A body-oriented intelligent navigation base module topology memory enhancement method for complex indoor environment

Abstract

The application discloses a kind of indoor complex environment-oriented embodied intelligent navigation base module topology memory enhancement method, belong to artificial intelligence and robot navigation technical field.The method is aimed at the deficiency of existing end-to-end navigation model in long-range planning, dynamic obstacle avoidance and cross-scene generalization ability, first, the semantic features of environment are extracted by vision-language model, and the environment base module containing spatial function and layout priori is constructed.Second, based on the similarity of base module, a hierarchical topological graph is dynamically maintained, the continuous space is discretized into semantic nodes, and the connectivity and travel cost between nodes are predicted using base module.Finally, combined with online reinforcement learning and graph search algorithm, the closed-loop control from macro-topological path planning to micro-action execution is realized.The present application introduces base module as prior knowledge to guide the construction and update of topological graph, significantly reduces invalid exploration, and improves the navigation success rate and path optimality of robot in unknown, dynamic and complex indoor environment.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robot navigation technology, specifically to an embodied intelligent navigation fundamental model topology memory enhancement method for complex indoor environments. Background Technology

[0002] With the widespread application of service robots, home companion robots, and warehousing and logistics AGVs in complex indoor scenarios, autonomous navigation technology has become a core supporting capability. Traditional indoor navigation solutions mainly rely on the combination of simultaneous localization and mapping (SLAM) technology and metric maps (such as occupancy grid maps or point cloud maps). This type of method uses sensors deployed in the environment to build a high-precision geometric map in real time and perform particle filtering for localization, supplemented by classic path planning algorithms such as A* and DWA to achieve obstacle avoidance and movement. However, this type of geometric metric-based method has significant inherent drawbacks: First, it has poor robustness to dynamic environmental changes. When faced with high-frequency movement of people, temporary stacks of goods, or changes in the position of furniture, static maps are prone to cumulative errors or even failure. Second, the computational overhead is huge. Maintaining a high-precision, dense map in real time places stringent demands on the robot's onboard computing power, limiting its deployment on low-power embedded platforms. Finally, pure geometric representation lacks an understanding of high-level semantics such as "kitchen" and "corridor," making it difficult for the robot to execute semantic-based natural language commands such as "go to the sofa," thus limiting its level of intelligence. To overcome the rigidity of traditional geometric methods, end-to-end navigation paradigms based on deep reinforcement learning (DRL) have gradually become a research hotspot in recent years. These methods attempt to directly establish a nonlinear mapping from raw sensor inputs (such as RGB-D images and LiDAR data) to underlying motion control outputs (linear velocity and angular velocity). While end-to-end methods have improved the robot's obstacle avoidance flexibility in unstructured environments to some extent, their practical application still faces significant challenges. The core problem is extremely low sample efficiency and weak generalization ability. DRL models typically require millions of steps of interaction data for training, and the learned policies often overfit to specific training environment textures, lighting, and layouts. Once the robot enters an unfamiliar room (such as switching from an office to a home environment) or the furniture layout changes, the model's performance drops sharply. Furthermore, end-to-end models lack explicit memory mechanisms, typically using RNNs or Transformers to implicitly encode historical information. They struggle to utilize long-term spatial priors like "elevators are usually located at the end of corridors," making them prone to "deadlock" or repetitive circling behaviors in long-range navigation tasks, and thus unsuitable for navigation in large-scale indoor spaces. To address the aforementioned shortcomings, academia has proposed a compromise combining topology navigation and semantic mapping, aiming to reduce reliance on precise geometry by constructing an abstract graph structure of nodes and edges. However, most existing topology navigation methods employ a node generation strategy based on distance thresholds, forcibly creating new nodes after the robot has moved a fixed distance. This mechanical approach ignores the semantic boundaries of the environment, easily leading to node redundancy or missed critical intersections. More importantly, existing solutions typically treat semantic information as an adjunct label to topology nodes, rather than the core driving force behind topological evolution, failing to effectively simulate the process by which humans use "schemas" to understand the world. When humans enter a new room, they don't start exploring from scratch but use existing "room layout schemas" in their brains to predict the possible locations of doors, windows, and furniture, thus quickly planning their exploration path. Current navigation algorithms lack this crucial intermediate cognitive layer—how to use prior knowledge to construct dynamic environmental schemas and use schema predictions to guide the construction, updating, and path planning of topological memory. Therefore, a new method is needed that integrates cognitive science and artificial intelligence technologies to enhance topological memory and solve the problems of long-range navigation and rapid generalization in complex indoor environments. To address this, we propose an embodied intelligent navigation fundamental model topology memory enhancement method for complex indoor environments. Summary of the Invention

