A Zero-Sample Target Navigation Method and System Based on Landmark Memory

Abstract

This invention relates to the field of target navigation technology, providing a zero-shot target navigation method and system based on landmark memory. The zero-shot target navigation method based on landmark memory includes: identifying landmarks within the current field of view at each time step of the navigation task; inferring the scene category label to which the landmark belongs by combining prior scene knowledge; constructing a landmark memory library to form a sparse topological map and extracting landmark feature vectors from it; forming an inference vector using a spatiotemporal attention mechanism; and then using gating coefficients to weightedly fuse the inference vector and the landmark feature vector to obtain a discriminative state representation for decision-making; and generating robot control commands based on the discriminative state representation for decision-making and a deep reinforcement learning framework. This method can both lightweightly construct the environmental topology and dynamically bind visual landmarks with semantic concepts, thereby supporting an explicit memory mechanism for logical reasoning and efficient exploration.

Description

Technical Field

[0001] This invention relates to the field of target navigation technology, and in particular to a zero-sample target navigation method and system based on landmark memory. Background Technology

[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.

[0003] Target navigation is a key task for robot autonomy, requiring agents to locate specific target objects in unknown environments relying solely on visual input. Traditional navigation systems are mostly based on Simultaneous Localization and Mapping (SLAM) technology. While these methods can build accurate metric maps, they often lack high-level semantic understanding capabilities. In recent years, end-to-end deep reinforcement learning methods have become mainstream because they can directly map visual perception into actions.

[0004] However, existing deep reinforcement learning-based navigation methods face severe challenges in zero-shot object navigation tasks. Zero-shot navigation requires agents to locate targets belonging to novel categories without any training samples. Existing technologies mainly have the following limitations: (1) Most existing deep reinforcement learning architectures rely on the hidden states of recurrent neural networks to store environmental information. This implicit memory representation is opaque, causing agents to tend to overfit specific spatial layouts and object distributions in the training set, rather than learning generalizable topological logic; and due to the overfitting problem, when faced with unknown environments or unknown targets, agents often rely on familiar training environment contexts and cannot effectively utilize general semantic prior knowledge for logical reasoning. (2) Although semantic embedding using pre-trained models such as CLIP has been successfully applied to zero-shot recognition in the field of computer vision, in the field of navigation, existing methods often focus on the generalization of target recognition, but fail to effectively combine this semantic generalization ability with the rapid adaptation ability of spatial structure, resulting in low exploration efficiency when searching for new targets. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a zero-sample target navigation method and system based on landmark memory, which can improve robot navigation performance.

[0006] To achieve the above objectives, the present invention adopts the following technical solution: The first aspect of the present invention provides a zero-sample target navigation method based on landmark memory.

[0007] In one or more embodiments, a zero-sample target navigation method based on landmark memory is provided, comprising: At each time step of the navigation task, the visual observations of the current field of view and the semantic category labels of objects in the field of view are received and encoded into feature vectors. Landmarks are identified from the visual observations of the current field of view, and then the scene category label to which the landmark belongs is inferred by combining the scene prior knowledge. Based on the scene category label of the landmark, the visual feature representation of the landmark location, and the semantic embedding associated with the landmark, an explicit externalized landmark memory bank is constructed to mimic the cognitive map mechanism of biological navigation, capture, store and dynamically update key locations in the environment online, form a sparse topological map and extract landmark feature vectors from it. Spatiotemporal cross-attention mechanism is used to extract spatiotemporal context vectors from local observation images and short-term historical path information by combining landmark memory. Target-oriented context vectors are extracted from landmark memory by using target perception attention mechanism and semantic embedding vectors of target objects, thus forming inference vectors. Then, gating coefficients are used to weight and fuse the inference vectors and landmark feature vectors to obtain a discriminative state representation for decision-making. Robot control commands are generated based on discriminative state representations used for decision-making and a deep reinforcement learning framework.

[0008] As one implementation method, the construction of the landmark memory bank is an online incremental process, which is as follows: At each step, the cosine similarity between the visual features of the detected objects in the field of view and the visual features of all the landmark nodes stored in the landmark memory bank is calculated, and the maximum similarity value is found. The maximum similarity value is compared with a preset novelty threshold to identify a new, representative object and write it into the landmark memory bank as a landmark point.

