A navigation method for multiple autonomous underwater robots based on embodied intelligence in a non-communication environment

Abstract

The present invention discloses a multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment, comprising: step 1: each autonomous underwater robot obtains multi-modal local environmental information through the sensor carried by itself; step 2: fusion of multi-modal local environmental information; step 3: extracting environmental feature values ​​and feature values ​​of adjacent autonomous underwater robots; step 4: learning environmental feature values, understanding the environment, and inferring the motion trajectory of adjacent autonomous underwater robots based on feature values ​​of adjacent autonomous underwater robots; step 5: collaborative relationship learning, generating navigation decisions; step 6: sending the generated navigation decisions to the controller of each autonomous underwater robot for compilation, generating control instructions, and realizing navigation control of the autonomous underwater robot. The present invention introduces embodied intelligence technology, designs a non-communication multi-autonomous underwater robot collaborative navigation scheme, and can realize autonomous underwater robot navigation based on embodied intelligence.

Description

Technical Field

[0001] The present invention belongs to the technical field of underwater robot control, and in particular relates to a multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment. Background Art

[0002] With the rapid development of technology, artificial intelligence has shifted from artificial intelligence that focuses on learning from data sets to embodied intelligence that learns through interaction with the surrounding environment. Collaborative navigation of multiple autonomous underwater robots remains a challenging task in complex and unknown underwater environments. Existing navigation strategies for multiple autonomous underwater robots have the following defects:

[0003] First, almost all multi-autonomous underwater robot collaboration models require a centralized system or local communication system to exchange information between autonomous underwater robots to achieve collaboration. However, existing underwater communication technology is still a bottleneck. Electromagnetic communication cannot be carried out in water, sonar communication has serious delays, and optical communication is limited by equipment. In many cases, autonomous underwater robots cannot communicate with each other and can only operate in a non-communication environment.

[0004] Second, due to the limitation of the sensors they carry, autonomous underwater robots rely on only a few types of sensors for navigation. However, sonar detection angles are too large and have poor directionality; visual sensors are affected by light and have limited observation distances. Therefore, autonomous underwater robots have limited ability to perceive the environment, making it difficult to obtain a large number of effective data sets.

[0005] Third, existing navigation strategies are usually implemented in an end-to-end manner through imitation learning or reinforcement learning. However, large-scale annotated data is a basic requirement for end-to-end training, which may not be feasible or practical in many cases. In addition, the ability to generalize trained knowledge and skills to new tasks or different robot platforms remains largely inefficient.

[0006] Therefore, based on the above difficulties, it is necessary to design an effective navigation method so that multi-autonomous underwater robots can have higher navigation efficiency and obtain safety guarantee in a non-communication environment, so that they can meet the needs of actual engineering. Summary of the invention

[0007] The technical problem to be solved by the present invention is to address the deficiencies of the above-mentioned prior art and to provide a multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment. By introducing embodied intelligence technology, a non-communication multi-autonomous underwater robot collaborative navigation scheme is designed, which can realize autonomous underwater robot navigation based on embodied intelligence.

[0008] In order to achieve the above technical objectives, the technical solution adopted by the present invention is:

[0009] A multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment comprises the following steps:

[0010] Step 1: Each autonomous underwater robot obtains multimodal local environment information through its own sensors;

[0011] Step 2: Fusing the corresponding multi-modal local environment information of each autonomous underwater robot;

[0012] Step 3: Each autonomous underwater robot extracts environmental feature values ​​and feature values ​​of neighboring autonomous underwater robots based on the fused local environmental information;

[0013] Step 4: Each autonomous underwater robot learns the environmental feature values ​​based on the embodied intelligent unit, understands the environment, and infers the motion trajectory of the neighboring autonomous underwater robot based on the feature values ​​of the neighboring autonomous underwater robot;

[0014] Step 5: Each autonomous underwater robot learns the collaborative relationship based on the environment understood by the embodied intelligent unit and the inferred motion trajectory of the adjacent autonomous underwater robot, so that multiple autonomous underwater robots can achieve collaboration without communication conditions, and correct the generated navigation decision;

[0015] Step 6: Send the generated navigation decision to the controller of each autonomous underwater robot for compilation, generate control instructions, and realize navigation control of the autonomous underwater robot.

[0016] To optimize the above technical solutions, the specific measures taken also include:

[0017] In the above step 3, a convolutional neural network (CNN) is used to extract spatial features from the fused local environmental information. At the same time, a long short-term memory (LSTM) network is used to extract the temporal information corresponding to different spatial features from local observations to obtain environmental feature values ​​containing spatial and temporal information.

