Image target navigation method and device based on potential space autoregression world model
By using an autoregressive world model for image target navigation in the latent space, the problems of high computational cost, weak generalization ability, and insufficient real-time performance in existing navigation technologies are solved, achieving efficient and stable navigation results.
Patent Information
- Application Number
- CN202511655172.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-12
- Publication Date
- 2026-02-10
AI Technical Summary
Existing navigation technologies have shortcomings in terms of computational efficiency, long-term time series modeling capabilities, and generalization ability. In particular, they are difficult to maintain navigation accuracy and robustness in dynamic or unknown environments, and the computational cost is high.
An image-based target navigation method based on a latent space autoregressive world model is adopted. By performing state prediction and planning in the latent space, pixel-level video reconstruction is abandoned and an autoregressive multi-frame prediction mechanism is introduced to achieve a lightweight navigation design.
It significantly reduces computational and storage overhead, improves the real-time performance of navigation and trajectory prediction accuracy, enhances navigation stability and generalization ability in dynamic environments, and achieves lightweight, robust and deployable capabilities.
Smart Images
Figure CN121492019A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of target navigation technology, specifically to an image target navigation method and apparatus based on a latent spatial autoregressive world model. Background Technology
[0002] Autonomous navigation is a core research direction in the field of robotics, and after years of development, a framework comprising mature components such as mapping, localization, planning, and control has been formed. However, such complex modular designs often result in complex system structures and high coupling between modules, making it easy for errors to accumulate during multi-level processing, leading to trajectory deviations or navigation failures. These methods require stable sensor input and high-precision environmental modeling, making it difficult to maintain global consistency and navigation accuracy in dynamic or unknown environments. The fundamental reason is that these methods rely on explicit spatial modeling and global coordinate constraints; when the environment changes or perceived noise increases, the global mapping and localization mechanisms struggle to maintain accuracy.
[0003] To overcome the limitations of traditional multi-module navigation, researchers have proposed end-to-end navigation methods based on reinforcement learning and imitation learning. However, reinforcement learning methods suffer from extremely low sample efficiency during training in high-dimensional continuous action spaces, requiring extensive trial and error for policy optimization, resulting in high computational costs and security risks in practical deployment. While imitation learning avoids random exploration, it relies excessively on expert demonstration data, limiting data coverage and leading to insufficient generalization ability in new environments. The technical bottleneck of these methods lies in the fact that they all fail to construct effective internal world models, lack explicit understanding of environmental dynamics, and rely solely on input distribution for policy fitting, thus failing to maintain robustness when environmental distribution changes.
[0004] In recent years, world model methods have emerged that attempt to improve the predictability and reasoning ability of navigation by learning environmental dynamics. However, existing solutions generally suffer from excessively high computational complexity. Many models use video generation or pixel-level future frame prediction to characterize environmental evolution. While these methods can reproduce realistic spatiotemporal changes, their training and inference costs are extremely high, making them unsuitable for resource-constrained robotic platforms. The technical reason for this is that the pixel space has extremely high dimensionality, requiring the model to process a large number of visual details unrelated to navigation decisions during the learning process, resulting in wasted computational resources and reduced real-time performance. Furthermore, while some world models using implicit feature representations can reduce dimensionality, they lack temporal modeling capabilities and can only capture static semantic features, making it difficult to effectively predict dynamic changes and long-term dependencies in the environment. Summary of the Invention
[0005] To address the significant shortcomings of existing navigation technologies in terms of computational efficiency, long-term temporal modeling capabilities, and generalization, this invention provides an image target navigation method and apparatus based on a latent spatial autoregressive world model. By predicting future potential states rather than pixel-level video sequences, end-to-end image target navigation is achieved, which can significantly reduce memory usage and computational requirements while ensuring navigation performance.