[0009] As one implementation method, during the construction of the landmark memory database, a gated update mechanism is used to perform feature fusion updates on the visited landmark nodes:

[0010] in, New observational features at the current moment; It is an adaptive update gate whose value is determined by the access confidence of the landmark; The landmark node at time t+1; Let be the landmark node at time t.

[0011] As one implementation method, the access confidence level of each landmark node is set to... For landmark nodes that are not visited or matched at the current moment, their confidence level is reduced according to a preset decay rate. Exponential decay: t represents time; when the confidence level of a landmark node is lower than the preset pruning threshold, the node is removed from the landmark memory.

[0012] As one implementation method, the process of extracting the spatiotemporal context vector is as follows: Use the current LSTM hidden state as the query vector. This includes encoding historical action trajectories and short-term time context; The landmark feature vectors corresponding to the landmark memory bank As keys and values; Calculate query vector Landmark feature vector The correlation generates a spatiotemporal context vector. :

[0013] in, This is the scaling factor; This is the transpose of the matrix.

[0014] As one implementation method, the process of extracting the target-oriented context vector is as follows: The semantic embedding vector of the target object is broadcast and concatenated with each landmark feature vector in the landmark memory to form a combined feature. A multilayer perceptron is used to process the combined features, learn the latent semantic associations between landmarks and targets, and then... Function normalization yields attention weights; The calculated attention weights are used to perform a weighted summation of the landmark feature vectors to obtain the target-oriented context vector.

[0015] As one implementation method, in the process of generating a discriminative state representation for decision-making, the spatiotemporal context vector and the goal-oriented context vector are added together to form a unified inference vector; a learnable gating coefficient is used to weight and fuse the inference vector and the landmark feature vector to obtain the discriminative state representation for decision-making. :

[0016]

[0017] in, This indicates element-wise multiplication; The landmark feature vector corresponding to the landmark memory database; The gating factor; for Activation function This is the weight matrix; This is the inference vector.

[0018] A second aspect of the present invention provides a zero-sample target navigation system based on landmark memory.

[0019] In one or more embodiments, a zero-sample target navigation system based on landmark memory includes: The scene category label inference module is used to receive the visual observation of the current field of view and the semantic category labels of objects in the field of view at each time step of the navigation task, encode them into feature vectors, identify landmarks from the visual observation of the current field of view, and then infer the scene category label to which the landmark belongs by combining the scene prior knowledge. The landmark memory bank construction module is used to build an explicit externalized landmark memory bank based on the scene category label to which the landmark belongs, the visual feature representation of the landmark location, and the semantic embedding associated with the landmark. This mimics the cognitive mapping mechanism of biological navigation, captures, stores, and dynamically updates key locations in the environment online, forms a sparse topological map, and extracts landmark feature vectors from it. The discriminative state representation generation module is used to extract spatiotemporal context vectors from local observation images and short-term historical path information by combining a spatiotemporal cross-attention mechanism with a landmark memory bank. It also uses a target perception attention mechanism and the semantic embedding vector of the target object to extract a target-oriented context vector from the landmark memory bank, thereby forming an inference vector. Finally, it uses a gating coefficient to weight and fuse the inference vector and the landmark feature vector to obtain a discriminative state representation for decision-making. The robot control command generation module generates robot control commands based on the discriminative state representation used for decision-making and a deep reinforcement learning framework.

[0020] A third aspect of the present invention provides a computer-readable storage medium.

[0021] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the steps of the zero-sample target navigation method based on landmark memory as described above.

[0022] A fourth aspect of the present invention provides an electronic device.

[0023] A robot includes a robot body, a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the zero-sample target navigation method based on landmark memory as described above.

[0024] Compared with the prior art, the beneficial effects of the present invention are: This invention proposes a zero-shot target navigation method and system based on landmark memory. By constructing an explicitly externalized landmark memory bank, mimicking the cognitive mapping mechanism of biological navigation, it captures, stores, and dynamically updates key locations in the environment online, significantly improving the generalization ability of zero-shot navigation. It uses a spatiotemporal attention mechanism to extract spatiotemporal context vectors and target-oriented context vectors, and finally uses weighted fusion of gating coefficients to determine the discriminative state representation for decision-making. This enables the robot to no longer blindly explore randomly, but to have planning and association abilities similar to humans, and to efficiently locate the target using semantic priors. Attached Figure Description