[0003] To achieve the above objectives, the present invention provides the following technical solution: a method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments, comprising the following steps: S1 Environment Model Construction Steps: Use a pre-trained visual-language model to process the robot's current multimodal observation data, extract high-level semantic features, and construct a dynamic environment model that includes spatial functional attributes, object layout priors, and physical constraints. S2 Topology Node Generation and Management Steps: Based on the semantic saliency and entropy change of the environmental schema, the continuous physical space is dynamically discretized into topology nodes with semantic labels, and a hierarchical topology memory graph is established. The S3 fundamental model-enhanced edge weight update steps are as follows: the environmental fundamental model is used to predict the connectivity and passage cost between topological nodes, and the weights of edges in the topological graph are dynamically updated in combination with real-time sensing data to form a representation of passable regions guided by the fundamental model. S4 Hybrid Planning and Decision-Making Steps: In response to navigation target instructions, a global path search is performed on the hierarchical topology memory graph to obtain a macroscopic path sequence, and combined with local real-time perception information, the underlying motion control instructions are output through a reinforcement learning strategy. S5 Closed-Loop Memory Correction Step: Based on the actual feedback results after the action is executed, the parameter confidence and topology structure of the environmental model are corrected and optimized online.

[0004] Preferably, the environmental fundamental model is represented by a three-layer probabilistic graphical structure, including: Semantic layer: used to represent the functional zoning categories of rooms and their confidence distribution, obtained by classifying images using VLM; Layout layer: used to describe the spatial co-occurrence relationship, relative pose and support relationship between objects, and is constructed by generating a scene graph network; Physical layer: used to characterize ground accessibility, friction coefficient, slope and kinematic constraints of dynamic obstacles, updated through depth information and temporal filtering.

[0005] Preferably, in the topology node generation and management step, the node creation conditions are determined by the fundamental module entropy value: When the KL divergence between the environmental schema corresponding to the current observation and the environmental schema associated with the current topology node exceeds a preset dynamic threshold. If a significant semantic change occurs in the environment, a new topological node is created; otherwise, the current observation data is simply fused into the fundamental model representation of the current node.

[0006] Preferably, in the edge weight update step of the fundamental model enhancement, the connecting nodes in the topology graph and Boundary rights The calculation method is as follows:

[0007] in, It is a semantic passage cost based on environmental schema prediction (including penalties for private areas and narrow passages). It is the average travel time based on historical experience. It is the probability of dynamic obstacles being occupied at the current moment. These are the weighting coefficients.

[0008] Preferably, the hybrid planning and decision-making steps employ a two-layer reinforcement learning architecture: Top-level policy network ( ): Taking the fundamental features of the current topology node and the global target as input, output the index of the next sub-target topology node to be reached; Low-level control network ( ): Taking the current local LiDAR point cloud or depth image and the relative pose of the sub-target as input, output the linear velocity and angular velocity control values ​​of the robot; The update frequency of the top-level policy network is lower than that of the bottom-level control network.

[0009] Preferably, the top-level policy network introduces an intrinsic reward function guided by the base model during the training phase: When the robot explores and establishes new and effective topological connections, it is given a positive reward based on the schema information gain of the new connection, in order to encourage the robot to use environmental schemas for efficient exploration rather than blindly wandering randomly.

[0010] Preferably, catastrophic forgetting defense mechanisms are also included: The system maintains a long-term memory to store schema snapshots and topology maps of different floors. When a robot crosses a geofence (such as entering an elevator or stairs) and causes a sudden change in the environmental schema, the system automatically archives the topology memory of the current scene and loads the corresponding historical schema snapshot according to the identified new scene type, thus achieving the isolation and rapid reuse of long-term memory.

