Brain-like navigation method and system based on multi-scale spatio-temporal representation and relation memory

Abstract

A kind of brain navigation method and system based on multi-scale space-time representation and relation memory, by collecting the environment observation information and corresponding history observation sequence of current time, action sequence or state sequence, and the hierarchical multi-scale space-time coding based on geometric cell model and self-attention mechanism fusion is carried out to environment observation information and history observation sequence, after obtaining the multi-scale state representation of current time, the node in relation structure and the connection between nodes are constructed according to multi-scale state representation;Then update relation structure and according to history relation structure, current state and candidate action, the future state or expected result is obtained by deep learning prediction to determine and execute current optimization action, finally, according to actual result, the iteration of state update, relation structure update is carried out, the brain navigation is realized, the layered representation of environment observation information is carried out by multi-scale space-time representation method, to improve the expression ability of local structure and global relation in complex scene;The correlation between states is constructed by the environment modeling method based on relation structure, and the organization and expression of environment information are realized.

Description

Technical Field

[0001] This invention relates to a technology in the field of autonomous robot navigation, specifically a brain-like navigation method and system based on multi-scale spatiotemporal representation and relational memory. Background Technology

[0002] Navigation in unknown environments involves not only estimating the current state but also utilizing historical experience and inferring future states. Existing technologies remain imperfect in multi-scale spatiotemporal information modeling, relational structure representation, and predictive reasoning based on internal representations, making it difficult to achieve stable navigation decisions within a unified framework. Existing navigation technologies based on spatiotemporal risk field prediction and deep reinforcement learning heavily rely on high-precision global prior data, preventing agents from achieving autonomous environmental cognition in unknown (e.g., without GPS or prior maps) complex environments. Furthermore, relying solely on black-box sequence networks to predict objective risk fields fails to quantify the agent's "cognitive uncertainty" in the face of unknown environments, hindering the trade-off between exploration and safe decision-making. Existing satellite navigation and positioning technologies based on deep learning enhancement heavily rely on external GNSS signals and aerial visual observation, preventing agents from achieving autonomous environmental cognition and navigation in complex environments without prior knowledge, such as indoors, underground, or where GPS is denied. Moreover, using deep learning models (such as CNN and LSTM) to weight and correct errors in objective signal quality fails to establish an internal cognitive representation of the environment's geometry, and further fails to quantify the "cognitive uncertainty" in the face of unknown areas, making it difficult to utilize Confidence World Models (BIW) for generative belief prediction and exploration decision-making trade-offs. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a brain-like navigation method and system based on multi-scale spatiotemporal representation and relational memory. By using a multi-scale spatiotemporal representation method to hierarchically represent environmental observation information, the ability to express local structures and global relationships in complex scenes is improved. Through an environment modeling approach based on relational structures, the association between states is constructed to achieve the organization and expression of environmental information.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a brain-like navigation method based on multi-scale spatiotemporal representation and relational memory. It collects environmental observation information and corresponding historical observation sequences, action sequences, or state sequences at the current moment. The environmental observation information and historical observation sequences are then subjected to hierarchical multi-scale spatiotemporal encoding based on a fusion of a geometric cell model and a self-attention mechanism to obtain the multi-scale state representation at the current moment. Based on this multi-scale state representation, nodes and connections between nodes in the relational structure are constructed. The relational structure is then updated, and based on the historical relational structure, the current state, and candidate actions, the future state or expected result is predicted through deep learning, thereby determining and executing the current optimized action. Finally, iterative updates of the state and relational structure are performed based on the actual results to achieve brain-like navigation.

[0006] The geometric cell coding model Taking into account both Gaussian distance decay and von Mises angle distribution, specifically: ,in: The two-dimensional coordinates of the cell's egocentric space. For distance, Let azimuth be the azimuth angle, and its domain be . For intelligent agents in an omnidirectional perspective The depth sensor detects the Euclidean distance to the environmental boundary, representing a physical measurement of local obstacles. The Gaussian variance parameter for distance sensing is used as a control term to adjust the system's tolerance to sensor ranging noise and uncertainty. The current average vector direction (MVD) of the agent, calculated using the head orientation cell (HDC) model, serves as a dynamic reference benchmark for spatial encoding. The angular concentration parameter is used to control the sharpness of the peak value in the von Mises distribution. It is a zero-order modified Bessel function of the first kind.

[0007] The hierarchical multi-scale spatiotemporal coding refers to: discretizing and mapping the physical boundary distance information and head direction information from the omnidirectional perspective of the agent using a geometric cell coding model to obtain a local representation matrix based on geometric cells. Then, multi-scale global aggregation and global structural attention encoding are performed, specifically including:

[0008] Step 1, Discretization Mapping: The agent acquires data in real time through sensors. A discrete perspective Euclidean distance of the environmental boundary and head direction angle reference value and build A two-dimensional geometric neuron sheet array serves as a spatial topological carrier; each coordinate unit is computed through traversal. The firing rate intensity of neurons at a given location is explicitly quantified to represent the physical boundary features of the agent from an omnidirectional perspective, ultimately outputting a complete and precise representation of the local spatial geometry. 3D matrix, as a local representation matrix based on geometric cells Specifically, it refers to the intensity of neuronal firing rate. ,in: Discrete position Geometric cell response intensity at the location Discrete quantity for environmental observation direction For the first Discrete observation directions For direction The detected environmental boundary distance For the first Discrete distance scale For the first Discrete azimuth angles Reference angle for the current head orientation of the agent Distance coding scale parameter Parameters in the direction encoding set For zero-order modified Bessel function For the angular discrete interval, satisfying These are the direction, distance, and angle indices, respectively.