[0006] To achieve the above objectives, this invention provides an image target navigation method based on a latent spatial autoregressive world model, comprising the following steps: Step 1: Construct a world model and train the world model, which includes: The world state embedding module is used to map visual observations and action information to a unified latent space to obtain the state vector corresponding to the visual observations and the action vector corresponding to the action information. The state transition module is used to predict the state based on the historical state vector sequence and action vector, and generate the predicted state vector. Step 2: Obtain the current visual observation and the target visual observation, and obtain the current state vector corresponding to the current visual observation and the target state vector corresponding to the target visual observation based on the world state embedding module; Step 3: Update the current state vector to the historical state vector sequence, optimize the action sequence based on the trained world model, and obtain the optimal predicted state vector that is closest to the target state vector, and the optimal action sequence corresponding to the optimal predicted state vector. Step 4: Extract the first action from the optimal action sequence and execute it. Then update the current visual observation and the current state vector. Determine whether the distance between the current state vector and the target state vector is less than a set threshold. If so, end the navigation; otherwise, return to step 3.
[0007] To achieve the above objectives, the present invention also provides an image target navigation device based on a latent spatial autoregressive world model, employing the above-described method for target navigation. The image target navigation device includes: A model building unit is used to build a world model and train the world model, the world model comprising: The world state embedding module is used to map visual observations and action information to a unified latent space to obtain the state vector corresponding to the visual observations and the action vector corresponding to the action information. The state transition module is used to predict the state based on the historical state vector sequence and action vector, and generate the predicted state vector. The feature extraction unit is used to acquire the current visual observation and the target visual observation, and to obtain the current state vector corresponding to the current visual observation and the target state vector corresponding to the target visual observation based on the world state embedding module. An action optimization unit is used to update the current state vector to the historical state vector sequence, optimize the action sequence based on the trained world model, and obtain the optimal predicted state vector that is closest to the target state vector, and the optimal action sequence corresponding to the optimal predicted state vector. The target navigation unit extracts and executes the first action in the optimal action sequence, updates the current visual observation and the current state vector, and ends navigation when the distance between the current state vector and the target state vector is less than a set threshold.
[0008] Compared with the prior art, the present invention has the following beneficial technical effects: 1. This invention achieves efficient prediction and planning of future navigation states by establishing an autoregressive world model in the potential space. Compared with traditional methods that rely on pixel-level video reconstruction or global mapping, it significantly reduces computation and storage overhead and significantly improves inference efficiency and real-time performance. 2. This invention effectively captures long-term spatiotemporal dependencies by performing state evolution modeling and multi-frame autoregressive prediction in a low-dimensional latent space, thereby enhancing the trajectory prediction accuracy and navigation stability of the model in dynamic and unknown environments. 3. This invention abandons the complex multi-module structure and reliance on high-precision perception, and realizes the integrated design of perception, prediction and decision-making. Thus, while ensuring navigation performance, it achieves a unity of lightweight, robustness and deployability, overcoming the technical bottlenecks of high computing cost, weak generalization ability and insufficient real-time performance of existing technologies. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the structures shown in these drawings without creative effort.
[0010] Figure 1 This is a flowchart of the image target navigation method based on the latent spatial autoregressive world model in an embodiment of the present invention; Figure 2 This is a structural block diagram of an image target navigation device based on a latent spatial autoregressive world model in an embodiment of the present invention.
[0011] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0012] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.
[0013] Furthermore, the technical solutions of the various embodiments of the present invention can be combined with each other, but only if they are feasible for those skilled in the art. If the combination of technical solutions is contradictory or cannot be implemented, it should be considered that such combination of technical solutions does not exist and is not within the scope of protection claimed by the present invention.
[0014] Example 1 This embodiment discloses an image target navigation method based on an autoregressive world model in latent space. By mapping visual observations and tasks to a latent space, and directly predicting future states and planning paths within that space, it eliminates the reliance on pixel-level video reconstruction and explicit map construction. Specifically, by modeling state evolution in a low-dimensional latent representation, it significantly reduces the computational load for training and inference. Simultaneously, it introduces an autoregressive multi-frame prediction mechanism to effectively capture long-term spatiotemporal dependencies, improving trajectory prediction accuracy and navigation stability in complex environments. This addresses the technical bottlenecks of existing navigation technologies, such as high computational cost, weak generalization ability, and difficulty in real-time deployment, thereby effectively meeting the navigation needs of autonomous robots in real-world scenarios.
