reconstruction quality without actually improving search and rescue decision-making performance.

[0118] 4) Integrate the current state information of the quadruped robot, the context information of the search and rescue mission, and the semantic information of the mission-related environment to construct embodied decision input, input it into the shared embodied motion strategy network, output the joint space residual motion, superimpose it with the reference joint position to generate the target joint command, and drive the quadruped robot to perform the corresponding motion;

[0119] Specifically, step 4) includes:

[0120] No. Each quadruped robot constructs embodied decision inputs based on its own state information, external environment perception information, search and rescue mission context information, and mission-related environmental semantic information obtained in step 3), as follows:

[0121] ;

[0122] in, Indicates the first A quadruped robot at any time Embodied decision input;

[0123] The embodied decision input is used as a shared embodied motion strategy network. The input is the continuous joint space residual motion;

[0124] To maintain the continuity of robot motion in complex terrain environments and to preserve appropriate exploration capabilities during the training phase, joint space residual motion can be sampled by constructing a multivariate Gaussian distribution based on the mean and covariance of the policy network output.

[0125] Joint space residual motion Reference joint position The superposition yields the following target joint commands:

[0126] ;

[0127] in, Indicates the first A quadruped robot at any time Target joint commands;

[0128] The proportional-derivative controller generates corresponding joint control quantities based on the target joint command, the current joint position, and the joint velocity, so that the target search, path adjustment, and cooperative obstacle avoidance intentions output by the upper controller are converted into joint movements that the robot can actually perform.

[0129] To guide the quadruped robot in developing a stable movement strategy for search and rescue missions, this embodiment employs a single-step task reward system consisting of multiple rewards and penalties. ,as follows:

[0130] ;

[0131] in, The target proximity reward is used to encourage the robot to shorten the distance to the target area; This indicates a heading alignment reward, used to encourage the robot to keep its current orientation consistent with the mission guidance direction; Indicates speed tracking reward; This indicates a reward for the progress of the search and rescue operation, used to evaluate the robot's current progress in the target search or environmental coverage task; This indicates a stagnation penalty; This represents the obstacle collision penalty. When the collision sensor or robot contact status detects an unpermitted collision, this item takes a non-zero value; otherwise, it takes zero. This indicates a termination penalty, used to suppress unsafe states such as robot falls, flips, posture instability, or joint control exceeding safety thresholds; when any termination condition is detected, a large negative reward is applied, and the current training is terminated; , , , , , , These are the weighting coefficients;

[0132] The robot receives a positive reward when it moves closer to the target area and a reduced reward when it moves further away from the target area, calculated based on the difference in target distance between two consecutive moments, as follows:

[0133] ;

[0134] in, Indicates time The distance between the robot and the current target area;

[0135] Based on the heading angle between the robot and the current target area The calculation is as follows:

[0136] ;

[0137] The reward is large when the robot is heading in the target direction; the reward gradually decreases as the heading deviation increases.

[0138] The calculation expression is:

[0139] ;

[0140] in, and These represent the robot's actual speed and its expected speed, respectively.

[0141] The calculation is based on changes in the area of ​​the newly added search region, changes in environmental coverage, or changes in target discovery confidence at adjacent time points, and is expressed as follows:

[0142] ;

[0143] in, and Representing time respectively and time The cumulative environmental coverage rate; positive rewards are given when the robot discovers new unexplored areas or new search and rescue targets; penalties are incurred when the robot's displacement is less than a stagnation threshold within a consecutive preset time window, as expressed as:

[0144] ;

[0145] in, and These represent the robot at time [time]. and time The position vector; The threshold for determining stagnation; This is an indicator function.

[0146] The above-mentioned reward mechanism enables the robot to approach the search and rescue target and expand the effective search range while taking into account the rationality of the path, the stability of the robot's posture, and the safety of obstacle avoidance.

[0147] After the robot performs the joint action, it causes changes in its own state and local environment observation. In the next time step, it re-perceives the environment, extracts task semantics, communicates with other robots, and makes motion decisions, thus forming a closed loop of perception-communication-decision-execution.

[0148] Furthermore, step 4) specifically includes:

[0149] During the strategy training phase, a global context evaluation network is introduced, and a centralized training and distributed execution mechanism is adopted to optimize the cooperative motion strategy of the multi-quadruped robot.

[0150] Built on time The global state includes the embodied decision inputs of each quadruped robot, global terrain projection, and global lossless semantic context information. ,as follows:

[0151] ;

[0152] in, This indicates that all quadruped robots are at time [time]. The set of embodied decision inputs, ; This represents the global terrain projection that can be obtained during the training phase. This represents the global lossless semantic context information available during the training phase;

[0153] The global state is input into the global value evaluation network. The expected cumulative discounted reward corresponding to the current state is estimated and used only during the intensive training phase. At the beginning of each training round, the search and rescue environment of multiple parallel quadruped robots is initialized or reset, and the initial self-state of each quadruped robot is obtained. At each time step, each robot acquires its own observation information, search and rescue task context information, and environmental semantic information related to the cooperative quadruped robot task, thereby constructing the first... A quadruped robot at any time Embodied decision input And through a network of shared embodied movement strategies Output joint space residual motion;

[0154] After the quadruped robot performs joint space residual actions, it obtains the task reward for the current time step and the state for the next time step. Store in the trajectory cache;

[0155] After completing the environmental interaction at the preset time step, the temporal difference residual is calculated using a global value evaluation network. ,as follows:

[0156] ;

[0157] in, For the reward function, Indicates the discount factor. The parameter is A global value evaluation network;

[0158] Furthermore, generalized advantage estimation is used to calculate the advantage value corresponding to the current action. ,as follows:

[0159] ;

[0160] in, This represents the generalized dominance estimation parameter, used to adjust the bias and variance in the dominance estimation process; This represents the time offset index starting from the current moment; Indicates time The timing difference error;

[0161] Based on the advantage estimate, a pruned proximal policy optimization target update shared embodied motion policy network is adopted to limit the magnitude of single policy parameter changes and reduce policy update oscillations caused by wireless semantic noise, environmental changes, and multi-robot interactions. At the same time, an entropy regularization term is added to the pruned proximal policy optimization target to enable the robot to maintain appropriate exploration capabilities during training.