[0009] Step 2, Multi-scale Global Aggregation: Extracting region-scale representations through spatial pooling and nonlinear activation functions. Aggregate neighborhood topology information, specifically: ,in: Current moment The regional scale representation matrix; Nonlinear activation functions are used to introduce nonlinear feature maps; Preset regional feature transformation weight matrix; Spatial pooling operators are used to downsample and reduce the spatial dimensionality of input features; Current moment The local representation matrix based on the geometric cell is used as the input to the spatial pooling operator; Preset region feature transformation bias vector.

[0010] Step 3, Global Structural Attention Encoding: An improved method is used to encode regional features based on a self-attention mechanism. Linear projection is used to calculate global topological connectivity features based on the query vector (Q), key vector (K), and value vector (V). Specifically: ,in: The scaling dimension of the feature vector is a scalar.

[0011] By using hierarchical multi-scale spatiotemporal coding within the same representation framework, the precise local geometric boundary features of the environment are seamlessly integrated with the macroscopic global topological relationships, providing high-quality, interference-resistant multi-scale feature inputs for subsequent generative belief prediction and relational memory storage.

[0012] The deep learning predictions include:

[0013] Step 1: Construct the spatiotemporally correlated generative state transition inputs, specifically including:

[0014] 1.1 Extracting the multi-scale spatial state representation at the current moment Furthermore, by combining the relational memory module, historical memory context features related to the current state topology are retrieved. .

[0015] 1.2 Acquiring candidate exploration actions of the agent in the action space, including motion commands such as linear velocity and angular velocity. ,Will , The features are stitched together.

[0016] 1.3 A multilayer perceptron (MLP) encoder is used to perform nonlinear mapping on the stitched features, fusing spatial topology, action intent, and historical experience to extract generative latent variable context vectors. ,in: These are the learnable weight matrix and bias term of the context encoding network, respectively. This indicates the concatenation operation of feature dimensions.

[0017] Step 2: Perform future state posterior distribution prediction based on variational inference, specifically including:

[0018] 2.1 Latent variable context vector The inputs are fed into a two-branch variational inference network to predict the mean distribution of future states. With log variance This is used to simulate the uncertainty of environmental state transitions, specifically: , ,in: To predict variance, For predicted future moments The mean vector of the state distribution; This is the learnable weight matrix for the mean prediction branch in the variational inference network; For a moment Encoded latent variable context vector; is the learnable bias vector for the mean prediction branch in the variational inference network; It is the natural logarithm operator; For predicted future moments The predicted variance vector of the state distribution; This is the learnable weight matrix for the log-variance prediction branch in the variational inference network; is the learnable bias vector for the log-variance prediction branch in the variational inference network.

[0019] The larger the variance, the less prior experience the agent has with the future state resulting from the candidate action.

[0020] 2.2 Employing the reparameterization trick, from the standard normal distribution... Medium sampling noise This leads to the generation of a predicted state representation that includes spatial geometry and topological information of the predicted environment. ,in: This indicates element-wise multiplication. For the generated future moment The predicted state is a vector; For predicted future moments The mean vector of the state distribution; This is a random noise vector sampled from a standard normal distribution; For element-wise multiplication (Hadamard product); For the natural constant Exponential operators with base 0; It is the natural logarithm operator; For predicted future moments The variance vector of the state distribution.

[0021] Step 3: Quantify cognitive uncertainty and output subjective confidence scores, specifically including:

[0022] 3.1 Based on prediction variance The L2 norm of the variance vector is extracted as a measure of epistemic uncertainty.

[0023] 3.2 Based on the exponential penalty function, calculate the subjective confidence score corresponding to this prediction and use it as the core weight for evaluating path safety and exploration value in subsequent navigation decisions, specifically: ,in: The scaling hyperparameter is used to adjust the sensitivity of the confidence level. .

[0024] The iterative process of updating the state and relational structure based on actual perception results includes the following specific iterative steps:

[0025] Step 1: Execute the target action and acquire real-world environment observations: Based on the optimal action output by the navigation decision module, the agent performs movement in the physical or simulated environment, triggering the next time step. Multi-source sensor data acquisition, and encoding through a multi-scale spatiotemporal representation module to obtain a true spatiotemporal state representation. .

[0026] Step 2: Calculate the joint loss and the gradient using the backpropagation algorithm. Update the network parameters of the confidence world model online to ensure continuous convergence of the model in unknown environments and improve the accuracy of state prediction. Specifically: Among them: state reconstruction loss Distribution divergence loss , The regularization coefficient is , i.e. The divergence weight coefficient is used to weigh the equilibrium state reconstruction accuracy against the probability distribution regularization constraint. To predict the state, This represents the actual state.

[0027] Step 3: Update the relational cognitive map and calibrate the experience memory retrieval: Calculate the current true state. Similarity to existing node states in the relational memory module. If the maximum similarity is lower than a preset novelty threshold. Then As a new node, it is stored in the relational cognitive graph, and new edge connections are established based on the executed action trajectory. If it exceeds the threshold, the access frequency weight of the corresponding node is increased, the memory topology graph is dynamically calibrated, and more accurate empirical context is provided for the generative belief prediction of the next cycle.