[0018] The extraction strategy of the feature value of the neighboring autonomous underwater robot described in step 3 above is:

[0019] First, a training detection network is used to mark the positions of adjacent autonomous underwater robots and detect their relative positions;

[0020] The training detection network consists of a cellular neural network with a residual structure connected to an MLP with a parallel architecture to extract high-dimensional features from an image-like representation;

[0021] An MLP performs the classification task to detect the number of n adjacent autonomous underwater vehicles (m1,m2,…,m n ), the non-existent autonomous underwater robot is represented by (0,…,0);

[0022] Another MLP outputs the relative positions of n adjacent autonomous underwater robots (loc1, loc2, ..., loc n ), the non-existent robot position is also represented by (0,…,0);

[0023] Secondly, by multiplying the number of adjacent autonomous underwater robots by the relative position, the marked position of the adjacent autonomous underwater robot at t = 0 is obtained, that is, the eigenvalue of the adjacent autonomous underwater robot: (mloc1, mloc2, ..., mloc m )=(m1,m2,…,m n )×(loc1,loc2,…,loc n ).

[0024] The embodied intelligent unit described in the above step 4 includes an embodied decision-making module and a multimodal storage module. The embodied decision-making module understands the input environmental feature values ​​through a large prediction model, and learns the results of the understanding in combination with the past environmental features and feature values ​​of nearby autonomous underwater robots stored in the multimodal storage module. Finally, the inference device is used to infer the motion trajectory of the nearby autonomous underwater robots, and the decision maker plans the navigation path of each autonomous underwater robot. At the same time, the embodied decision-making module sends the current environmental features, motion features of the autonomous underwater robot, and trajectory information to the multimodal storage module; the multimodal storage module classifies and stores the environmental features, motion features of the autonomous underwater robot, and trajectory information that are decided and understood in the embodied decision-making module, and provides the stored past environmental features, motion features of the autonomous underwater robot, and trajectory information to the embodied decision-making module for decision-making at the next moment.

[0025] The above-mentioned embodied decision module includes LLM, learner, reasoner and decision maker;

[0026] LLM understands the environmental information and the actions of other autonomous underwater robots around it based on the environmental feature values, and sends the understood information to the learner and reasoner;

[0027] The learner uses the environmental features, motion features, and trajectory information of the autonomous underwater robot understood by the LLM and the past environmental features, motion features, and trajectory information stored in the multimodal storage module to learn and make navigation plans;

[0028] The reasoner infers the further movement of other nearby autonomous underwater robots based on the environmental features, motion features and trajectory information of the autonomous underwater robot understood by the navigation plan and LLM;

[0029] The decision maker makes specific navigation decisions based on the motion results of other nearby autonomous underwater robots deduced by the reasoner.

[0030] The above-mentioned multimodal storage module adopts a memory tagging strategy to store environmental features, autonomous underwater robot motion features, and trajectory information. The memory tagging strategy encodes the multimodal experience and its narrative into the multimodal storage module, which catalogs the various aspects of data included in the multimodal scenario memory fragment and its corresponding labels in the environmental storage unit of the multimodal storage module. The label is stored together with the autonomous underwater robot motion memory type in the autonomous underwater robot motion storage unit of the multimodal storage module to achieve efficient recall; after locating the label of the multimodal plot record, the label and its associated multimodal content are fed to the embodied decision module.

[0031] The above step 5 specifically includes the following steps:

[0032] Step 5.1: Collaborative information aggregation of multiple autonomous underwater robots;

[0033] Step 5.2: Multi-robot cooperation strategy learning based on aggregated information.

[0034] The collaborative information aggregation of multiple autonomous underwater robots described in step 5.1 above is achieved by utilizing a graph attention network applied to graph representation, specifically:

[0035] Based on the understanding of the spatial and temporal information of the adjacent environment by the embodied intelligent decision-making module and the derived motion trajectory of the autonomous underwater robot in the adjacent area, according to the characteristics of the environment where communication is impossible, a single autonomous underwater robot achieves the collaboration of multiple autonomous underwater robots through learning based on the positions of other autonomous underwater robots to avoid collisions during navigation. Specifically, the relationship between the decision-making autonomous underwater robot and its adjacent autonomous underwater robots is described as an undirected graph G. In the kth graph convolutional layer, the following message passing and information aggregation operations are performed:

[0036]

[0037] Among them, H k Graph convolution operation in layer k with non-linear activation function H k-1 Represents the graph convolution operation in the k-1th layer with a non-linear activation function A represents the connection matrix of the direct relationship between the autonomous underwater robot and its neighboring robots, Ψ k represents the learnable weight matrix of the kth graph convolution layer, K represents the number of graph convolution layers, and γ k is the trainable bias, a mn represents the value of the mth row and nth column of the connection matrix A, LR represents the leaky rectifier linear unit activation function, δ represents the trainable attention kernel, (H k Ψ k ) m Represents Hk Ψ k The mth row of (H k Ψ k ) n Represents H k Ψ k The nth column of (H k Ψ k ) l Represents H k Ψ k In the lth column of , exp(·) represents the exponential operation.