[0011] Preferably, the layout layer update employs a Bayesian inference framework: When the robot observes an object placement that is inconsistent with the prediction of the basic model layout layer (e.g., a chair is moved), it dynamically adjusts the posterior probability of the corresponding relationship in the layout layer based on the confidence level of the observation; when the posterior probability is lower than a preset threshold, it triggers the correction or splitting of the basic model structure. Preferably, the closed-loop memory correction step specifically includes: When performing a global path search, if multiple attempts to access a certain topology node fail, the system marks the node as "unreachable" or "occupied by a dynamic obstacle". Then, it uses the reasoning ability of the environment schema to find an alternative path. After the alternative path is successfully traversed, the connectivity state of the original failed node is updated in reverse to distinguish between "permanent structural changes" and "temporary dynamic obstacles".

[0012] Preferably, an embodied intelligent navigation system includes: The perception module is configured to acquire images and depth information of the robot's surrounding environment; A basic model construction module, coupled to the perception module, is configured to execute any of the environmental basic model construction steps described herein; The topology memory module, coupled to the base module, is configured as any of the topology node generation, edge weight update and memory correction steps described above; The decision planning module is coupled to the topology memory module and the perception module respectively, and is configured to execute any of the hybrid planning and decision steps described above; The actuator interface is configured to send the instructions generated by the decision planning module to the robot's motor driver.

[0013] Compared with existing technologies, this invention provides a method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments, which has the following beneficial effects: 1. This embodied intelligent navigation method for complex indoor environments, based on a fundamental model of topological memory, fundamentally changes the robot's spatial exploration mode by introducing an "environmental model" as a high-level cognitive prior. It solves the inefficiency and failure problems caused by the lack of a global perspective in traditional end-to-end methods during long-range planning. Specifically, utilizing the priors of "room functional layout" and "object spatial co-occurrence relationships" contained in the model, the robot can predict the optimal direction of travel based on semantic reasoning when facing intersections or open areas, rather than relying on random walks or simple geometric distances. This proactive exploration mechanism based on model prediction significantly reduces invalid paths and backtracking times. Experimental data shows that compared to traditional distance-based topology mapping methods, this invention reduces the average path length by more than 15% and improves the task success rate by more than 20% in long-range indoor navigation tasks at the 100-meter level. Especially in complex layouts such as mazes or multi-room apartments, the macroscopic guidance provided by the model effectively prevents the agent from getting stuck in local loops, significantly enhancing the determinism of navigation.

[0014] 2. This embodied intelligent navigation method for complex indoor environments, utilizing topological memory enhancement, significantly improves the system's fault tolerance and robustness by decoupling the schema from the topological memory, addressing the highly dynamic characteristics of indoor environments (such as pedestrian movement and temporary furniture relocation). Unlike traditional SLAM's stringent requirements for precise geometric consistency, this invention's topological map relies solely on semantic connectivity. When local physical changes occur (such as minor adjustments to sofa positions), the schema's layout layer corrects its confidence level through Bayesian updates without overturning the entire map structure. When encountering complete blockage (such as a closed door), the schema-enhanced edge weight mechanism quickly identifies connectivity failures and rapidly switches to alternative paths based on common sense principles such as "windows do not lead to other rooms." Furthermore, through the node management mechanism triggered by "fundamental mode entropy change", the robot is no longer limited to a fixed distance threshold, but adaptively adjusts the memory granularity according to semantic changes. This prevents it from getting lost in low-texture areas such as corridors due to odometry drift, and from wasting computing power due to excessive mapping in cluttered environments, thus maintaining stable navigation performance under various extreme indoor conditions.

[0015] 3. This embodied intelligent navigation model topology memory enhancement method for complex indoor environments overcomes the bottleneck of weak generalization ability in existing learning-based navigation algorithms, endowing robots with human-like ability to learn by analogy. By archiving and isolating the topology memories of different floors or different building types, and using a common "environmental model" as a bridge for cross-scene transfer, the robot can transfer the "elevator-corridor-workstation" spatial structure learned in the office building to a completely new office building environment with zero samples, without the need to re-collect a large amount of data for training. This transfer learning capability based on cognitive models significantly reduces deployment costs and time. At the same time, this invention solves the pain point of the difficulty in debugging the "black box" decision-making of deep learning models. Since path planning occurs on an explicit topology graph with semantic labels, human maintenance personnel can intuitively trace the logical chain of robot decision-making by visualizing topology nodes and model states (e.g., the system chooses to detour because the model predicts congestion in the corridor ahead). This not only facilitates fault diagnosis and algorithm optimization, but also provides a transparent interactive basis for establishing trust in human-machine collaboration scenarios. Detailed Implementation