[0028] This invention relates to a brain-like navigation system based on multi-scale spatiotemporal representation and relational memory to implement the above-mentioned method. The system includes: a multi-source observation acquisition module, a multi-scale spatiotemporal representation module, a relational cognitive graph construction module, a relational memory module, a generative belief prediction module, a navigation decision module, and an online update module. Specifically: the multi-source observation acquisition module acquires observational information of the agent in the environment and performs hierarchical multi-scale representation to express environmental information within different ranges; the multi-scale spatiotemporal representation module encodes current observational information and historical information to obtain state representations at different spatial scales and time ranges; the relational cognitive graph construction module constructs a relational structure composed of nodes and edges based on the state representations; the relational memory module writes, retrieves, and updates the relational structure to manage historical information; the generative belief prediction module predicts state change trends, target proximity, or probability of behavioral success based on the current state, relational memory, and candidate actions; the navigation decision module determines the navigation action at the current moment based on the prediction results; and the online update module updates the state representation, relational structure, and relational memory during navigation based on newly acquired observational information.

[0029] The observation information includes: visual observation, depth observation, laser observation, inertial observation, odometer information, or historical trajectory information.

[0030] In the relational structure, nodes represent the state at each moment, and edges represent temporal continuity, spatial proximity, or action transition relationships, and weights or confidence levels can be set.

[0031] The multi-scale spatiotemporal representation module includes: a local geometric perception unit, a regional topology extraction unit, a global relation association unit, and a multi-scale spatiotemporal fusion unit. Specifically: the local geometric perception unit, based on depth data or laser data and dynamic reference direction information from the current environmental observation information, performs brain-like geometric cell feature reconstruction and polar coordinate discretization feature mapping in polar coordinates to obtain a local-scale state representation result that explicitly quantifies local physical boundary features; the regional topology extraction unit, based on the local-scale state representation result output by the local geometric perception unit, performs neighborhood feature space pooling dimensionality reduction and regional topological connectivity nonlinear feature activation processing to obtain a representation describing a specific neighborhood and... The regional scale state representation results of the connectivity structure relationship are obtained. The global relationship association unit performs long-distance spatial connectivity encoding and global environment layout modeling based on the regional scale state representation results output by the regional topology extraction unit to obtain the global scale state representation results describing the global topological connectivity relationship of the overall environment. The multi-scale spatiotemporal fusion unit performs multi-level feature concatenation and spatiotemporal correlation transformation feature fusion processing based on the local scale state representation results, regional scale state representation results, global scale state representation results and input historical observation sequence and action sequence information to obtain the complete multi-scale spatiotemporal state representation results representing the multi-scale topological features inside the environment at the current moment.

[0032] The state representation includes: local scale representation Intermediate scale representation and global scale representation , where: local scale representation Used to describe the geometric information of physical boundaries and passable obstacles within the neighborhood; intermediate scale representation Used to describe spatial connectivity topology within a specific local area; global scale representation. Used to describe the macroscopic spatial topology and global connectivity of the entire environment in which an intelligent agent exists.

[0033] The relational cognitive graph construction module includes: a node similarity measurement unit, a node dynamic management unit, a topology edge connection establishment unit, and a topology update calibration unit. Specifically: the node similarity measurement unit performs high-dimensional feature space similarity measurement processing (e.g., calculating feature cosine similarity) based on the current multi-scale spatiotemporal state representation and the historical topology node features already existing in the relational memory module, obtaining a similarity evaluation result between the current state and existing nodes; the node dynamic management unit performs conditional branch routing processing based on whether the similarity evaluation result meets a preset similarity threshold condition. If it meets the threshold, it outputs an update instruction for existing nodes; otherwise, it outputs an update instruction for existing nodes. If the conditions are met, a new node creation instruction is output and a unique global topology node index is dynamically assigned, resulting in a dynamic management result for the topology graph nodes. The topology edge connection establishment unit performs topology reachability assessment based on the dynamic management result of the topology graph nodes, the historical time step flow, and the agent's action execution records, and dynamically establishes topology edge connections between the current node and existing neighboring nodes, representing temporal continuity, spatial proximity, or action transfer relationships. The topology update calibration unit performs adaptive adjustment of edge weights and confidence based on newly acquired environmental interaction feedback information and online optimization control information, resulting in a corrected and improved topology structure update result for the relational cognitive graph.

[0034] The relational memory module includes a writing unit, a retrieval unit, and an update unit. The writing unit stores the nodes, topological edge connections, and corresponding feature matrices of the new topological graph generated by the relational cognitive graph construction module. The retrieval unit uses the current state representation as an index to perform local association matching retrieval in the existing topological database to obtain the local historical experience context information most relevant to the current state topology. The update unit corrects and dynamically covers the node feature representations or edge topological connection weights in the existing topological graph based on the observation residuals extracted in the new interaction cycle.

[0035] The generative belief prediction module includes: a variational posterior probability model, a spatiotemporal context encoding unit, a future state generation sampling unit, a prediction variance measurement unit, and a subjective confidence scoring unit. Specifically: the variational posterior probability model probabilistically represents the uncertainty of environmental state transitions, describing the multivariate joint state transition distribution from history to the future in a continuous probability space; the spatiotemporal context encoding unit performs multi-feature dimension concatenation and context embedding encoding based on the current state representation, historical experience context information returned by the retrieval unit, and a given sequence of candidate exploration actions, obtaining a feature latent variable context vector representation; and the future state generation sampling unit outputs the mean vector of the future state distribution through a multilayer perceptron (MLP) prediction network based on the latent variable context vector representation. With log-variance vector The predicted future multi-scale state representation is obtained by combining standard normal distribution noise with reparameterization sampling. The prediction variance metric unit outputs the prediction variance vector based on the variational posterior probability model during the aforementioned computation process. Calculate the L2 norm of the variance vector. The result serves as a metric reflecting the strength of the agent's "epistemological uncertainty" regarding the candidate decision path. Based on the metric result, the subjective confidence scoring unit performs nonlinear confidence smoothing calculation through a negative exponential penalty function term to obtain a subjective confidence score that indicates the strength of the predictive certainty of future behavior outcomes.