[0038] The reward function r used in the above step 5.2 cooperative strategy learning t for:

[0039] r t = s r t + d r t + p r t (4)

[0040] in s r t Represents direction reward, d r t Represents collision reward, p r t Represents navigation reward;

[0041] Said s r t The expression is:

[0042] s r t =β s (d t -d t-1 ) (5)

[0043] where d t represents the yaw angle between the current position of the autonomous underwater robot and the position of its target at time t, d t-1 represents the yaw angle between the current position of the autonomous underwater robot and the position of its target at time t-1, β s represents a hyperparameter;

[0044] Said d r t The expression is:

[0045]

[0046] where p t represents the position of the autonomous underwater robot at time t, Ld and β d Respectively represent the obstacle position, safety distance and collision penalty factor, r collsion Indicates the penalty coefficient for collision, L c Indicates the distance to the obstacle;

[0047] Said p r t The expression is:

[0048]

[0049] Where g represents the location of the target point, p t represents the position of the autonomous underwater robot at time t, p t-1 represents the position of the autonomous underwater robot at time t-1, β g Indicates the penalty factor for not reaching the target point, r goal Represents the reward for reaching the target point, L g Indicates the distance to the target point.

[0050] The present invention has the following beneficial effects:

[0051] (1) The present invention proposes a collaborative navigation scheme for multiple autonomous underwater robots in a non-communication environment. First, each autonomous underwater robot uses its own sensors to detect and fuse information about the environment and perceive the surrounding information. Secondly, the fused information is subjected to feature value extraction. Next, the features of local observations are learned based on the embodied intelligence method, and spatial features are extracted from the environmental representation to help the underwater robot understand the changes in the environmental state and enhance its short-term memory of the environment. At the same time, the features of the movements of other autonomous underwater robots are extracted from the information of the observed local environment, that is, individuals infer the behavior of other autonomous underwater robots through self-observation. Then, the graph attention network is used to realize the collaboration of multiple autonomous underwater robots in a non-communication environment and make control decisions for the autonomous underwater robots. Finally, the autonomous underwater robot controller generates control instructions based on the control decisions to control the underwater robot. Collaborative navigation of multiple autonomous underwater robots can be realized in a non-communication environment.

[0052] (2) The present invention proposes an embodied intelligent unit for autonomous underwater robot navigation. The embodied intelligent unit includes two parts: an embodied decision-making module and a multimodal storage module. The navigation information of the autonomous underwater robot is stored in the multimodal information storage module. A memory marking strategy is adopted to simplify the reflection and retrieval of situational multimodal memory.

[0053] (3) This paper proposes a new continuous reward function for multi-robot cooperative strategy learning, which integrates expert guidance to guide the robots to the local endpoint while considering potential collisions, promoting efficient actions by balancing environmental avoidance and endpoint proximity, thereby improving navigation efficiency. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A block diagram of cooperative navigation of multiple autonomous underwater robots without communication provided by the present invention;

[0055] Figure 2 The autonomous underwater robot navigation decision framework based on embodied intelligence provided by the present invention;

[0056] Figure 3 Framework diagram of collaborative relationship learning for multiple autonomous underwater robots. DETAILED DESCRIPTION

[0057] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0058] Although the steps in the present invention are arranged with numbers, they are not used to limit the order of the steps. Unless the order of the steps is clearly stated or the execution of a certain step requires other steps as a basis, the relative order of the steps can be adjusted. It is understood that the term "and / or" used in this article involves and covers any and all possible combinations of one or more of the associated listed items.