[0162] The global value assessment network updates its value by minimizing the value estimation error between the state value estimation result and the actual sampled cumulative return.

[0163] Once training is complete, privileged information such as the global terrain projection and global lossless semantic context used during training are removed, and the global value evaluation network no longer participates in actual deployment. Each quadruped robot operates a shared or parameter-consistent lightweight embodied movement strategy based solely on its own state information, search and rescue mission context information, and mission-related environmental semantic information, thereby achieving distributed collaborative search and rescue in post-disaster environments with limited communication.

[0164] This invention has many specific applications. The above description is only a preferred embodiment of this invention. It should be noted that for those skilled in the art, several improvements can be made without departing from the principle of this invention, and these improvements should also be considered within the scope of protection of this invention.

Claims

1. A multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication, characterized in that, The steps are as follows: 1) Obtain the external environment perception information, self-state information and search and rescue mission context information of each quadruped robot; 2) Convert the deep observation information in the collected external environment perception information into spatial semantic tokens, and combine them with the current target search, path navigation and obstacle avoidance tasks to dynamically evaluate and filter the importance of each spatial semantic token to obtain a set of high-value semantic tokens. 3) The high-value semantic token set is feature-aggregated, and the semantic feature vector is mapped to complex-valued continuous channel symbols through a source channel coding network and sent to the cooperative quadruped robot; the cooperative quadruped robot performs channel decoding and semantic recovery on the received signal to obtain task-related environmental semantic information; 4) Integrate the current state information of the quadruped robot, the context information of the search and rescue mission, and the semantic information of the mission-related environment to construct embodied decision input, and input it into the shared embodied motion strategy network to output joint space residual motion. Superimpose it with the reference joint position to generate target joint command, and drive the quadruped robot to perform the corresponding motion.

2. The multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication according to claim 1, characterized in that, Step 1) specifically includes: Collect information about the external environment and the quadruped robot's own state. The external environment perception information is a local depth map collected by a depth camera, which describes the terrain geometry, obstacle distribution and passable area information within the current field of view of the quadruped robot. The quadruped robot's self-state information describes the quadruped robot's own motion state, joint state, and foot contact state; Construct search and rescue mission context information, the first A quadruped robot at any time Search and rescue mission context Represented as: ； in, This refers to the search and rescue mission guidance information obtained based on current local observation information and existing environmental semantics, used to describe the spatial relationship between the candidate target area, the search front area, or the high-value exploration area and the current quadruped robot; Indicating a collaborative quadruped robot Compared to current quadruped robots Status information; ； ； In the formula, and These represent the projections of the task guidance area onto the horizontal and vertical axes of the quadruped robot's body coordinate system, respectively. This indicates the heading deviation between the current orientation of the quadruped robot and the mission guidance direction; and These represent collaborative quadruped robots. Compared to current quadruped robots Relative displacement in the horizontal and vertical directions of the body coordinate system This indicates the relative heading angle between the two.

3. The multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication according to claim 1, characterized in that, Step 2) specifically includes: 21) Divide the depth observation information in the external environment perception information into multiple image blocks, perform block embedding on each image block, and add category tokens and position codes to obtain an initial visual semantic token sequence; 22) Determine the global dependency relationship of the visual semantic tokens in the initial visual semantic token sequence; extract the local semantic features of each visual semantic token, and aggregate the local semantic features of the currently retained visual semantic tokens to obtain global context features; 23) By integrating local semantic features, global context features, and current search and rescue mission context information, the correlation between each visual semantic token and the current search and rescue mission is evaluated to obtain the corresponding retention probability or importance score. After screening, a set of high-value semantic tokens is obtained.

4. The multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication according to claim 1, characterized in that, Step 3) specifically includes: 31) Pool and aggregate the obtained high-value semantic token set to obtain the semantic feature vector of the cooperative quadruped robot; 32) The semantic feature vector is mapped to a complex-valued continuous channel symbol through a channel coding network, and the channel symbol is power normalized according to the average transmit power constraint; 33) The normalized continuous channel symbols are transmitted to the cooperative quadruped robot through the wireless channel between the quadruped robots; 34) The cooperative quadruped robot sequentially decodes and recovers the received continuous channel symbols through the channel decoder and semantic decoder to obtain task semantic information related to target search, path navigation, obstacle avoidance and environmental occupancy status.

5. The multi-quadruped robot cooperative search and rescue method based on task-aware semantic communication according to claim 1, characterized in that, Step 4) specifically includes: 41) Integrate the current state information of the quadruped robot, the context information of the search and rescue mission, and the semantic information of the mission-related environment to construct embodied decision input; 42) The embodied decision input is input into the embodied motion strategy network, the probability distribution corresponding to the joint space residual action is output, and the joint space residual action is obtained by sampling according to the probability distribution; 43) The joint space residual motion is superimposed with the reference joint position to generate the target joint command, and the corresponding joint control quantity is generated by the proportional differential controller according to the target joint command and the current joint state, and the quadruped robot is driven to perform the corresponding motion.

Images