[0015] refer to Figure 1 The image target navigation method based on the latent spatial autoregressive world model in this embodiment specifically includes the following steps: Step 1: Construct and train the world model. The world model includes: The world state embedding module is used to map visual observations and action information to a unified latent space to obtain the state vector corresponding to the visual observations and the action vector corresponding to the action information. The state transition module is used to predict the state based on the historical state vector sequence and action vector, and generate the predicted state vector. Step 2: Obtain the current visual observation and the target visual observation, and obtain the current state vector corresponding to the current visual observation and the target state vector corresponding to the target visual observation based on the world state embedding module; Step 3: Update the current state vector to the historical state vector sequence, optimize the action sequence based on the trained world model, and obtain the optimal predicted state vector that is closest to the target state vector, as well as the optimal action sequence corresponding to the optimal predicted state vector. Step 4: Extract the first action from the optimal action sequence and execute it. Then update the current visual observation and the current state vector. Determine whether the distance between the current state vector and the target state vector is less than a set threshold. If so, end the navigation; otherwise, return to step 3.
[0016] In this embodiment, the world model takes image sequences and corresponding navigation actions as input. It can directly learn the mapping relationship from historical observations and actions to future states from video-based interactive data, thereby achieving environmental state prediction and navigation planning in the latent space. In navigation tasks, pixel-level image prediction is usually computationally complex and redundant. However, for navigation behavior, the key is to obtain semantically consistent latent state representations. Therefore, this embodiment transfers the prediction target from the pixel space to the latent space, and achieves the prediction of future latent states through a state transition model. The latent space is a commonly used concept in deep learning, referring to the low-dimensional feature space formed after compressing high-dimensional input data (such as images) through an encoder. This space retains semantic information useful for the task. Let the observation sequence be... , representing time frame Images acquired by an autonomous robot; let the action sequence be... , representing time frame The actions of the autonomous robot. Each action... Including forward displacement Lateral displacement and rotation angle ,in, , , They are time steps The world model calculates the forward displacement, lateral displacement, and rotation angle. It extracts latent state vectors from visual observation and motion information using an observation encoder and a motion encoder, respectively. Implicit action vectors And based on this, establish the state transition relationship as follows:
[0017] in, For time step The state vector below, This is the state transition function. This represents the potential spatial representation sequence of historical observations.
[0018] In this embodiment, the world state embedding module uses a pre-trained variational autoencoder (VAE) to encode visual observations, thereby mapping the visual observations to the latent space and obtaining the corresponding state vectors. This not only achieves globally consistent feature representations but also accelerates training convergence. Specifically, the variational autoencoder is a generative deep learning model composed of an encoder and a decoder, which compresses and reconstructs input data through probability distribution modeling. Action information is embedded periodically. By applying sine and cosine transforms to each action component, the model can effectively capture periodic and spatial variation features. After processing by a multilayer perceptron, the embedding features of the three action components are concatenated to form a unified action vector, enhancing the model's ability to represent complex motion patterns. In the specific implementation, the process of mapping action information to the latent space to obtain action vectors includes: First, apply sine and cosine transforms to each motion component of the motion information to obtain the forward displacement. Embedding features Lateral displacement Embedding features and rotation angle Embedding features ; Then, the embedded features of each action component are mapped using a multilayer perceptron and concatenated to obtain the action vector, as follows:
[0019]
[0020]
[0021]
[0022]
[0023] in, For time step The action vector below, For feature splicing operations, , , These are the mapping functions for the multilayer perceptron. For frequency components, For the maximum period range, This represents the number of frequency bases.