[0059] The present invention provides a multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment, and the specific contents are as follows:

[0060] (1) A collaborative navigation system architecture for multiple autonomous underwater robots without communication. Traditional collaborative navigation for multiple robots relies on communication, where robots exchange information with each other to cooperate. However, it is difficult to achieve real-time and efficient communication in an underwater environment, and communication is often interrupted in scenes with complex terrain. The present invention designs a collaborative navigation system architecture for multiple autonomous underwater robots in a non-communication environment, such as Figure 1As shown in the figure. First, each autonomous underwater robot uses its own sensors to detect and fuse the environment information and perceive the surrounding information. Secondly, the fused information is extracted for feature values. Next, the features of local observations are learned based on the embodied intelligence method, and spatial features are extracted from the environmental representation to help the underwater robot understand the changes in the environmental state and enhance its short-term memory of the environment. At the same time, the features of the movements of other autonomous underwater robots are extracted from the information of the observed local environment, that is, individuals infer the behavior of other autonomous underwater robots through self-observation. Then, the Graph Attention Networks (GAT) are used to realize the collaboration of multiple autonomous underwater robots under non-communication conditions and make autonomous underwater robot control decisions. Finally, the autonomous underwater robot controller generates control instructions based on the control decision to control the underwater robot.

[0061] (2) Autonomous underwater robot navigation strategy based on embodied intelligence. The autonomous underwater robot navigation strategy based on embodied intelligence consists of two parts: an embodied decision-making module and a multimodal storage module, such as Figure 2 As shown. The embodied decision module contains the interacting LLM (Large Language Model), learner, reasoner and decision maker. The environmental information features detected by the sensor and the human instructions are sent to the LLM component to understand the environmental information and the actions of other autonomous underwater robots around it, and the results of the understanding are handed over to the learner and reasoner. The learner uses the information understood by the LLM and the information previously stored in the multimodal memory to make navigation plans. The reasoner deduce the further movement of other nearby autonomous underwater robots based on the navigation plan and the information understood by the LLM. Finally, the reasoning results of each autonomous underwater robot are passed to the decision maker to make specific navigation decisions. The multimodal storage module is to classify and store the information made and perceived in the embodied decision module, such as: environmental storage, motion control instructions, etc. At the same time, the stored information is provided to the embodied decision module for decision-making at the next moment.

[0062] (3) Collaborative relationship learning. The local environment features learned by the autonomous underwater robot through CNN (Convolutional Neural Networks) and LSTM (Long Short-Term Memory) methods are combined with the robot's next action through positioning and LSIL (Local Spatiotemporal Information Learning) reasoning, and the navigation decision made by the embodied intelligence algorithm. Then, the collaboration between autonomous underwater robots is learned using the CRL (Cooperation Relationship Learning) method based on the decision made by each autonomous underwater robot.

[0063] Finally, according to the collaborative navigation system architecture of multi-autonomous underwater robots without communication and the navigation strategy of autonomous underwater robots based on embodied intelligence, the collaborative navigation algorithm of multi-autonomous underwater robots was designed, such as Figure 3 shown.

[0064] like Figure 1-3 As shown, the specific steps of a multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment proposed by the present invention are as follows:

[0065] Step 1: Acquisition of environmental information: Each autonomous underwater robot obtains multi-modal local environmental information through its own sensors (such as sonar, camera, inertial navigation, depth meter, current meter, etc.).

[0066] Step 2: Multimodal information fusion: The multimodal information sensed by the sensor is fused and the information is enhanced to facilitate subsequent feature value extraction.

[0067] Step 3: Feature extraction. The autonomous underwater robot extracts feature values ​​from the fused information. Feature extraction is divided into two categories: environmental feature extraction and neighboring autonomous underwater robot feature extraction.

[0068] Step 4: Autonomous underwater robot navigation decision-making based on embodied intelligence. Based on the embodied intelligence method, the environmental feature values ​​are learned to understand the environment; based on the embodied intelligence method, the movement of the neighboring autonomous underwater robot is inferred according to its feature values.

[0069] Step 5: Collaborative relationship learning. Each autonomous underwater robot learns collaborative relationships based on the environment understood by the embodied intelligence method and the inferred motion trajectory of the adjacent autonomous underwater robots, so that multiple autonomous underwater robots can achieve collaboration without communication and correct navigation decisions.

[0070] Step 6: Control command generation: The generated navigation decision is sent to the controller of each autonomous underwater robot for compilation, and control commands are generated to realize navigation control of the autonomous underwater robot.

[0071] In the embodiment, step 3 specifically includes the following steps:

[0072] Step 3.1: Extraction of environmental feature information. The present invention uses a convolutional neural network (CNN) to extract spatial features from the environmental representation. At the same time, a long short-term memory (LSTM) network is used to extract the time information corresponding to different spatial features from previous local observations to obtain environmental feature values ​​containing spatial features and time information. This helps the autonomous underwater robot understand the change of state and enhance its short-term memory of the environment.