[0016] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] Example An embodiment of an embodied intelligent navigation fundamental schema topology memory enhancement method for complex indoor environments. A method for enhancing the topological memory of embodied intelligent navigation fundamental schema for complex indoor environments includes the following steps: S1 Environment Model Construction Steps: Use a pre-trained visual-language model to process the robot's current multimodal observation data, extract high-level semantic features, and construct a dynamic environment model that includes spatial functional attributes, object layout priors, and physical constraints. S2 Topology Node Generation and Management Steps: Based on the semantic saliency of the environmental schema and the schema entropy change, the continuous physical space is dynamically discretized into topology nodes with semantic labels, and a hierarchical topology memory graph is established. The S3 fundamental model-enhanced edge weight update steps are as follows: the environmental fundamental model is used to predict the connectivity and passage cost between topological nodes, and the weights of edges in the topological graph are dynamically updated in combination with real-time sensing data to form a representation of passable regions guided by the fundamental model. S4 Hybrid Planning and Decision-Making Steps: In response to navigation target instructions, a global path search is performed on a hierarchical topological memory graph to obtain a macroscopic path sequence, and combined with local real-time perception information, the underlying motion control instructions are output through a reinforcement learning strategy. S5 Closed-Loop Memory Correction Steps: Based on the actual feedback results after the action is executed, the parameter confidence and topology structure of the environmental model are corrected and optimized online.

[0018] Specifically, the environmental fundamental model is represented using a three-layer probabilistic graphical structure, including: Semantic layer: used to represent the functional zoning categories of rooms and their confidence distribution, obtained by classifying images using VLM; Layout layer: used to describe the spatial co-occurrence relationship, relative pose and support relationship between objects, and is constructed by generating a scene graph network; Physical layer: used to characterize ground accessibility, friction coefficient, slope and kinematic constraints of dynamic obstacles, updated through depth information and temporal filtering.

[0019] Specifically, in the topology node generation and management steps, the node creation conditions are determined by the fundamental module entropy value: When the KL divergence between the environmental schema corresponding to the current observation and the environmental schema associated with the current topology node exceeds a preset dynamic threshold. If a significant semantic change occurs in the environment, a new topological node is created; otherwise, the current observation data is simply fused into the fundamental model representation of the current node.

[0020] Specifically, in the edge weight update step of the fundamental model enhancement, the connecting nodes in the topology graph... and Boundary rights The calculation method is as follows:

[0021] in, It is a semantic passage cost based on environmental schema prediction (including penalties for private areas and narrow passages). It is the average travel time based on historical experience. It is the probability of dynamic obstacles being occupied at the current moment. These are the weighting coefficients.

[0022] Specifically, the hybrid planning and decision-making process employs a two-layer reinforcement learning architecture: Top-level policy network ( ): Taking the fundamental features of the current topology node and the global target as input, output the index of the next sub-target topology node to be reached; Low-level control network ( ): Taking the current local LiDAR point cloud or depth image and the relative pose of the sub-target as input, output the linear velocity and angular velocity control values ​​of the robot; The update frequency of the top-level policy network is lower than that of the bottom-level control network.

[0023] Specifically, the top-level policy network introduces an intrinsic reward function guided by the schema during the training phase: When the robot explores and establishes new and effective topological connections, it is given a positive reward based on the schema information gain of the new connection, in order to encourage the robot to use environmental schemas for efficient exploration rather than blindly wandering randomly.

[0024] Specifically, this also includes catastrophic forgetting defense mechanisms: The system maintains a long-term memory to store schema snapshots and topology maps of different floors. When a robot crosses a geofence (such as entering an elevator or stairs) and causes a sudden change in the environmental schema, the system automatically archives the topology memory of the current scene and loads the corresponding historical schema snapshot according to the identified new scene type, thus achieving the isolation and rapid reuse of long-term memory.