[0024] In this embodiment, a state transition module fuses historical observations and action features to achieve future state prediction based on action conditions. In specific implementation, the operation of the state transition module includes: First, obtain the vector containing the current state. Historical state vector sequence and the current time step Action vectors under ,in, The length of the historical state vector sequence; Then, the feature adaptive fusion mechanism SELayer (Squeeze-and-Excitation Layer) is used to process the historical state vector sequence. With action vectors Joint modulation is performed to obtain the joint modulation feature sequence. That is, by introducing a dynamic weighting mechanism, the input feature tensor is adaptively scaled to highlight environmental information related to the current action. The joint modulation process is as follows:
[0025] in, It is the Sigmoid activation function. This represents element-wise multiplication. It is a multilayer perceptron; Subsequently, in order to reduce redundant features and improve computational efficiency, a token-based feature selection mechanism was introduced to filter the jointly modulated feature sequences. Feature filtering is performed to obtain a compact set of key tokens. ; In terms of temporal modeling, to capture the robot's environmental dependencies during long-term navigation, this embodiment introduces a self-attention mechanism on the historical feature sequence: after obtaining the latent representation of the current world state, the state transition module fuses historical observations with action features to achieve action-based prediction of future states, i.e., prediction of the key token set. Perform temporal dependency modeling to obtain feature sequences integrated into the global temporal context. ,for:
[0026] in, This is a self-attention mechanism, a deep network structure used to assign different weights based on the importance of input features, thereby enhancing the expression of key information. Furthermore, the historical state vector sequence is fused through a cross-time step feature fusion mechanism. With characteristic sequence Feature fusion is performed to obtain a fused feature sequence. ,for:
[0027] in, , For learnable parameters, This is the transpose of the matrix; Finally, the fused feature sequences By performing convolution and mapping with a multilayer perceptron, the predicted state vector for the next time step is obtained. ,for:
[0028] in, This is a convolution operation.
[0029] In this embodiment, the joint modulation feature sequence Feature filtering can significantly reduce the computational complexity of a model while preserving key information. The specific implementation process includes: First, in the joint modulation feature sequence Based on this, attention weights are calculated using a multilayer perceptron. ; Then, the softmax function is used to adjust the attention weights. Normalization is performed to obtain normalized attention weights. ; Finally, based on normalized attention weights For joint modulation feature sequences Perform a weighted summation to obtain a compact set of key tokens. ,for:
[0030] in, The number of tokens in the input space. Normalized attention weights The Middle The output token is the first Attention weights are assigned to each input token. This outputs the number of tokens.
[0031] In this embodiment, the world model training process employs an autoregressive multi-step prediction training strategy. Autoregressive prediction is a sequence modeling method whose basic idea is to use historical outputs as inputs to progressively generate subsequent prediction results. Unlike pixel-based supervision, the world model in this embodiment directly supervises the prediction state in the latent space, thereby significantly reducing computational complexity. During training, the world model uses the predicted output and the actual action as inputs for the next step at each step and continues to make multi-step predictions. By performing mean squared error (MSE) supervision at multiple time steps, the world model can maintain the consistency and stability of long-term predictions. Simultaneously, due to the lightweight structure of the world model and the absence of image-level reconstruction, the overall training process has high computational efficiency. The loss function during training... for:
[0032]
[0033]
[0034] in, For time step The predicted state vector below, For time step The true state vector below, To predict the number of steps, = For those with The historical state vector sequence for predicting the state vector step by step. This refers to the state transition function (i.e., the execution process of the aforementioned state transition module). Using these parameters as model parameters, this strategy effectively improves the model's stability and generalization ability in long-term prediction scenarios.
[0035] In image target navigation tasks, the robot plans the optimal sequence of actions to reach the target location based on the currently observed image and the target image. This embodiment, based on a world model of latent space, can predict future state sequences through continuous actions, starting from an initial state, thereby gradually completing image target navigation. In specific implementation, optimizing the action sequence based on the trained world model includes: First, obtain the current state vector. Target state vector And several candidate action sequences, where the candidate action sequences are represented as The candidate action sequence is then mapped to the latent space to obtain the candidate action vector sequence. ,in, , For the current candidate action and the current candidate action vector, , For the future Candidate actions and candidate action vectors for each step; The current state vector Update to the historical state vector sequence, and generate the future autoregressively based on the world model. Predicted state vector sequence of steps That is, the prediction is made using the methods of equations (13) and (14); Based on the target state vector With the predicted state vector Construct the energy function as follows:
[0036] in, Let be the energy function. The target state vector With prediction vector Euclidean distance loss; With the goal of minimizing the energy function, the cross-entropy method is used to sample and iteratively optimize candidate action sequences until the closest target state vector is obtained. The optimal predicted state vector and the optimal predicted state vector The corresponding optimal action sequence. The cross-entropy method is a stochastic optimization algorithm that gradually optimizes the strategy or parameter distribution by sampling and filtering high-performance samples. Its specific implementation process is a conventional technique in this field, and therefore will not be described in detail in this embodiment.