[0073] Step 3.2: Local Neighborhood Robot Feature Extraction. First, a detection network is trained to mark the locations of neighboring autonomous underwater robots. The trained detection network consists of a CNN with a residual structure connected to an MLP (Marked Location Protocol) with a parallel architecture to extract high-dimensional features from image-like representations. An MLP performs a classification task to detect the number of n neighboring autonomous underwater robots (m1, m2, …, m ) visible at t = 0. n ), and the non-existent autonomous underwater robot is represented by (0,…,0). Another MLP outputs the relative positions of the n adjacent autonomous underwater robots (loc1,loc2,…,loc n ). The position of a non-existent robot is also represented by (0,…,0). By multiplying the number of adjacent autonomous underwater robots by the relative position, the marked position of the adjacent autonomous underwater robot at t=0 is obtained, that is, the eigenvalue of the adjacent autonomous underwater robot: (mloc1,mloc2,…,mloc m )=(m1,m2,…,m n )×(loc1,loc2,…,loc n ). The parallel structure improves coordinate position accuracy by integrating elements of the classification task, usually associated with the regression task. This approach effectively reduces the number of false detections of autonomous underwater vehicles due to inherent regression errors.

[0074] In the embodiment, in step 4, each autonomous underwater robot learns the environmental feature values ​​based on the embodied intelligent unit, understands the environment, and infers the motion trajectory of the adjacent autonomous underwater robot based on the feature values ​​of the adjacent autonomous underwater robot. The embodied intelligent unit includes an embodied decision module and a multimodal storage module. The embodied decision module understands the input environmental feature values ​​through a large prediction model, combines the understanding results with the past environmental features and the feature values ​​of the adjacent autonomous underwater robots stored in the multimodal storage module for learning, and finally uses the inference device to infer the motion trajectory of the adjacent autonomous underwater robot, and the decision maker plans the navigation path of each autonomous underwater robot; at the same time, the embodied decision module sends the current environmental features, autonomous underwater robot motion features, trajectory and other information to the multimodal storage module; the multimodal storage module classifies and stores the environmental features, autonomous underwater robot motion features, trajectory and other information made and understood in the embodied decision module, and provides the stored past environmental features, autonomous underwater robot motion features, trajectory and other information to the embodied decision module for decision-making at the next moment;

[0075] The embodied decision module includes an LLM, a learner, a reasoner and a decision maker;

[0076] LLM understands the environmental information and the actions of other autonomous underwater robots around it based on the environmental feature values, and sends the understood information to the learner and reasoner;

[0077] The learner uses the environmental features, motion features, and trajectory of the autonomous underwater robot understood by the LLM and the past environmental features, motion features, and trajectory of the autonomous underwater robot stored in the multimodal storage module to learn and make navigation plans;

[0078] The reasoner infers the further movement of other nearby autonomous underwater robots based on the navigation plan and the environmental characteristics, motion characteristics, and trajectory of the autonomous underwater robot understood by the LLM;

[0079] The decision maker makes specific navigation decisions based on the motion results of other nearby autonomous underwater robots deduced by the reasoner.

[0080] The specific steps include:

[0081] Step 4.1: The information perceived by the perceptron is fed into the LLM, and the generative artificial intelligence function based on the language large model effectively monitors and processes multimodal scenario inputs.

[0082] Step 4.2: The information understood by the language model is sent to the embodied learner. The learner is designed to match a specific combination of payloads and describe a set of feasible actions. When the autonomous underwater vehicle breaks down the plan into specific actions, the first step is to consult the past information stored in the multimodal storage module to determine the appropriate action for the plan. Subsequently, the selected action activates a predefined workflow to infer, process and determine the necessary parameters, effectively guiding the learner to select an action function. This action function indexes a specific operation flow. The process contains access to sensor results and operations on the learner, as well as actively calling specific sensors. After the process ends, the parameters of the action function are determined and then the final action operation is executed. Step 4.3: The autonomous underwater vehicle reasoner integrates multimodal data streams and temporal comparisons across sequence frames to inform its decision-making process, resulting in task-oriented planning. This plan is represented by "monitoring rapid environmental changes", which is translated into specific actions to maintain position or call other sensors to obtain more information, rather than the "imagined" actions of the reinforcement learning agent.

[0083] Step 4.4: Decision Maker Learns with Few Shots via Multimodal Memory. The AUV leverages the vast amount of sensory data and records in the multimodal information database. It estimates its current position by cross-referencing recently captured images in the multimodal memory database with previously encountered images, assessing changes in visual details. In addition, the AUV intelligently integrates temporal and auxiliary data to make real-time decisions.