[0025] Specifically, the layout layer updates employ a Bayesian inference framework: When the robot observes an object placement that is inconsistent with the predictions of the basic schema layout layer (e.g., a chair is moved), it dynamically adjusts the posterior probability of the corresponding relationship in the layout layer based on the confidence level of the observation. When the posterior probability falls below a preset threshold, it triggers the correction or splitting of the basic schema structure. Specifically, the closed-loop memory correction steps include: When performing a global path search, if multiple attempts to access a certain topology node fail, the system marks the node as "unreachable" or "occupied by a dynamic obstacle". Then, it uses the reasoning ability of the environment schema to find an alternative path. After the alternative path is successfully traversed, the connectivity state of the original failed node is updated in reverse to distinguish between "permanent structural changes" and "temporary dynamic obstacles".

[0026] Specifically, an embodied intelligent navigation system includes: The perception module is configured to acquire images and depth information of the robot's surrounding environment; The schema construction module, coupled with the perception module, is configured to execute any one of the environment schema construction steps. The topology memory module, coupled with the base module, can be configured as any of the topology node generation, edge weight update, and memory correction steps. The decision planning module is coupled with the topology memory module and the perception module respectively, and is configured to execute any one of the hybrid planning and decision steps; The actuator interface is configured to send instructions generated by the decision planning module to the robot's motor drivers.

[0027] Through the above technical solution, this invention fundamentally changes the robot's spatial exploration mode by introducing an "environmental schema" as a high-level cognitive prior, solving the inefficiency and failure problems caused by the lack of a global perspective in traditional end-to-end methods for long-range planning. Specifically, by utilizing the priors of "room functional layout" and "object spatial co-occurrence relationships" contained in the schema, the robot can predict the optimal direction of travel based on semantic reasoning when facing intersections or open areas, rather than relying on random walks or simple geometric distances. This proactive exploration mechanism based on schema prediction significantly reduces invalid paths and backtracking times. Experimental data shows that, compared with traditional distance-based topology mapping methods, this invention reduces the average path length by more than 15% and improves the task success rate by more than 20% in indoor long-range navigation tasks at the 100-meter level. Especially in complex floor plans with maze-like layouts or multiple rooms, the macro-level guidance provided by the schema effectively prevents the agent from getting stuck in local loops, significantly enhancing the determinism of navigation. For the highly dynamic characteristics of indoor environments (such as pedestrian movement and temporary furniture relocation), this invention greatly improves the system's fault tolerance and robustness through a decoupled design between the schema and topological memory. Unlike traditional SLAM's stringent requirements for precise geometric consistency, the topological graph of this invention relies solely on semantic connectivity. When local physical changes occur in the environment (such as minor adjustments to the sofa's position), the schema's layout layer corrects the confidence level through Bayesian updates without overturning the entire map structure. When encountering complete blockage (such as a closed door), the schema-enhanced edge weight mechanism can quickly identify connectivity failures and rapidly switch to alternative paths based on common sense such as "windows do not lead to other rooms." Furthermore, through a node management mechanism triggered by "archetype entropy change," the robot is no longer limited to a fixed distance threshold but adaptively adjusts its memory granularity based on semantic changes. This prevents it from getting lost in low-texture areas such as corridors due to odometry drift and avoids wasting computational power on excessive mapping in cluttered environments. Consequently, it maintains stable navigation performance under various extreme indoor conditions. By overcoming the bottleneck of weak generalization ability in existing learning-based navigation algorithms, it endows the robot with human-like ability to learn by analogy. By archiving and isolating the topological memories of different floors or building types, and using a universal "environmental archetype" as a bridge for cross-scene transfer, the robot can transfer the "elevator-corridor-workstation" spatial structure learned in the office building to a completely new office building environment with zero samples, without needing to re-collect large amounts of data for training. This transfer learning capability based on cognitive archetypes significantly reduces deployment costs and time. Simultaneously, this invention addresses the pain point of difficult-to-debug "black box" decision-making in deep learning models.Since path planning occurs on an explicit, semantically labeled topology graph, human operators can intuitively trace the logical chain of robot decisions by visualizing topology nodes and schema states (e.g., the system chooses to detour because the schema predicts congestion in the corridor ahead). This not only facilitates troubleshooting and algorithm optimization, but also provides a transparent interactive basis for establishing trust in human-machine collaboration scenarios.

[0028] 1. System Initialization and Environment Model Construction After the robot is powered on, the system loads a pre-trained visual-language model (VLM, such as BLIP-2) as the perceptual backbone network and initializes the hierarchical topology memory graph. .

[0029] Step S110: Instantiate the Environmental Schema.