[0037] It is worth noting that, although this embodiment Figure 1 The steps are shown sequentially as indicated by the arrows, but they are not necessarily executed in the order indicated by the arrows. Unless otherwise specified in this document, there is no strict order in which these steps are performed; they can be executed in other orders. Figure 1 At least some of the steps in the process may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be executed in turn or alternately with other steps or at least some of the sub-steps or stages of other steps.
[0038] Example 2 Based on the image target navigation method based on the latent spatial autoregressive world model in Embodiment 1, this embodiment discloses an image target navigation device based on the latent spatial autoregressive world model, referencing... Figure 2The image target navigation device includes a model building unit, a feature extraction unit, an action optimization unit, and a target navigation unit, specifically: The model building unit is used to build and train the world model, which includes: The world state embedding module is used to map visual observations and action information to a unified latent space, obtaining the state vector corresponding to the visual observations and the action vector corresponding to the action information. The state transition module is used to predict the state based on the historical state vector sequence and action vector, and generate the predicted state vector. The feature extraction unit is used to obtain the current visual observation and the target visual observation, and obtains the current state vector corresponding to the current visual observation and the target state vector corresponding to the target visual observation based on the world state embedding module; The action optimization unit is used to update the current state vector to the historical state vector sequence, optimize the action sequence based on the trained world model, and obtain the optimal predicted state vector that is closest to the target state vector, as well as the optimal action sequence corresponding to the optimal predicted state vector. After the target navigation unit extracts and executes the first action in the optimal action sequence, it updates the current visual observation and the current state vector, and ends navigation when the distance between the current state vector and the target state vector is less than a set threshold.
[0039] In this embodiment, the specific working process and working principle of the model building unit, feature extraction unit, action optimization unit, and target navigation unit are the same as those in Embodiment 1, and therefore will not be described again in this embodiment. Each unit module can be implemented entirely or partially through software, hardware, or a combination thereof. Each unit module can be embedded in or independent of the processor in a computer device in hardware form, or it can be stored in the memory of a computer device in software form, so that the processor can call and execute the operations corresponding to each of the above unit modules.
[0040] The above description is only a preferred embodiment of the present invention and does not limit the scope of protection of the present invention. All equivalent structural transformations made under the inventive concept of the present invention using the contents of the present invention specification and drawings, or direct / indirect applications in other related technical fields, are included within the scope of protection of the present invention.
Claims
1. An image target navigation method based on a latent spatial autoregressive world model, characterized in that, Includes the following steps: Step 1: Construct a world model and train the world model, which includes: The world state embedding module is used to map visual observations and action information to a unified latent space to obtain the state vector corresponding to the visual observations and the action vector corresponding to the action information. The state transition module is used to predict the state based on the historical state vector sequence and action vector, and generate the predicted state vector. Step 2: Obtain the current visual observation and the target visual observation, and obtain the current state vector corresponding to the current visual observation and the target state vector corresponding to the target visual observation based on the world state embedding module; Step 3: Update the current state vector to the historical state vector sequence, optimize the action sequence based on the trained world model, and obtain the optimal predicted state vector that is closest to the target state vector, and the optimal action sequence corresponding to the optimal predicted state vector. Step 4: Extract the first action from the optimal action sequence and execute it. Then update the current visual observation and the current state vector. Determine whether the distance between the current state vector and the target state vector is less than a set threshold. If so, end the navigation; otherwise, return to step 3.
2. The image target navigation method based on a latent spatial autoregressive world model according to claim 1, characterized in that, The world state embedding module uses a pre-trained variational autoencoder to map visual observations to the latent space, obtaining the state vector corresponding to the visual observation.
3. The image target navigation method based on a latent spatial autoregressive world model according to claim 1, characterized in that, The action information includes three action components: forward displacement, lateral displacement, and rotation angle. The world state embedding module maps the action information to the latent space to obtain the action vector corresponding to the action information. The process includes: First, apply sine and cosine transforms to each motion component of the motion information to obtain the forward displacement. Embedding features Lateral displacement Embedding features and rotation angle Embedding features ,in, , , They are time steps The forward displacement, lateral displacement, and rotation angle are measured. Then, the embedded features of each action component are mapped using a multilayer perceptron and concatenated to obtain the action vector, as follows: in, For time steps The action vector below, For feature splicing operations, , , These are the mapping functions for the multilayer perceptron. For frequency components, For the maximum period range, This represents the number of frequency bases.