[0025] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0026] Figure 1 This is a flowchart of the zero-sample target navigation method based on landmark memory according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the spatiotemporal attention mechanism structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the zero-sample target navigation system based on landmark memory according to an embodiment of the present invention. Detailed Implementation

[0027] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0028] It should be noted that the following detailed description is illustrative and intended to provide further explanation of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0029] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0030] As the background technology indicates, existing navigation methods based on deep reinforcement learning suffer from the high computational cost of existing structured map methods. While some research has attempted to introduce structured neural maps or graph-based methods to enhance memory capabilities, these methods typically involve high computational overhead or construct overly dense scene graphs, making it difficult to meet the demands of real-time, efficient navigation. In summary, current technologies lack an explicit memory mechanism that can both lightweightly construct environmental topology and dynamically bind visual landmarks with semantic concepts, thereby supporting logical reasoning and efficient exploration.

[0031] Figure 1 A schematic diagram of the zero-sample target navigation method based on landmark memory, according to an embodiment of the present invention, is provided. Figure 1 The zero-sample target navigation method based on landmark memory in this embodiment may include the following steps S101 to S104.

[0032] The specific implementation process of steps S101 to S104 is as follows: Step S101: At each time step of the navigation task, receive the visual observations of the current field of view and the semantic category labels of objects in the field of view and encode them into feature vectors. Identify landmarks from the visual observations of the current field of view and then infer the scene category label to which the landmarks belong by combining the scene prior knowledge.

[0033] At each time step of the navigation task As an intelligent agent, the robot receives current egocentric visual observations. and semantic category description of the target object .

[0034] To achieve generalization under zero-shot conditions, the unstructured visual input first needs to be transformed into a structured feature representation. This is done by using a pre-trained deep convolutional neural network, ResNet-50, as the visual encoder to process the observed images. Processing is performed to extract high-dimensional visual feature vectors. This feature vector captures the scene spatial layout information from the current viewpoint.

[0035] Classifying target objects using the pre-trained text vector model GloVe Mapped to semantic embedding vector This embedding vector encodes the conceptual features of the target object in the semantic space, enabling visually unseen objects to be associated through semantic similarity.

[0036] The object detection model identifies salient objects within the current field of view and, combined with prior scene knowledge, infers the current scene category label. (For example, bedrooms, living rooms, etc.), which are encoded as one-hot vectors. The object detection model here can be selected according to the specific situation, and will not be detailed here.

[0037] Step S102: Based on the scene category label of the landmark, the visual feature representation of the landmark location, and the semantic embedding associated with the landmark, construct an explicit externalized landmark memory bank to mimic the cognitive map mechanism of biological navigation, capture, store and dynamically update key locations in the environment online, form a sparse topological map and extract landmark feature vectors from it.

[0038] To overcome the gradient vanishing and information forgetting problems inherent in traditional recurrent neural networks during long-sequence navigation, this invention designs an explicit, externalized landmark memory. .

[0039] Landmark Memory Bank Each landmark node in Defined as a triple containing multimodal information: .in, The visual features representing the location of the landmark are used for subsequent visual relocation matching; Represents the semantic embedding associated with the landmark, used to establish semantic associations; This represents the scene category to which the landmark belongs and is used to provide topological constraints for higher levels.

[0040] The construction of the landmark memory bank is an online incremental process, which is as follows: At every step Calculate the visual features of objects detected in the field of view Visual features of all landmark nodes already stored in the landmark memory bank Calculate the cosine similarity between them and find the maximum similarity value. ; ; The maximum similarity value is compared with the preset novelty threshold. The comparison is used to identify a new, representative object and write it into the landmark memory bank as a landmark.

[0041] if This location is then identified as a new, representative environmental location. At this point, the system initializes a new landmark node. Write it into the memory bank This mechanism ensures the sparsity of the memory bank, avoiding the storage of a large number of redundant frames and retaining only the key topology nodes.

[0042] Landmark feature update based on gating mechanism: To improve the robustness of landmark representation, a gating update mechanism is used to perform feature fusion updates on visited landmark nodes during the construction of the landmark memory database.

[0043] in, New observational features at the current moment; It is an adaptive update gate whose value is determined by the access confidence of the landmark; The landmark node at time t+1; Let be the landmark node at time t.