[0084] Step 4.5: The decision maker exploits observations and actions hierarchically. Agents built on LLMs can absorb environmental inputs of varying sources and semantic nature in a hierarchical manner, reflecting the sequential focus of human attention, where certain stimuli prompt the prioritization of relevant information. By synthesizing data from different sources, the agent has the ability to infer and even proactively seek out complementary focused observations. Furthermore, the action repertoire of LLM-based agents is not limited to predefined choices or static responses. The act of interrogating a sensor can itself constitute an action, and the determination of the parameters of the action can occur through logical reasoning, rather than being limited to optimization via policy gradients.

[0085] Step 4.6: The decision maker compiles the action decision into commands that the controller can recognize and passes it to collaborative relationship learning.

[0086] Step 4.7: Store the autonomous underwater robot navigation information into the multimodal information storage module. A memory tagging strategy is used to store information such as environmental features, autonomous underwater robot motion features, and trajectories. The memory tagging strategy encodes the multimodal experience and its narrative into the multimodal storage module, which catalogs the various aspects of the multimodal episodic memory fragment and its corresponding tags in the environmental storage unit of the multimodal storage module. The tags are stored together with the autonomous underwater robot motion memory type in the autonomous underwater robot motion storage unit of the multimodal storage module to achieve efficient recall. After locating the tags of the multimodal plot record, the tags and their associated multimodal content are fed into the embodied decision module for processing.

[0087] In the embodiment, step 5 specifically comprises the following steps:

[0088] Step 5.1: Collaborative information aggregation of multiple autonomous underwater robots.

[0089] Based on the understanding of the spatial and temporal information of the adjacent environment by the embodied intelligent decision-making module and the derived motion trajectory of the autonomous underwater robot in the adjacent area, and the characteristics of the environment where communication is impossible, a single autonomous underwater robot can achieve the collaboration of multiple autonomous underwater robots through learning based on the positions of other autonomous underwater robots to avoid collisions during navigation. Specifically:

[0090] Multi-AUV collaboration is achieved by leveraging a graph attention network applied to graph representation. The relationship between the decision-making autonomous underwater robot and its neighboring autonomous underwater robots can be described as an undirected graph G. In the kth graph convolutional layer, the following message passing and information integration operations are performed:

[0091]

[0092] Among them, H kGraph convolution operation in layer k with non-linear activation function H k-1 Represents the graph convolution operation in the k-1th layer with a non-linear activation function A represents the connection matrix of the direct relationship between the autonomous underwater robot and its neighboring robots, Ψ k represents the learnable weight matrix of the kth graph convolution layer, and K represents the number of graph convolution layers. k is the trainable bias. a mn represents the value of the mth row and nth column of the connection matrix A, LR represents the leaky rectified linear unit activation function, and δ represents the trainable attention kernel. (H k Ψ k ) m Represents H k Ψ k The mth row of (H k Ψ k ) n Represents H k Ψ k The nth column of (H k Ψ k ) l Represents H k Ψ k In the lth column of , exp(·) represents the exponential operation.

[0093] Step 5.2: Multi-robot cooperative strategy learning. This paper proposes a new continuous reward function that integrates expert guidance to guide the robot to the local endpoint while considering potential collisions. The reward design promotes efficient actions by balancing environmental avoidance and endpoint proximity, thereby improving navigation efficiency. The reward function r t It is expressed as:

[0094] r t = s r t + d r t + p r t (4)

[0095] in s r t Represents the direction reward, guiding the autonomous underwater robot to learn to turn in the correct direction, and its expression is:

[0096] s r t =β s (d t -d t-1 ) (5)

[0097] where d trepresents the yaw angle between the current position of the autonomous underwater vehicle and the position of its target at time t. t-1 represents the yaw angle between the current position of the autonomous underwater vehicle and the position of its target at time t-1. s represents a hyperparameter.

[0098] d r t Represents the collision reward. In order to facilitate the autonomous underwater robot to learn to avoid obstacles, the learning process penalizes the state of the autonomous underwater robot approaching obstacles. d r t Defined as

[0099]

[0100] where p t represents the position of the autonomous underwater robot at time t, L d and β d Represents the position of the obstacle, the safety distance and the collision penalty factor respectively. collsion Indicates the penalty coefficient for collision, which is a constant, L c Indicates the distance to the obstacle. The distance between the autonomous underwater robot and the obstacle is less than L c The vehicle is considered to have collided with an obstacle. Based on the safety distance L d The gradient reward function avoids the rigid expansion of the obstacle collision volume. This enables the autonomous underwater robot to have a larger range of motion, thereby achieving more flexible actions based on the physical environment.