[0036] The negative exponential penalty function term .

[0037] The navigation decision module includes: a value gain prediction unit, a confidence regularization adjustment unit, a candidate action comprehensive scoring unit, and an optimal action decision execution unit, wherein: the value gain prediction unit is based on the future multi-scale state representation results. In addition to the global navigation mission objective, the system performs target proximity estimation, feasible path cost calculation, and physical collision risk factor prediction to obtain a navigation decision evaluation score; the confidence regularization adjustment unit adjusts the subjective confidence score output by the generative belief prediction module. The original value return estimation results are subjected to multiplicative weighting or nonlinear regularization constraint adjustment as a penalty control term to restrict the agent from blindly entering the unknown region with extremely low confidence. The candidate action comprehensive scoring unit performs comprehensive algebraic superposition calculation based on the various indicators after confidence regularization adjustment to obtain the expected action score distribution results of all currently selectable candidate actions. The optimal action decision execution unit performs maximum value retrieval (Argmax) or constrained optimal probability distribution action selection processing based on the expected action score distribution results, selects the action that meets the preset conditions as the best navigation action at the current moment, and outputs it to the chassis execution mechanism.

[0038] The navigation decision evaluation score is obtained by constructing a multi-factor coupled action evaluation objective function model for each candidate action. The derived future predicted state is quantitatively scored from multiple dimensions, specifically: ,in: Indicates the current time step. Indicates the navigation decision evaluation score; Indicates time Candidate navigation actions; Indicates the degree of proximity to the target; Indicates path cost; This indicates the environmental risk cost; The subjective confidence score serves as an adaptive adjustment coefficient for value gains, enabling the agent to automatically weigh exploration risks and path energy consumption in unknown and complex environments, thereby achieving safe and efficient navigation decisions that best align with biological cognitive intuition.

[0039] The target proximity refers to the geometric deviation between the actual location reached by the agent along the path and the preset endpoint. Specifically, it is a quantitative measure of the geometric progress of the agent's spatial coordinates in approaching the task navigation target point after the agent moves to the future predicted location derived from the candidate action. These are the current two-dimensional physical space coordinates; The coordinates of the global target point for the task; The two-dimensional predicted physical spatial coordinates of the agent's future moment, obtained from the confidence world model; This represents the Euclidean distance L2 norm.

[0040] The path cost includes the energy and time costs incurred by the agent after traversing the trajectory. Specifically, it refers to the physical overhead penalties incurred by the agent in kinetic energy consumption and maintaining smooth motion while reducing detours during action, in order to drive the agent to generate a straighter trajectory. Candidate exploration actions The linear velocity and angular velocity control command parameters included; This is the vector of the actual historical actions executed at the previous moment; , This is the preset weighted penalty coefficient.

[0041] The risk factors include: uneven roads, potential collisions, etc., which are measured by a physical collision risk factor. Specifically, this is an explicit quantitative penalty for the potential danger of the agent's physical outline colliding with the boundary of irregular obstacles in the environment at the predicted spatial location. Predicting states for future moments The included geometric cell model in the azimuth angle Predicted Euclidean distance down to the environmental boundary; The hard threshold radius parameter for safe collisions; For sensitivity variance; This is a conditional indicator function. The function outputs 1 when the predicted boundary distance is less than the safe radius, and 0 otherwise.

[0042] Technical effect

[0043] This invention innovatively constructs a brain-like geometric cell model with Gaussian and von Mises product integral properties through local geometric sensing units. This mechanism can directly and explicitly quantify physical boundary features in polar coordinates and seamlessly integrate precise local geometric boundaries with macroscopic global topological relationships in a hierarchical manner. This spatial coding technique completely eliminates the dependence on external GNSS signals, high-precision global maps, and prior knowledge of the macroscopic environment, enabling agents to have extremely high tolerance to sensor ranging noise and occlusion interference in underground, indoor, or GPS-denied blind zones, ensuring the stable construction of internal spatial representations.

[0044] To address the limitation of existing black-box sequence prediction networks in quantifying epistemological uncertainty, this invention constructs a generative belief prediction module based on the Confidence World Model (BIW). This module not only infers future multi-scale state transitions but also extracts variance-norm 2 to quantify the prediction uncertainty arising from the agent's lack of experience, thereby adaptively outputting a subjective confidence score. By using the confidence score as a dynamic regularization adjustment coefficient for the action scoring objective function, the agent can keenly identify "cognitive blind spots" in unfamiliar environments and automatically make intelligent trade-offs between avoiding collision risks, saving action energy consumption, and acquiring highly novel boundary information, significantly improving the decision-making stability and safety of online autonomous exploration.

[0045] This invention utilizes a spatiotemporal dual-dimensional adaptive compression mechanism in the relational memory module. In the spatial dimension, it leverages motion information feature functions. The system dynamically determines spatial neighborhoods and performs adaptive sparsity management on redundant nodes. In the time dimension, a closed-loop timestamp mechanism is used to merge and batch-process nearby topological edges and scene nodes. This improvement not only significantly reduces the storage scale of map nodes and edges (e.g., the number of nodes in the experimental experience map is reduced by nearly 70%), but also completely solves the computational paralysis problem caused by historical information redundancy in complex, large-scale scenes, ensuring the system's online real-time operation capability on low-computing-power intelligent agent edge devices. In navigation control decision-making, this invention uses a target guidance vector generation unit to calculate adaptive vectors based on the geometric center of the intelligent agent's moving body. Furthermore, it is coupled with the geometric cell boundary perception mechanism for closed-loop control. In the scoring and evaluation model, the path cost term and the environmental risk term are coupled and cooperate in a multi-factor manner, explicitly limiting the abrupt changes and frequent U-turns of control commands. This mechanism drives the agent to generate a more straight and smooth global motion trajectory in narrow, multi-structured, unknown long-distance environments, effectively eliminating the phenomenon of ineffective detours, making the motion performance highly consistent with the biological cognitive intuition of saving energy, and greatly improving the long-distance operation efficiency of the embodied agent. Attached Figure Description