[0044] Landmark feature updates based on gating mechanisms allow the model to adaptively balance the stability of historical memory with the plasticity of new observations, making landmark features gradually more stable and accurate with repeated visits.

[0045] To control the size of the memory bank and simulate the forgetting characteristics of biological memory, the access confidence of each landmark node is maintained.

[0046] The access confidence level for each landmark node is set to For landmark nodes that are not visited or matched at the current moment, their confidence level is reduced according to a preset decay rate. Exponential decay: t represents time; when the confidence level of a landmark node falls below a preset pruning threshold, the node is considered outdated information that is no longer important and is thus removed from memory. Remove from the topology to ensure that the memory always retains the information most relevant to the current navigation task.

[0047] Step S103: Using a spatiotemporal cross-attention mechanism combined with a landmark memory bank, spatiotemporal context vectors are extracted from local observation images and short-term historical path information. Using a target perception attention mechanism and the semantic embedding vector of the target object, target-oriented context vectors are extracted from the landmark memory bank to form inference vectors. Then, the inference vectors and landmark feature vectors are weighted and fused using gating coefficients to obtain a discriminative state representation for decision-making.

[0048] To efficiently utilize this information in complex navigation tasks, such as Figure 2 As shown, this invention proposes a spatiotemporal attention mechanism, which employs a two-stream architecture, including a cross-attention mechanism and a target perception attention mechanism. These mechanisms handle long-term temporal spatial dependencies and target semantic relevance respectively, and are finally adaptively fused through gating units. Traditional LSTM networks can only capture short-term temporal dependencies and struggle to utilize global spatial information. This invention utilizes the cross-attention mechanism to align the robot's current short-term temporal state with its global long-term spatial memory.

[0049] The process of extracting the spatiotemporal context vector is as follows: Use the current LSTM hidden state as the query vector. This includes encoding historical action trajectories and short-term time context; The landmark feature vectors corresponding to the landmark memory bank As keys and values; where The number of landmarks in the memory bank. For feature dimensions; Calculate query vector Landmark feature vector The correlation generates a spatiotemporal context vector. :

[0050] in, This is the scaling factor; This is the transpose of the matrix. Through this step, the model can automatically focus on historical landmarks that are spatially adjacent to or topologically related to the current location, achieving localization and loop closure detection.

[0051] Unlike traditional methods that rely solely on visual similarity to find targets, this approach introduces semantic reasoning to explicitly identify landmarks that co-occur with the target object semantically.

[0052] The process of extracting the target-oriented context vector is as follows: First, embed the semantic vector of the target object. Broadcast and compare with each landmark feature vector in the memory bank. By splicing together, a combined feature is formed. ; Then, a multilayer perceptron is used to process the combined features, learn the latent semantic association between landmarks and targets, and through... Attention weights are obtained by function normalization. : ; Finally, the calculated weights are used to perform a weighted summation of the landmark memories to obtain the goal-oriented context vector. :

[0053] Among them, This represents the number of landmarks in the memory bank.

[0054] This mechanism allows the robot to prioritize exploring areas containing potentially relevant objects, rather than blindly searching randomly.

[0055] In order to dynamically integrate the two inference results mentioned above and preserve the integrity of the original environmental features, this invention designs a gating fusion mechanism. In the process of generating the discriminative state representation for decision-making, the spatiotemporal context vector and the target-oriented context vector are added together to form a unified inference vector. ; Calculate a learnable gating coefficient This is used to balance the preservation of original memory features with the utilization of reasoning context: By using a learnable gating coefficient to weight and fuse the inference vector and the landmark feature vector, a discriminative state representation for decision-making is obtained. :

[0056]

[0057] in, This indicates element-wise multiplication; The landmark feature vector corresponding to the landmark memory database; The gating factor; for Activation function This is the weight matrix; This is the inference vector.

[0058] The discriminative state representation used for decision-making includes both global spatial positioning information and semantic navigation cues for the current target.

[0059] Step S104: Generate robot control instructions based on the discriminative state representation used for decision-making and the deep reinforcement learning framework.

[0060] After the discriminative state representation used for decision-making, specific robot control instructions are generated through a deep reinforcement learning framework.

[0061] The deep reinforcement learning framework here uses the A3C architecture as the decision core.