[0101] p r t Represents navigation reward. It aims to motivate the autonomous underwater robot to quickly approach the local endpoint using the shortest possible action through rewards. p r t It is expressed as:

[0102]

[0103] Where g represents the location of the target point, p t represents the position of the autonomous underwater robot at time t, p t-1 represents the position of the autonomous underwater robot at time t-1, β g Indicates the penalty factor for not reaching the target point, r goal Represents the reward for reaching the target point, which is a constant, L g Indicates the distance to the target point. The distance between the autonomous underwater robot and the target point is less than L g Take charge to reach the target point.

[0104] It will be apparent to those skilled in the art that the invention is not limited to the details of the exemplary embodiments described above and that the invention can be implemented in other specific forms without departing from the spirit or essential features of the invention. Therefore, the embodiments should be considered exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description, and it is intended that all variations falling within the meaning and scope of the equivalent elements of the claims be included in the invention. Any reference numeral in a claim should not be considered as limiting the claim to which it relates.

[0105] In addition, it should be understood that although the present specification is described according to implementation modes, not every implementation mode contains only one independent technical solution. This description of the specification is only for the sake of clarity. Those skilled in the art should regard the specification as a whole. The technical solutions in each embodiment may also be appropriately combined to form other implementation modes that can be understood by those skilled in the art.

Claims

1. A multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment, characterized in that: The steps include: Step 1: Each autonomous underwater robot obtains multimodal local environment information through its own sensors; Step 2: Fusing the corresponding multi-modal local environment information of each autonomous underwater robot; Step 3: Each autonomous underwater robot extracts environmental feature values ​​and feature values ​​of neighboring autonomous underwater robots based on the fused local environmental information; Step 4: Each autonomous underwater robot learns the environmental feature values ​​based on the embodied intelligent unit, understands the environment, and infers the motion trajectory of the neighboring autonomous underwater robot based on the feature values ​​of the neighboring autonomous underwater robot; Step 5: Each autonomous underwater robot learns the collaborative relationship based on the environment understood by the embodied intelligent unit and the inferred motion trajectory of the adjacent autonomous underwater robot, so that multiple autonomous underwater robots can achieve collaboration without communication conditions, and correct the generated navigation decision; Step 6: Send the generated navigation decision to the controller of each autonomous underwater robot for compilation, generate control instructions, and realize navigation control of the autonomous underwater robot; The step 5 specifically comprises the following steps: Step 5.1: Collaborative information aggregation of multiple autonomous underwater robots; Step 5.2: Multi-robot cooperation strategy learning based on aggregated information; The collaborative information aggregation of multiple autonomous underwater robots described in step 5.1 is achieved by utilizing a graph attention network applied to graph representation, specifically: Based on the understanding of the spatial and temporal information of the adjacent environment by the embodied intelligent decision-making module and the derived motion trajectory of the autonomous underwater robot in the adjacent area, according to the characteristics of the environment where communication is impossible, a single autonomous underwater robot achieves the collaboration of multiple autonomous underwater robots through learning based on the positions of other autonomous underwater robots to avoid collisions during navigation. Specifically, the relationship between the decision-making autonomous underwater robot and its adjacent autonomous underwater robots is described as an undirected graph G. In the kth graph convolutional layer, the following message passing and information aggregation operations are performed: Among them, H k Graph convolution operation in layer k with non-linear activation function H k-1 Represents the graph convolution operation in the k-1th layer with a non-linear activation function A represents the connection matrix of the direct relationship between the autonomous underwater robot and its neighboring robots, Ψ k represents the learnable weight matrix of the kth graph convolution layer, K represents the number of graph convolution layers, and γ k is the trainable bias, a mn represents the value of the mth row and nth column of the connection matrix A, LR represents the leaky rectifier linear unit activation function, δ represents the trainable attention kernel, (H k Ψ k ) m Represents H k Ψ k The mth row of (H k Ψ k ) n Represents H k Ψ k The nth column of (H k Ψ k ) l Represents H k Ψ k In the lth column of , exp(·) indicates exponential operation; The reward function r used in step 5.2 of cooperative strategy learning t for: r t ＝ s r t + d r t + p r t (4) in s r t Represents direction reward, d r t Represents collision reward, p r t Represents navigation reward; Said s r t The expression is: s r t =b s (d t -d t-1 ) (5) where d t represents the yaw angle between the current position of the autonomous underwater robot and the position of its target at time t, d t-1 represents the yaw angle between the current position of the autonomous underwater robot and the position of its target at time t-1, β s represents a hyperparameter; Said d r t The expression is: where p t represents the position of the autonomous underwater robot at time t, L d and β d Represent the obstacle position, safety distance and collision penalty factor, r collsion Indicates the penalty coefficient for collision, L c Indicates the distance to the obstacle; Said p r t The expression is: Where g represents the location of the target point, p t represents the position of the autonomous underwater robot at time t, p t-1 represents the position of the autonomous underwater robot at time t-1, β g Indicates the penalty factor for not reaching the target point, r goal Represents the reward for reaching the target point, L g Indicates the distance to the target point.