[0046] Figure 1 This is a flowchart of the present invention;

[0047] Figure 2This is a flowchart of an implementation example;

[0048] Figure 3 A schematic diagram illustrating relation structure construction and relation memory update;

[0049] Figure 4 This is a schematic diagram of the generative prediction and navigation decision-making process;

[0050] Figure 5 This is a schematic diagram illustrating the effect of an embodiment after the downstream task is completed;

[0051] Figure 6 A schematic diagram illustrating the effect of an example of constructing a cognitive relationship graph. Detailed Implementation Example 1

[0052] like Figure 5 As shown, this embodiment relates to target navigation in an unknown indoor environment, i.e., the intelligent agent is deployed in an unknown indoor environment, which includes structures such as corridors, rooms and corners, and may have obstructions.

[0053] To demonstrate the technical effectiveness of this invention, a comparative experiment on long-distance autonomous exploration mapping and target-driven navigation was conducted on a robot operating system simulation platform. In this environment, the present invention (BIW) was quantitatively compared with existing mainstream non-brain-like world models (Nomad, NWM). Specific experimental results are as follows:

[0054] In a completely unknown indoor environment, the success rate of this invention can reach 90%. The above quantitative simulation comparison data fully demonstrates that this invention can plan a more direct, smoother, and detour-free globally optimal navigation trajectory in a completely unknown environment with extremely low computational and storage resource consumption.

[0055] During the actual execution of this indoor downstream task, the online closed-loop iterative update of this invention specifically includes the following top-down operation steps:

[0056] Step 1: During navigation, acquire environmental observation information, including visual information, depth information, and motion information, and execute navigation according to the mission objective.

[0057] Step 2: Extract features from the observed information and fuse them with historical observation information and action information to obtain the state representation at the current moment. Based on this, map the state representation to multiple scales, including: local scale representation, regional scale representation, and global scale representation.

[0058] Step 3: Based on the multi-scale state representation, construct the nodes in the relational structure and match them with existing nodes. When the similarity condition is met, update the corresponding node; when not, create a new node and establish a connection with the existing node. Simultaneously, adjust the connection relationships based on the action execution results.

[0059] Step 4: Retrieve historical information related to the current state from relational memory, and combine the current state with candidate actions to make predictions and obtain the result estimates corresponding to different actions.

[0060] Step 5: The generative belief prediction module predicts the current state. The retrieval unit returns local historical experience context information and a given sequence of candidate actions. The input is used in a Confidence World Model (BIW) for probabilistic variational inference prediction, and the multi-scale state prediction results for future time steps are output through reparameterized sampling. and the prediction variance vector The subjective confidence scoring unit extracts the L2 norm of the variance vector and calculates the subjective confidence score, which quantifies the strength of certainty in future predictions, based on the negative exponential penalty function. .

[0061] Step 6: The navigation decision module reads the predicted state. The algorithm calculates the target approach progress of candidate actions, linear velocity energy consumption and path cost of action abrupt changes, as well as the risk penalty for collision with obstacles ahead; subsequently, it calculates the subjective confidence score. An adaptive adjustment coefficient, representing the value gain, is injected into the comprehensive scoring objective function to comprehensively score each candidate action. The action with the highest score is selected as the best navigation action at the current moment and output to the chassis hardware actuator.

[0062] Step 7: The agent executes the selected optimal action, causing a positional shift in physical space and proceeding to the next time step. The online update module calculates the state reconstruction loss and divergence loss between the predicted state and the actual environmental observation online, and updates the parameters of the confidence world model through the backpropagation algorithm, so as to realize the incremental and continuous accumulation and utilization of environmental information and complete the closed-loop iteration of the whole process. Example 2

[0063] like Figure 6 As shown, this embodiment relates to the construction of a cognitive map in an unknown indoor or outdoor open environment. This environment is characterized by its large size, open structure, and uneven information distribution.

[0064] This embodiment specifically includes: importing an unknown campus scene spanning hundreds of meters into a simulation platform to quantitatively verify the BIW (Browser-Based Learning) relational cognitive graph construction. After completing the full closed-loop exploration, this method reduces redundant storage by 50% through spatiotemporal dual-dimensional compression, and the graph construction time is also lower than that of the comparative methods, demonstrating efficient cognitive graph construction and resistance to memory overload.

[0065] Step 1: Real-time acquisition of multi-source observations and motion sequences, encoding and generating multi-scale spatiotemporal state representations containing medium- and long-term spatial displacement information. ;

[0066] Step 2: Based on Perform online node topology incremental expansion and execute two-dimensional adaptive pruning clustering in the spatial and temporal domains to construct a sparse topology cognitive graph.

[0067] Step 3: Predict future states using the Confidence World Model (BIW), combine subjective confidence scores and multifactor evaluation functions to output the optimal action, and dynamically update the experience memory topology online.

[0068] Step 4: In this scenario, the state representation integrates local observation information with medium- to long-term location information based on odometry and online integration of historical motion trajectories, explicitly incorporating macroscopic understanding of long-distance outdoor displacement and environmental boundaries.