[0062] Will The input is fed into an A3C network, which contains two output headers: The action policy head is responsible for outputting the probability distribution in the action space. ;in, Represents the action space; Representing the state space; This indicates the probability of outputting a certain action under a given state.

[0063] The motion space includes: ; The value estimation head is responsible for outputting the value estimate of the current state, which is used to assist in the calculation of the advantage function during the training process, thereby reducing the variance of the policy gradient and accelerating convergence.

[0064] in, These respectively represent moving forward, turning left, turning right, looking up, looking down, and ending.

[0065] Finally, the action at the current moment is sampled based on the action probability distribution. And it is executed by the robot. If the action to be executed is If the target object is within the field of view and meets the distance threshold (e.g., 1.5 meters), then navigation is considered successful.

[0066] To enable the model to adapt to unknown environments and targets, an A3C-based parallel training framework is employed. Empirical data is collected simultaneously across multiple parallel environment instances, and asynchronous gradient updates are used to optimize network parameters. The training objective is to maximize long-term cumulative reward while minimizing navigation path length, enabling the robot to learn a navigation strategy that successfully locates the target while maintaining path efficiency.

[0067] In other embodiments, the deep reinforcement learning framework may also be implemented using other existing frameworks, which will not be detailed here.

[0068] Compared with the prior art, the present invention has the following significant advantages: Significantly improves the generalization ability of zero-shot navigation: By constructing explicit landmark memory, this invention overcomes the drawback of traditional implicit memory, which is prone to overfitting to the training environment layout features. Experiments demonstrate that, in completely unseen object categories and unseen scenes, the method of this invention significantly outperforms existing state-of-the-art models in both success rate and path efficiency, proving its powerful transfer and generalization capabilities.

[0069] This invention achieves semantic association-based logical reasoning navigation: by introducing "target-aware attention flow" from the spatiotemporal attention mechanism, it can retrieve semantically related landmarks from a memory bank using the semantic embedding of target objects. This mechanism enables the robot to move beyond blind, random exploration and instead possess human-like planning and associative abilities, allowing it to efficiently locate targets using semantic priors.

[0070] A lightweight and robust topological cognitive map is constructed: This invention employs online mapping based on novelty and a forgetting mechanism based on confidence, ensuring the sparsity and efficiency of the memory bank. Compared to dense metric maps or computationally expensive scene graph methods, the method of this invention not only has lower computational overhead but also enhances robustness to changes in ambient lighting and viewpoint through the fusion of multimodal features.

[0071] The system solves the challenge of integrating long-term spatial memory with short-term temporal reasoning: by using a spatiotemporal cross-attention mechanism, it successfully aligns the robot's current instantaneous state with its global historical trajectory, solving the problem of getting lost during long-distance navigation and improving the robot's positioning accuracy and loop closure detection capability in complex indoor environments.

[0072] like Figure 3 As shown, the zero-sample target navigation system based on landmark memory provided in this embodiment of the invention can be implemented in software. The zero-sample target navigation system based on landmark memory includes the following software modules: scene category label inference module 301, landmark memory database construction module 302, discriminative state representation generation module 303, and robot control command generation module 304.

[0073] The functions of each software module in the landmark-based zero-sample target navigation system are described below: The scene category label inference module 301 is used to receive the visual observation of the current field of view and the semantic category label of the object in the field of view at each time step of the navigation task, encode and convert them into feature vectors, identify landmarks from the visual observation of the current field of view, and then infer the scene category label to which the landmark belongs by combining the scene prior knowledge. The landmark memory bank construction module 302 is used to construct an explicit externalized landmark memory bank based on the scene category label to which the landmark belongs, the visual feature representation of the landmark location, and the semantic embedding associated with the landmark. This mimics the cognitive map mechanism of biological navigation, captures, stores, and dynamically updates key locations in the environment online, forms a sparse topological map, and extracts landmark feature vectors from it. The discriminative state representation generation module 303 is used to extract spatiotemporal context vectors from local observation images and short-term historical path information by combining a spatiotemporal cross-attention mechanism with a landmark memory bank, extract target-oriented context vectors from the landmark memory bank by using a target perception attention mechanism and the semantic embedding vector of the target object, and then form an inference vector. The inference vector and the landmark feature vector are then weighted and fused using a gating coefficient to obtain a discriminative state representation for decision-making. The robot control instruction generation module 304 is used to generate robot control instructions based on the discriminative state representation used for decision-making and the deep reinforcement learning framework.