4. The image target navigation method based on a latent spatial autoregressive world model according to claim 1, 2, or 3, characterized in that, The operation of the state transition module includes: Get the vector containing the current state Historical state vector sequence and the current time step Action vectors under ,in, The length of the historical state vector sequence; A feature adaptive fusion mechanism is used to process the historical state vector sequence. With action vectors Perform joint modulation to obtain the joint modulation feature sequence. ,in, It is the Sigmoid activation function. This represents element-wise multiplication. It is a multilayer perceptron; Introducing a token-based feature selection mechanism for joint modulation feature sequences Feature filtering is performed to obtain a compact set of key tokens. ; For key token set Perform temporal dependency modeling to obtain feature sequences integrated into the global temporal context. ,in, This is a self-attention mechanism; The historical state vector sequence is obtained through a cross-time step feature fusion mechanism. With characteristic sequence Feature fusion is performed to obtain a fused feature sequence. ,in, , For learnable parameters, This is the transpose of the matrix; For fusion feature sequences By performing convolution and mapping with a multilayer perceptron, the predicted state vector for the next time step is obtained. ,in, This is a convolution operation.
5. The image target navigation method based on a latent spatial autoregressive world model according to claim 4, characterized in that, Introducing a token-based feature selection mechanism for joint modulation feature sequences Feature filtering is performed to obtain a compact set of key tokens. The process includes: Attention weights are calculated using a multilayer perceptron. ; Use the softmax function to adjust attention weights Normalization is performed to obtain normalized attention weights. ; Based on normalized attention weights For joint modulation feature sequences Perform a weighted summation to obtain a compact set of key tokens. ,for: in, The number of tokens in the input space. Normalized attention weights Middle The output token is the first Attention weights are assigned to each input token. This outputs the number of tokens.
6. The image target navigation method based on a latent spatial autoregressive world model according to claim 4, characterized in that, The world model is trained using an autoregressive multi-step prediction training strategy, and the loss function during the training process... for: in, For time steps The predicted state vector below, For time steps The true state vector below, To predict the number of steps, = For those with The historical state vector sequence for predicting the state vector step by step. This is the state transition function.
7. The image target navigation method based on a latent spatial autoregressive world model according to claim 1, 2, or 3, characterized in that, Step 3 specifically includes: Get the current state vector Target state vector And several candidate action sequences, where the candidate action sequences are represented as The candidate action sequence is then mapped to the latent space to obtain the candidate action vector sequence. ,in, , For the current candidate action and the current candidate action vector, , For the future Candidate actions and candidate action vectors for each step; The current state vector Update to the historical state vector sequence, and generate the future autoregressively based on the world model. Predicted state vector sequence of steps ; Based on the target state vector With the predicted state vector Construct the energy function as follows: in, Let be the energy function. The target state vector With prediction vector Euclidean distance loss; With the goal of minimizing the energy function, the cross-entropy method is used to sample and iteratively optimize candidate action sequences until the closest target state vector is obtained. The optimal predicted state vector and the optimal predicted state vector The corresponding optimal action sequence.
8. An image target navigation device based on a latent spatial autoregressive world model, characterized in that, The image target navigation device comprises: (The method described in any one of claims 1 to 7 is used for target navigation.) A model building unit is used to build a world model and train the world model, the world model comprising: The world state embedding module is used to map visual observations and action information to a unified latent space to obtain the state vector corresponding to the visual observations and the action vector corresponding to the action information. The state transition module is used to predict the state based on the historical state vector sequence and action vector, and generate the predicted state vector. The feature extraction unit is used to acquire the current visual observation and the target visual observation, and to obtain the current state vector corresponding to the current visual observation and the target state vector corresponding to the target visual observation based on the world state embedding module. An action optimization unit is used to update the current state vector to the historical state vector sequence, optimize the action sequence based on the trained world model, and obtain the optimal predicted state vector that is closest to the target state vector, and the optimal action sequence corresponding to the optimal predicted state vector. The target navigation unit extracts and executes the first action in the optimal action sequence, updates the current visual observation and the current state vector, and ends navigation when the distance between the current state vector and the target state vector is less than a set threshold.