[0069] When constructing the relational structure, the similarity measurement unit evaluates the feature overlap between the current state and existing nodes. If the similarity threshold is not met, the dynamic routing unit creates new nodes and assigns indices, while the edge building unit establishes topological edges based on the trajectory. Subsequently, motion information feature functions are used in the spatial domain. Neighborhood determination is performed to eliminate local redundant points; in the time domain, edge and scene nodes belonging to the same closed-loop region are merged and batch-processed using loopback timestamps, highly compressed, and stored in the global experience map. Through the above methods, environmental space and abstract information are gradually constructed in an unknown environment.

[0070] Compared with existing technologies, this method adopts a motion feature function-based approach in cognitive map construction. This involves spatially adaptive sparse technology and time-domain topological edge clustering technology based on loopback timestamps. It reduces graph storage redundancy by 50%, significantly accelerating coverage speed while fundamentally addressing memory redundancy and storage bottlenecks in large, complex scenarios with long-term operation.

[0071] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A brain-like navigation method based on multi-scale spatio-temporal representation and relational memory, characterized in that, The system collects environmental observation information and corresponding historical observation sequences, action sequences, or state sequences at the current moment. It then performs hierarchical multi-scale spatiotemporal encoding on the environmental observation information and historical observation sequences based on the fusion of a geometric cell model and a self-attention mechanism to obtain the multi-scale state representation at the current moment. Based on the multi-scale state representation, it constructs the nodes in the relational structure and the connections between nodes. The relational structure is then updated, and based on the historical relational structure, the current state, and candidate actions, the system uses deep learning to predict the future state or expected result, thereby determining and executing the current optimized action. Finally, it iterates through state updates and relational structure updates based on the actual results to achieve brain-like navigation. The geometric cell coding model is specifically as follows: ,in: The two-dimensional coordinates of the cell's egocentric space. For distance, Let azimuth be the azimuth angle, and its domain be . For intelligent agents in an omnidirectional perspective The depth sensor detects the Euclidean distance to the environmental boundary, representing a physical measurement of local obstacles. The Gaussian variance parameter for distance sensing is used as a control term to adjust the system's tolerance to sensor ranging noise and uncertainty. The current average vector direction (MVD) of the agent, calculated using the head orientation cell (HDC) model, serves as a dynamic reference benchmark for spatial encoding. The angular concentration parameter is used to control the sharpness of the peak value in the von Mises distribution. It is a zero-order modified Bessel function of the first kind.

2. The brain-like navigation method based on multi-scale spatiotemporal representation and relational memory according to claim 1, characterized in that, The hierarchical multi-scale spatiotemporal coding refers to: discretizing and mapping the physical boundary distance information and head direction information from the omnidirectional perspective of the agent using a geometric cell coding model to obtain a local representation matrix based on geometric cells. Then, multi-scale global aggregation and global structural attention encoding are performed, specifically including: Step A, Discretization Mapping: The agent acquires data in real time through sensors. A discrete perspective Euclidean distance of the environmental boundary and head direction angle reference value and build A two-dimensional geometric neuron sheet array serves as a spatial topological carrier; each coordinate unit is computed through traversal. The firing rate intensity of neurons at a given location is explicitly quantified to represent the physical boundary features of the agent from an omnidirectional perspective, ultimately outputting a complete and precise representation of the local spatial geometry. 3D matrix, as a local representation matrix based on geometric cells Specifically, it refers to the intensity of neuronal firing rate. ,in: Discrete position Geometric cell response intensity at the location Discrete number of environmental observation directions For the first Discrete observation directions For direction The detected environmental boundary distance For the first Discrete distance scale For the first Discrete azimuth angles Reference angle for the current head orientation of the agent Distance coding scale parameter Parameters in the direction encoding set For zero-order modified Bessel function For the angular discrete interval, satisfying These are indexes for direction, distance, and angle, respectively. Step B, Multi-scale Global Aggregation: Extracting regional scale representations through spatial pooling and nonlinear activation functions. Aggregate neighborhood topology information, specifically: ,in: Current moment The regional scale representation matrix; Nonlinear activation functions are used to introduce nonlinear feature maps; Preset regional feature transformation weight matrix; Spatial pooling operators are used to downsample and reduce the spatial dimensionality of input features; Current moment The local representation matrix based on the geometric cell is used as the input to the spatial pooling operator; Preset region feature transformation bias vector Step C, Global Structural Attention Encoding: An improved method is used to encode regional features based on a self-attention mechanism. Linear projection is used to calculate global topological connectivity features based on the query vector (Q), key vector (K), and value vector (V). Specifically: ,in: The scaling dimension of the feature vector is a scalar.