[0074] It should be noted that each module in the landmark memory-based zero-sample target navigation system of this invention corresponds one-to-one with each step in the landmark memory-based zero-sample target navigation method in the above embodiments, and their specific implementation processes are the same, so they will not be repeated here.

[0075] In one or more embodiments, a robot is also provided, including a robot body and at least one processor, memory, user interface, and at least one network interface mounted on the robot body. The various components in the landmark-memory-based zero-sample target navigation system are coupled together via a bus system. It is understood that the bus system is used to enable communication between these components. In addition to a data bus, the bus system also includes a power bus, a control bus, and a status signal bus. The user interface may include a display, keyboard, mouse, trackball, click wheel, buttons, a touchpad, or a touchscreen, etc.

[0076] It is understood that the memory can be volatile memory or non-volatile memory, or both. The memory in this embodiment of the invention is capable of storing data to support the operation of the terminal. Examples of this data include any computer programs used to operate on the terminal, such as operating systems and applications. The operating system includes various system programs, such as the framework layer, core library layer, driver layer, etc., used to implement various basic services and handle hardware-based tasks. Applications can include various applications.

[0077] In some embodiments, the landmark-memory-based zero-shot target navigation system provided in this invention can be implemented using a combination of hardware and software. For example, the landmark-memory-based zero-shot target navigation system provided in this invention can be a processor in the form of a hardware decoding processor, programmed to execute the landmark-memory-based zero-shot target navigation method provided in this invention. For instance, the hardware decoding processor can employ one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components.

[0078] As an example, a processor can be an integrated circuit chip with signal processing capabilities, such as a general-purpose processor, a digital signal processor (DSP), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc., where a general-purpose processor can be a microprocessor or any conventional processor, etc.

[0079] As an example of the hardware implementation of the zero-sample target navigation system based on landmark memory provided in this embodiment of the invention, the device provided in this embodiment of the invention can be directly executed by a processor in the form of a hardware decoding processor. For example, it can be executed by one or more application-specific integrated circuits (ASICs), DSPs, programmable logic devices (PLDs), complex programmable logic devices (CPLDs), field-programmable gate arrays (FPGAs), or other electronic components to implement the zero-sample target navigation method based on landmark memory provided in this embodiment of the invention.

[0080] The memory in this embodiment of the invention is used to store various types of data to support the operation of a landmark-memory-based zero-sample target navigation system, or to store data for execution. Figure 1 The program code for the method shown. Examples of this data include: any executable instructions for operation on a landmark memory-based zero-sample target navigation system, such as executable instructions that can be included in the executable instructions to implement the landmark memory-based zero-sample target navigation method of the embodiments of the present invention.

[0081] Specifically, according to embodiments of this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program including functions for executing... Figure 1 The program code for the method shown. In such an embodiment, the computer program can be downloaded and installed from a network via a communication component, and / or installed from a removable medium. When the computer program is executed by the central processing unit, it performs the various functions defined in the apparatus of this application.

[0082] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart. Figure 1 One or more processes and / or boxes Figure 1A device that provides the functions specified in one or more boxes.

[0083] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A zero-sample target navigation method based on landmark memory, characterized in that, include: At each time step of the navigation task, the visual observations of the current field of view and the semantic category labels of objects in the field of view are received and encoded into feature vectors. Landmarks are identified from the visual observations of the current field of view, and then the scene category label to which the landmark belongs is inferred by combining the scene prior knowledge. Based on the scene category label of the landmark, the visual feature representation of the landmark location, and the semantic embedding associated with the landmark, an explicit externalized landmark memory bank is constructed to mimic the cognitive map mechanism of biological navigation, capture, store and dynamically update key locations in the environment online, form a sparse topological map and extract landmark feature vectors from it. Spatiotemporal cross-attention mechanism is used to extract spatiotemporal context vectors from local observation images and short-term historical path information by combining landmark memory. Target-oriented context vectors are extracted from landmark memory by using target perception attention mechanism and semantic embedding vectors of target objects, thus forming inference vectors. Then, gating coefficients are used to weight and fuse the inference vectors and landmark feature vectors to obtain a discriminative state representation for decision-making. Robot control commands are generated based on discriminative state representations used for decision-making and a deep reinforcement learning framework.