[0030] The robot is located in the center of the living room and acquires the current RGB-D image frame. and depth map .

[0031] Semantic layer parsing: Inputting VLM, VLM identifies the image as containing "sofa," "coffee table," "television," and "carpet." Based on semantic co-occurrence probabilities, the system determines the current environment's fundamental schema type as "living room," with a confidence level of 0.92.

[0032] Layout layer construction: The scene graph generation network analyzes object relationships and infers that: the "TV" is placed against the wall, the "coffee table" is located about 0.5 meters in front of the "sofa", and the "carpet" is located below the "coffee table". These relationships are encoded into a sparse adjacency matrix as the initial state of the layout layer.

[0033] Physical layer calibration: Based on the depth map, the system detected that the height of the "carpet" area is slightly lower than the surrounding floor (approximately 1 cm difference). The physical layer records that the coefficient of friction and passage difficulty of this area are slightly higher than those of the smooth floor, but do not constitute an obstacle.

[0034] At this point, the system creates the first topology node. The aforementioned basic model information is stored as an attribute of the node.

[0035] 2. Node generation and management based on fundamental mode entropy change The robot begins to move slowly in the set direction. The system monitors changes in the fundamental modulus entropy in real time to determine whether to create a new node.

[0036] Step S210: Calculation of fundamental mode entropy and node triggering.

[0037] Define the fundamental mode entropy For current observation data and nodes KL divergence between stored fundamental modules.

[0038]

[0039] in The probability distribution representing the fundamental mode.

[0040] Scenario A (Not Triggered): The robot moves 2 meters in a straight line in the living room. Although the odometry shows a large displacement, the semantic content recognized by the VLM is still "sofa" and "TV", and the fundamental mode entropy is low. Less than the preset threshold The system determines that the region is still within the same semantic region and only updates the status. The node is positioned at the center and no new node is created. This avoids node redundancy caused by simple distance thresholds in traditional methods.

[0041] Scenario B (Trigger): The robot moves to the boundary between the living room and dining room (doorway). At this moment, the VLM observes that the "sofa" has disappeared and a "dining table" and "dining chairs" have appeared. The fundamental schema type changes from "living room" to "dining room," and the fundamental schema entropy... Far exceeding the threshold The system immediately triggers the creation of a new topology node. .

[0042] Step S220: Assign values ​​to node attributes.

[0043] New node A new environmental schema (of type "restaurant") is assigned, and its relative environment is recorded. Relative pose transformation At the same time, the system establishes a directed edge. Connect the two.

[0044] 3. Edge weight update and dynamic correction with fundamental schema enhancement The edges of a topological graph are not just geometric connections, but also a carrier for evaluating the passage cost under the guidance of the fundamental model.

[0045] Step S310: Initial edge weight prediction.

[0046] For the edge (Living room → Dining room), the basic model module analyzes the physical connection between the two. Since it requires passage through a doorway approximately 0.8 meters wide, and the basic model predicts that the doorway may contain a door, the system calculates the initial passage cost. (There is a penalty compared to version 1.0, which is different from the open ground).

[0047] Step S320: Dynamic weight fusion.

[0048] The robot attempts to pass through a doorway. If a person stands beside the doorway, partially blocking the passage, the lidar detects the dynamic obstacle.

[0049] System update edge weights W12:

[0050] Substitute the values: .

[0051] This weight reflects the overall cost of passing through the doorway at the current moment. If the dynamic obstacle disappears, When the weight is reduced to zero, it will automatically decrease.

[0052] 4. Hybrid Planning and Decision Execution Suppose a user issues a voice command: "Go to the study and get my water glass." VLM parses the command into a semantic target: "study."

[0053] Step S410: Global topology path planning.

[0054] System in topology diagram Search for the target. Currently known nodes include: (living room), (Dining room), (corridor), (bedroom).

[0055] Top-level policy network ( ) with the current node The target "study room" is the input.

[0056] Since the "study" has not yet been explored, the basic module infers based on prior knowledge: "The study is usually adjacent to the bedroom and needs to be reached through a corridor."

[0057] The system performs an A* search and plans a macroscopic path sequence: [Unknown area]

[0058] Step S420: Low-level control and obstacle avoidance.

[0059] Top-level strategy selection (Corridor) is the current sub-target.

[0060] Low-level control network ( It takes over control. It receives lidar data and discovers pathways to... There was a chair that had been left lying around on the path.