3. The brain-like navigation method and system based on multi-scale spatiotemporal representation and relational memory according to claim 1, characterized in that, The deep learning predictions include: Step 1: Construct the spatiotemporally correlated generative state transition inputs, specifically including: 1.1 Extracting the multi-scale spatial state representation at the current moment Furthermore, by combining the relational memory module, historical memory context features related to the current state topology are retrieved. ; 1.2 Acquiring candidate exploration actions of the agent in the action space, including motion commands such as linear velocity and angular velocity. ,Will , Concatenate along the feature dimension; 1.3 A multilayer perceptron (MLP) encoder is used to perform nonlinear mapping on the stitched features, fusing spatial topology, action intent, and historical experience to extract generative latent variable context vectors. ,in: These are the learnable weight matrix and bias term of the context encoding network, respectively. This indicates a concatenation operation of feature dimensions; Step 2: Perform future state posterior distribution prediction based on variational inference, specifically including: 2.1 Latent variable context vector The inputs are fed into a two-branch variational inference network to predict the mean distribution of future states. With log variance This is used to simulate the uncertainty of environmental state transitions, specifically: , ,in: To predict variance, For predicted future moments The mean vector of the state distribution; This is the learnable weight matrix for the mean prediction branch in the variational inference network; For a moment Encoded latent variable context vector; is the learnable bias vector for the mean prediction branch in the variational inference network; It is the natural logarithm operator; For predicted future moments The predicted variance vector of the state distribution; This is the learnable weight matrix for the log-variance prediction branch in the variational inference network; This is the learnable bias vector for the log-variance prediction branch in the variational inference network; 2.2 Employing reparameterization techniques, from the standard normal distribution Medium sampling noise This leads to the generation of a predicted state representation that includes spatial geometry and topological information of the predicted environment. ,in: This indicates element-wise multiplication. For the generated future moment The predicted state is a vector; For predicted future moments The mean vector of the state distribution; This is a random noise vector sampled from a standard normal distribution; For element-wise multiplication (Hadamard product); For the natural constant Exponential operators with base 0; It is the natural logarithm operator; For predicted future moments The variance vector of the state distribution; Step 3: Quantify cognitive uncertainty and output subjective confidence scores, specifically including: 3.1 Based on prediction variance The L2 norm of this variance vector is extracted as the basis for measuring epistemic uncertainty. 3.2 Based on the exponential penalty function, calculate the subjective confidence score corresponding to this prediction and use it as the core weight for evaluating path safety and exploration value in subsequent navigation decisions, specifically: ,in: The scaling hyperparameter is used to adjust the sensitivity of the confidence level. .

4. The brain-like navigation method and system based on multi-scale spatiotemporal representation and relational memory according to claim 1, characterized in that, The iterative process of updating the state and relational structure based on actual perception results includes the following specific iterative steps: Step 1: Execute the target action and acquire real-world environment observations: Based on the optimal action output by the navigation decision module, the agent performs movement in the physical or simulated environment, triggering the next time step. Multi-source sensor data acquisition, and encoding through a multi-scale spatiotemporal representation module to obtain a true spatiotemporal state representation. ; Step 2: Calculate the joint loss and the gradient using the backpropagation algorithm. Update the network parameters of the confidence world model online to ensure continuous convergence of the model in unknown environments and improve the accuracy of state prediction. Specifically: Among them: state reconstruction loss Distribution divergence loss , The regularization coefficient is , i.e. The divergence weight coefficient is used to weigh the equilibrium state reconstruction accuracy against the probability distribution regularization constraint. To predict the state, This represents the actual state; Step 3: Update the relational cognitive map and calibrate the experience memory retrieval: Calculate the current true state. The similarity to existing node states in the relational memory module is considered if the maximum similarity is lower than a preset novelty threshold. Then As a new node, it is stored in the relational cognitive graph, and new edge connections are established based on the executed action trajectory. If it exceeds the threshold, the access frequency weight of the corresponding node is increased, the memory topology graph is dynamically calibrated, and more accurate empirical context is provided for the generative belief prediction of the next cycle.

5. A brain-like navigation system based on multi-scale spatiotemporal representation and relational memory, implementing the method of any one of claims 1-4, characterized in that, include: The system comprises a multi-source observation acquisition module, a multi-scale spatiotemporal representation module, a relational cognitive graph construction module, a relational memory module, a generative belief prediction module, a navigation decision module, and an online update module. Specifically: the multi-source observation acquisition module acquires observational information of the agent in the environment and performs hierarchical multi-scale representation to express environmental information within different ranges; the multi-scale spatiotemporal representation module encodes current and historical information to obtain state representations at different spatial scales and temporal ranges; the relational cognitive graph construction module constructs a relational structure composed of nodes and edges based on the state representation; the relational memory module writes, retrieves, and updates the relational structure to manage historical information; the generative belief prediction module predicts state change trends, target proximity, or probability of action success based on the current state, relational memory, and candidate actions; the navigation decision module determines the navigation action at the current moment based on the prediction results; and the online update module updates the state representation, relational structure, and relational memory during navigation based on newly acquired observational information. The observation information includes: visual observation, depth observation, laser observation, inertial observation, odometry information, or historical trajectory information; In the relational structure, nodes are used to represent the state at each moment, and edges are used to represent temporal continuity, spatial proximity, or action transition relationships, with weights or confidence levels set. The target proximity refers to the geometric progress quantification of the agent's spatial coordinates approaching the task navigation target point after the agent moves to the future predicted position derived from the candidate action inference, specifically: These are the current two-dimensional physical space coordinates; The coordinates of the global target point for the task; The two-dimensional predicted physical spatial coordinates of the agent's future moment, obtained from the confidence world model; This represents the Euclidean distance L2 norm.

6. The neuromorphic navigation system according to claim 5, characterized in that, The multi-scale spatiotemporal representation module includes: a local geometric perception unit, a regional topology extraction unit, a global relation association unit, and a multi-scale spatiotemporal fusion unit. Specifically: the local geometric perception unit, based on depth data or laser data and dynamic reference direction information from the current environmental observation information, performs brain-like geometric cell feature reconstruction and polar coordinate discretization feature mapping in polar coordinates to obtain a local-scale state representation result that explicitly quantifies local physical boundary features; the regional topology extraction unit, based on the local-scale state representation result output by the local geometric perception unit, performs neighborhood feature space pooling dimensionality reduction and regional topological connectivity nonlinear feature activation processing to obtain a representation describing a specific neighborhood and... The regional scale state representation results of the connectivity structure relationship are obtained. The global relationship association unit performs long-distance spatial connectivity encoding and global environment layout modeling based on the regional scale state representation results output by the regional topology extraction unit to obtain the global scale state representation results describing the global topological connectivity relationship of the overall environment. The multi-scale spatiotemporal fusion unit performs multi-level feature concatenation and spatiotemporal correlation transformation feature fusion processing based on the local scale state representation results, regional scale state representation results, global scale state representation results and input historical observation sequence and action sequence information to obtain the complete multi-scale spatiotemporal state representation results representing the multi-scale topological features inside the environment at the current moment. The state representation includes: local scale representation Intermediate scale representation and global scale representation , where: local scale representation Used to describe the geometric information of physical boundaries and passable obstacles within the neighborhood; intermediate scale representation Used to describe spatial connectivity topology within a specific local area; global scale representation. Used to describe the macroscopic spatial topology and global connectivity of the entire environment in which an intelligent agent exists.