2. The zero-sample target navigation method based on landmark memory as described in claim 1, characterized in that, The construction of the landmark memory bank is an online incremental process, which is as follows: At each step, the cosine similarity between the visual features of the detected objects in the field of view and the visual features of all the landmark nodes stored in the landmark memory bank is calculated, and the maximum similarity value is found. The maximum similarity value is compared with a preset novelty threshold to identify a new, representative object and write it into the landmark memory bank as a landmark point.

3. The zero-sample target navigation method based on landmark memory as described in claim 1, characterized in that, During the construction of the landmark memory bank, a gating update mechanism is used to perform feature fusion updates on the visited landmark nodes: in, New observational features at the current moment; It is an adaptive update gate whose value is determined by the access confidence of the landmark; The landmark node at time t+1; Let be the landmark node at time t.

4. The zero-sample target navigation method based on landmark memory as described in claim 3, characterized in that, The access confidence level for each landmark node is set to For landmark nodes that are not visited or matched at the current moment, their confidence level is reduced according to a preset decay rate. Exponential decay: t represents time; when the confidence level of a landmark node is lower than the preset pruning threshold, the node is removed from the landmark memory.

5. The zero-sample target navigation method based on landmark memory as described in claim 1, characterized in that, The process of extracting the spatiotemporal context vector is as follows: Use the current LSTM hidden state as the query vector. This includes encoding historical action trajectories and short-term time context; The landmark feature vectors corresponding to the landmark memory bank As keys and values; Calculate query vector Landmark feature vector The correlation generates a spatiotemporal context vector. : in, This is the scaling factor; This is the transpose of the matrix.

6. The zero-sample target navigation method based on landmark memory as described in claim 1, characterized in that, The process of extracting the target-oriented context vector is as follows: The semantic embedding vector of the target object is broadcast and concatenated with each landmark feature vector in the landmark memory to form a combined feature. A multilayer perceptron is used to process the combined features, learn the latent semantic association between landmarks and targets, and then... Function normalization yields attention weights; The calculated attention weights are used to perform a weighted summation of the landmark feature vectors to obtain the target-oriented context vector.

7. The zero-sample target navigation method based on landmark memory as described in claim 1, characterized in that, In generating the discriminative state representation for decision-making, the spatiotemporal context vector and the goal-oriented context vector are added together to form a unified inference vector. A learnable gating coefficient is then used to weight and fuse the inference vector and the landmark feature vector to obtain the discriminative state representation for decision-making. : in, This indicates element-wise multiplication; The landmark feature vector corresponding to the landmark memory database; The gating coefficient; for Activation function This is the weight matrix; This is the inference vector.

8. A zero-sample target navigation system based on landmark memory, characterized in that, The zero-sample target navigation method based on landmark memory as described in any one of claims 1-7 includes: The scene category label inference module is used to receive the visual observation of the current field of view and the semantic category labels of objects in the field of view at each time step of the navigation task, encode them into feature vectors, identify landmarks from the visual observation of the current field of view, and then infer the scene category label to which the landmark belongs by combining the scene prior knowledge. The landmark memory bank construction module is used to build an explicit externalized landmark memory bank based on the scene category label to which the landmark belongs, the visual feature representation of the landmark location, and the semantic embedding associated with the landmark. This mimics the cognitive mapping mechanism of biological navigation, captures, stores, and dynamically updates key locations in the environment online, forms a sparse topological map, and extracts landmark feature vectors from it. The discriminative state representation generation module is used to extract spatiotemporal context vectors from local observation images and short-term historical path information by combining a spatiotemporal cross-attention mechanism with a landmark memory bank. It also uses a target perception attention mechanism and the semantic embedding vector of the target object to extract a target-oriented context vector from the landmark memory bank, thereby forming an inference vector. Finally, it uses a gating coefficient to weight and fuse the inference vector and the landmark feature vector to obtain a discriminative state representation for decision-making. The robot control command generation module generates robot control commands based on the discriminative state representation used for decision-making and a deep reinforcement learning framework.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the landmark memory-based zero-sample target navigation method as described in any one of claims 1-7.

10. A robot, comprising a robot body, a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the zero-sample target navigation method based on landmark memory as described in any one of claims 1-7.

Images