[0061] Low-level network output control commands: (Slight left turn to detour).

[0062] The robot successfully bypassed the chair and arrived at... Nearby. The visual sensor confirms arrival in the corridor, triggering a node arrival event.

[0063] Step S430: Exploration of unknown areas and reuse of basic models.

[0064] arrive The robot then discovers an unrecorded door at the end of the corridor. The top-level strategy activates exploration mode. The base schema module indicates: "There's a high probability there's a room behind the door." The robot pushes the door open, creating a new node. VLM recognized this as a "study". The system will... Incorporate into the topology graph and connect to it. At this point, the global path is established: → → → .

[0065] 5. Closed-loop memory correction and catastrophic forgetting defense Step S510: Closed-loop correction.

[0066] When the robot returned from its mission, it found that the chair that was originally in the living room had been moved.

[0067] The underlying control network found that the original path was unobstructed.

[0068] The system passes this feedback to the schema module. The layout layer updates the data using Bayesian methods, reducing the confidence that "the chair is fixed in one place" and updating the occupancy state of the physical layer.

[0069] Topological edge weights The corresponding decrease reflects a reduction in toll costs.

[0070] Step S520: Cross-scene switching (Catastrophic amnesia defense).

[0071] When the robot enters the elevator and travels from the 1st floor to the 2nd floor bedroom area: The fundamental entropy undergoes a dramatic change (the metallic texture inside the elevator is vastly different from the wooden furniture in the bedroom).

[0072] The system automatically generated a topology map of the first floor. Compress the archive.

[0073] The system loads a historical schema snapshot from the second floor (if it exists), or initializes a new blank schema.

[0074] Because the basic structure of "bedroom-bathroom-walk-in closet" is similar across different floors, the robot can use transfer learning to build a new topology map on the second floor with only a small amount of exploration. And thus, it will not be interfered with by the memory of the first floor (i.e., catastrophic amnesia is avoided).

[0075] 6. Pseudocode Example The following is the pseudocode for the core logic of node generation and edge weight update: # Pseudocode: Primitive schema-driven topology memory update class SchemaTopoNav: def update_memory(self, observation, odometry): # 1. Construct the current environment schema current_schema = self.vlm_inference(observation) # 2. Calculate the fundamental entropy (compared with the current node) kl_divergence=KL_divergence(current_schema, self.current_node.schema) # 3. Determine if a new node needs to be created. if kl_divergence > self.entropy_threshold: # Create a new node new_node = TopoNode( id=self.node_count, schema=current_schema, pose=self.estimate_pose(odometry) ) # Connecting edges edge = TopoEdge( start=self.current_node, end=new_node ) # 4. Edge weight initialization with fundamental schema enhancement edge.weight = self.predict_cost_by_schema( current_schema, self.current_node.schema ) self.graph.add_node(new_node) self.graph.add_edge(edge) self.current_node = new_node else: # Update observation fusion only for the current node self.current_node.fuse_observation(current_schema) def predict_cost_by_schema(self, schema_A, schema_B): # Predicting travel cost based on fundamental models base_cost = 1.0 if schema_A.type == "corridor" and schema_B.type == "room": # Predicting that it requires going through a gate increases the cost. base_cost += 0.2 if self.detect_dynamic_obstacle(schema_A): base_cost += 0.5 return base_cost Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for enhancing the topological memory of embodied intelligent navigation fundamental models for complex indoor environments, characterized in that: Includes the following steps: S1 Environment Model Construction Steps: Use a pre-trained visual-language model to process the robot's current multimodal observation data, extract high-level semantic features, and construct a dynamic environment model that includes spatial functional attributes, object layout priors, and physical constraints. S2 Topology Node Generation and Management Steps: Based on the semantic saliency and entropy change of the environmental schema, the continuous physical space is dynamically discretized into topology nodes with semantic labels, and a hierarchical topology memory graph is established. The S3 fundamental model-enhanced edge weight update steps are as follows: the environmental fundamental model is used to predict the connectivity and passage cost between topological nodes, and the weights of edges in the topological graph are dynamically updated in combination with real-time sensing data to form a representation of passable regions guided by the fundamental model. S4 Hybrid Planning and Decision-Making Steps: In response to navigation target instructions, a global path search is performed on the hierarchical topology memory graph to obtain a macroscopic path sequence, and combined with local real-time perception information, the underlying motion control instructions are output through a reinforcement learning strategy. S5 Closed-Loop Memory Correction Step: Based on the actual feedback results after the action is executed, the parameter confidence and topology structure of the environmental model are corrected and optimized online.

2. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 1, characterized in that: The environmental fundamental model is represented by a three-layer probabilistic graphical structure, including: Semantic layer: used to represent the functional zoning categories of rooms and their confidence distribution, obtained by classifying images using VLM; Layout layer: used to describe the spatial co-occurrence relationship, relative pose and support relationship between objects, and is constructed by generating a scene graph network; Physical layer: used to characterize ground accessibility, friction coefficient, slope and kinematic constraints of dynamic obstacles, updated through depth information and temporal filtering.

3. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 1, characterized in that: In the topology node generation and management steps, the node creation conditions are determined by the fundamental module entropy value: When the KL divergence between the environmental schema corresponding to the current observation and the environmental schema associated with the current topology node exceeds a preset dynamic threshold. If a significant semantic change occurs in the environment, a new topological node is created; otherwise, the current observation data is simply fused into the fundamental model representation of the current node.

4. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 1, characterized in that: In the edge weight update step of the fundamental schema enhancement, the connecting nodes in the topology graph and Boundary rights The calculation method is as follows: in, It is a semantic passage cost based on environmental schema prediction (including penalties for private areas and narrow passages). It is the average travel time based on historical experience. It is the probability of dynamic obstacles being occupied at the current moment. These are the weighting coefficients.

5. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 1, characterized in that: The hybrid planning and decision-making steps employ a two-layer reinforcement learning architecture: Top-level policy network ( ): Taking the fundamental features of the current topology node and the global target as input, output the index of the next sub-target topology node to be reached; Low-level control network ( ): Taking the current local LiDAR point cloud or depth image and the relative pose of the sub-target as input, output the linear velocity and angular velocity control values ​​of the robot; The update frequency of the top-level policy network is lower than that of the bottom-level control network.

6. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 5, characterized in that: The top-level policy network introduces an intrinsic reward function guided by the schema during the training phase: When the robot explores and establishes new and effective topological connections, it is given a positive reward based on the schema information gain of the new connection, in order to encourage the robot to use environmental schemas for efficient exploration rather than blindly wandering randomly.

7. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 1, characterized in that: It also includes catastrophic forgetting defense mechanisms: The system maintains a long-term memory to store schema snapshots and topology maps of different floors. When a robot crosses a geofence (such as entering an elevator or stairs) and causes a sudden change in the environmental schema, the system automatically archives the topology memory of the current scene and loads the corresponding historical schema snapshot according to the identified new scene type, thus achieving the isolation and rapid reuse of long-term memory.

8. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 2, characterized in that: The layout layer update employs a Bayesian inference framework: When the robot observes an object placement that is inconsistent with the prediction of the basic model layout layer (e.g., a chair is moved), it dynamically adjusts the posterior probability of the corresponding relationship in the layout layer based on the confidence level of the observation; when the posterior probability is lower than a preset threshold, it triggers the correction or splitting of the basic model structure.

9. The method for enhancing the topological memory of embodied intelligent navigation fundamental model for complex indoor environments according to claim 1, characterized in that: The closed-loop memory correction step specifically includes: When performing a global path search, if multiple attempts to access a certain topology node fail, the system marks the node as "unreachable" or "occupied by a dynamic obstacle". Then, it uses the reasoning ability of the environment schema to find an alternative path. After the alternative path is successfully traversed, the connectivity state of the original failed node is updated in reverse to distinguish between "permanent structural changes" and "temporary dynamic obstacles".

10. A personal intelligent navigation system, characterized in that, include: The perception module is configured to acquire images and depth information of the robot's surrounding environment; A basic model construction module, coupled to the perception module, is configured to perform the environmental basic model construction step as described in any one of claims 1-9; A topology memory module, coupled to the base module, is configured to perform the topology node generation, edge weight update and memory correction steps as described in any one of claims 1-9; The decision planning module is coupled to the topology memory module and the perception module respectively, and is configured to execute the hybrid planning and decision steps as described in any one of claims 1-9; The actuator interface is configured to send the instructions generated by the decision planning module to the robot's motor driver.

Images