7. The neuromorphic navigation system according to claim 5, characterized in that, The relational cognitive graph construction module includes: a node similarity measurement unit, a node dynamic management unit, a topology edge connection establishment unit, and a topology update calibration unit. Specifically: the node similarity measurement unit performs high-dimensional feature space similarity measurement based on the current multi-scale spatiotemporal state representation and the historical topology node features already existing in the relational memory module, obtaining a similarity evaluation result between the current state and existing nodes; the node dynamic management unit performs conditional branch routing based on whether the similarity evaluation result meets a preset similarity threshold condition. If it does, it outputs an existing node update instruction; otherwise, it outputs a new node creation instruction and dynamically assigns a unique global topology node index, obtaining the topology graph node dynamic management result; the topology edge connection establishment unit performs topology reachability evaluation based on the topology graph node dynamic management result, historical time step flow, and agent action execution records, dynamically establishing topology edge connections representing temporal continuity, spatial proximity, or action transfer relationships between the current node and existing adjacent nodes; the topology update calibration unit performs adaptive adjustment of edge weights and confidence based on newly acquired environmental interaction feedback information and online optimization control information, obtaining a corrected and improved relational cognitive graph topology structure update result.

8. The neuromorphic navigation system according to claim 5, characterized in that, The relational memory module includes a writing unit, a retrieval unit, and an update unit. The writing unit stores the nodes, topological edge connections, and corresponding feature matrices of the new topological graph generated by the relational cognitive graph construction module. The retrieval unit uses the current state representation as an index to perform local association matching retrieval in the existing topological database to obtain the local historical experience context information most relevant to the current state topology. The update unit corrects and dynamically covers the node feature representations or edge topological connection weights in the existing topological graph based on the observation residuals extracted in the new interaction cycle.

9. The neuromorphic navigation system according to claim 5, characterized in that, The generative belief prediction module includes: a variational posterior probability model, a spatiotemporal context encoding unit, a future state generation sampling unit, a prediction variance measurement unit, and a subjective confidence scoring unit. Specifically: the variational posterior probability model probabilistically represents the uncertainty of environmental state transitions, describing the multivariate joint state transition distribution from history to the future in a continuous probability space; the spatiotemporal context encoding unit performs multi-feature dimension concatenation and context embedding encoding based on the current state representation, historical experience context information returned by the retrieval unit, and a given sequence of candidate exploration actions, obtaining a feature latent variable context vector representation; and the future state generation sampling unit outputs the mean vector of the future state distribution through a multilayer perceptron (MLP) prediction network based on the latent variable context vector representation. With log-variance vector The predicted future multi-scale state representation is obtained by combining standard normal distribution noise with reparameterization sampling. The prediction variance metric unit outputs the prediction variance vector based on the variational posterior probability model during the aforementioned computation process. Calculate the L2 norm of the variance vector. The result serves as a metric reflecting the strength of the agent's "epistemological uncertainty" regarding the candidate decision path. Based on the metric result, the subjective confidence scoring unit performs nonlinear confidence smoothing calculation through a negative exponential penalty function term to obtain a subjective confidence score that indicates the strength of the predictive certainty of future behavior outcomes. The negative exponential penalty function term .

10. The neuromorphic navigation system according to claim 5, characterized in that, The navigation decision module includes: a value gain prediction unit, a confidence regularization adjustment unit, a candidate action comprehensive scoring unit, and an optimal action decision execution unit, wherein: the value gain prediction unit is based on the future multi-scale state representation results. In addition to the global navigation mission objective, the system performs target proximity estimation, feasible path cost calculation, and physical collision risk factor prediction to obtain a navigation decision evaluation score; the confidence regularization adjustment unit adjusts the subjective confidence score output by the generative belief prediction module. The original value return estimation results are subjected to multiplicative weighting or nonlinear regularization constraint adjustment as a penalty control term to restrict the agent from blindly entering the unknown region with extremely low confidence. The candidate action comprehensive scoring unit performs comprehensive algebraic superposition calculation based on the various indicators after confidence regularization adjustment to obtain the expected action score distribution results of all currently selectable candidate actions. The optimal action decision execution unit performs maximum value retrieval (Argmax) or constrained optimal probability distribution action selection processing based on the expected action score distribution results, selects the action that meets the preset conditions as the best navigation action at the current moment and outputs it to the chassis execution mechanism. The navigation decision evaluation score is obtained by constructing a multi-factor coupled action evaluation objective function model for each candidate action. The derived future predicted state is quantitatively scored from multiple dimensions, specifically: ,in: Indicates the current time step. Indicates the navigation decision evaluation score; Indicates time Candidate navigation actions; Indicates the degree of proximity to the target; Indicates path cost; This indicates the environmental risk cost; The subjective confidence score serves as an adaptive adjustment coefficient for value gains, enabling the agent to automatically weigh exploration risks and path energy consumption in unknown and complex environments, thereby achieving safe and efficient navigation decisions that best align with biological cognitive intuition.

Images