2. According to claim 1, a multi-autonomous underwater robot navigation method based on embodied intelligence in a non-communication environment is characterized in that: The step 3 uses a convolutional neural network (CNN) to extract spatial features from the fused local environmental information. At the same time, a long short-term memory (LSTM) network is used to extract time information corresponding to different spatial features from local observations to obtain environmental feature values ​​containing spatial features and time information.

3. The method for multi-autonomous underwater robot navigation based on embodied intelligence in a non-communication environment according to claim 1, characterized in that: The extraction strategy of the feature value of the neighboring autonomous underwater robot in step 3 is: First, a training detection network is used to mark the positions of adjacent autonomous underwater robots and detect their relative positions; The training detection network consists of a cellular neural network with a residual structure connected to an MLP with a parallel architecture to extract high-dimensional features from an image-like representation; An MLP performs the classification task to detect the number of n adjacent autonomous underwater vehicles (m1,m2,…,m n ), the non-existent autonomous underwater robot is represented by (0,…,0); Another MLP outputs the relative positions of n adjacent autonomous underwater robots (loc1, loc2, ..., loc n ), the non-existent robot position is also represented by (0,…,0); Secondly, by multiplying the number of adjacent autonomous underwater robots by the relative position, the marked position of the adjacent autonomous underwater robot at t = 0 is obtained, that is, the eigenvalue of the adjacent autonomous underwater robot: (mloc1, mloc2, ..., mloc m )=(m1,m2,…,m n )×(loc1,loc2,…,loc n ).

4. The method for multi-autonomous underwater robot navigation based on embodied intelligence in a non-communication environment according to claim 1, characterized in that: The embodied intelligent unit in step 4 includes an embodied decision-making module and a multimodal storage module. The embodied decision-making module understands the input environmental feature values ​​through a large prediction model, and learns the results of the understanding in combination with the past environmental features and feature values ​​of nearby autonomous underwater robots stored in the multimodal storage module. Finally, the motion trajectory of the nearby autonomous underwater robots is inferred by the inference device, and the decision maker plans the navigation path of each autonomous underwater robot. At the same time, the embodied decision-making module sends the current environmental features, motion features of the autonomous underwater robot, and trajectory information to the multimodal storage module. The multimodal storage module classifies and stores the environmental features, motion features of the autonomous underwater robot, and trajectory information that are decided and understood in the embodied decision-making module, and provides the stored past environmental features, motion features of the autonomous underwater robot, and trajectory information to the embodied decision-making module for decision-making at the next moment.

5. The method for multi-autonomous underwater robot navigation based on embodied intelligence in a non-communication environment according to claim 4, characterized in that: The embodied decision module includes an LLM, a learner, a reasoner and a decision maker; LLM understands the environmental information and the actions of other autonomous underwater robots around it based on the environmental feature values, and sends the understood information to the learner and reasoner; The learner uses the environmental features, motion features, and trajectory information of the autonomous underwater robot understood by the LLM and the past environmental features, motion features, and trajectory information stored in the multimodal storage module to learn and make navigation plans; The reasoner infers the further movement of other nearby autonomous underwater robots based on the environmental features, motion features and trajectory information understood by the navigation plan and LLM; The decision maker makes specific navigation decisions based on the motion results of other nearby autonomous underwater robots deduced by the reasoner.

6. The method for multi-autonomous underwater robot navigation based on embodied intelligence in a non-communication environment according to claim 4, characterized in that: The multimodal storage module adopts a memory tagging strategy to store environmental features, autonomous underwater robot motion features, and trajectory information. The memory tagging strategy encodes the multimodal experience and its narrative into the multimodal storage module, which catalogs the various aspects of data included in the multimodal scenario memory fragment and its corresponding labels in the environmental storage unit of the multimodal storage module. The label is stored together with the autonomous underwater robot motion memory type in the autonomous underwater robot motion storage unit of the multimodal storage module to achieve efficient recall; after locating the label of the multimodal plot record, the label and its associated multimodal content are fed to the embodied decision module.

Images