A large language model-based adaptive interactive navigation method and device

Abstract

This invention relates to an adaptive interactive navigation method and apparatus based on a large language model. The method includes: a robot capturing environmental information through sensors; constructing a primitive skill tree and generating an optimal task sequence using modules based on a large language model based on target indication and environmental information; the robot calling upon an atomic skill library to execute the optimal skill sequence; the robot receiving new observation information during movement and determining whether replanning needs to be triggered through an adaptive replanning mechanism; and the robot sequentially executing control commands based on a task decomposition scheme and making real-time replanning adjustments according to the replanning task decomposition. This application effectively overcomes the problem that single planning schemes are prone to failure due to misunderstandings of the large language model or environmental uncertainties, and can proactively and adaptively respond to environmental changes. By directly modifying the skill tree, it greatly improves the computational efficiency of replanning and achieves real-time rapid response in unknown environments.

Description

Technical Field

[0001] This application belongs to the field of autonomous navigation, specifically, it relates to an adaptive interactive navigation method and device based on a large language model. Background Technology

[0002] In the field of robotics, achieving autonomous navigation of robots in diverse and complex environments is a core and crucial research problem, requiring robots to possess efficient and adaptive decision-making and planning capabilities. In recent years, platforms represented by quadruped robots have demonstrated enormous potential in tasks involving traversing complex terrains due to their superior agility and terrain adaptability. Despite significant progress, traditional navigation strategies are essentially passive path planning. The core idea of ​​this approach is to plan the optimal trajectory within a given, static, free workspace. However, this strategy has serious limitations in many real-world scenarios (such as disaster relief sites, crowded warehouses, etc.) because readily available feasible paths to the destination may not exist beforehand. In such cases, robots must shift from passively adapting to the environment to active interactive navigation, that is, dynamically changing the workspace by actively interacting with objects in the environment (e.g., pushing aside obstacles, using objects as tools, etc.) to create feasible routes for themselves.

[0003] Interactive navigation can generally be viewed as a complex Task and Motion Planning (TAMP) problem, comprising two main technical aspects: high-level task planning and low-level motion planning. However, current technologies face bottlenecks in both areas. First, in high-level task planning—the decision-making process of determining which objects to interact with and the order of interaction—existing methods typically require an omniscient global scene description as input. This is difficult to achieve in the real world because robots usually only obtain local observation information through their onboard sensors, severely limiting their adaptability in dynamic and partially observable environments. Second, in low-level motion planning and execution, while research has demonstrated robots' excellent obstacle-crossing capabilities (such as robot parkour), the core challenge of interactive navigation is not simply adapting to terrain, but actively manipulating and utilizing environmental objects, which places higher demands on the coupling of planning and control. Furthermore, to cope with environmental uncertainties, the system must possess closed-loop feedback and replanning capabilities. However, most existing closed-loop planning methods are passive, focusing on error recovery after plan execution failures, lacking the ability to proactively utilize newly discovered information to optimize the current plan. Therefore, the core challenge facing the field today is how to create effective navigation plans for robots in unknown interactive environments, while also being able to adaptively and proactively respond to the new information they acquire during the task.

[0004] Navigation in Movable Obstacle Environments (NAMO) is an existing approach that allows robots to actively interact with movable obstacles in their environment to create feasible paths. NAMO methods primarily focus on obstacle removal, failing to utilize obstacles as tools—for example, using shorter obstacles as stepping stones to cross taller ones that would otherwise be insurmountable. This limits their ability to solve long-term, interactive navigation tasks requiring multi-step strategic utilization of environmental objects.

[0005] Benefiting from the state-of-the-art capabilities of Large Language Models (LLMs) in understanding and reasoning, recent work combines LLMs with pre-trained atomic skill sets to solve complex, long-term interaction tasks. These problems are typically modeled as Task and Motion Planning (TAMP) problems. Task planning is responsible for decomposing the final goal into a series of subtasks, such as choosing which objects to interact with and determining the order of interactions; while motion planning is responsible for designing specific controllers to execute these high-level instructions. Such methods suffer from two main drawbacks. First, they typically require an omniscient global scene description as input, which is difficult to obtain in the real world because robots usually rely on their own sensors to obtain first-person, local observations. Second, most of these methods perform open-loop planning, meaning they generate a complete plan and then do not change it, which is impractical for navigation tasks that require continuous adjustments to the strategy based on new information during the task.

[0006] To compensate for the shortcomings of open-loop planning, some works have introduced closed-loop planning, responding to new observations through iterative replanning. To improve planning quality, researchers have adopted tree-based structures to enable large language models to reason, selecting the optimal solution by generating and evaluating multiple possible planning paths. Existing closed-loop replanning mechanisms have significant limitations. First, they are usually passive, mostly triggering replanning only when the current plan fails. This mechanism focuses on error recovery while neglecting the possibility of proactively using new information to optimize the current plan; for example, when the robot discovers a new object that can more easily complete the task, the system cannot proactively adjust its strategy. Second, the flexibility of planning is limited. For tree-based planning methods, once the initial planning tree is generated, if the optimal solution is not included in the original tree, or if the robot needs to switch to a completely new strategy based on new observations, the fixed tree structure will hinder it from finding the optimal solution. Summary of the Invention

[0007] This application provides an adaptive interactive navigation method and apparatus based on a large language model, which actively interacts with the environment through task and motion planning to create paths and can adaptively respond to new observations to solve navigation problems in complex environments.

[0008] In a first aspect, embodiments of this application provide an adaptive interactive navigation method based on a large language model, comprising: S1, the robot captures environmental information through sensors; S2, based on the target indication and environmental information, constructs a primitive skill tree using modules based on a large language model and generates an optimal task sequence; S3, the robot calls the atomic skill library and executes the optimal skill sequence; S4, the robot receives new observation information during movement and determines whether replanning needs to be triggered through an adaptive replanning mechanism; S5, based on the task decomposition scheme generated in step S2, the robot executes control commands and performs replanning adjustments in real time according to the replanning task decomposition in step S4.

[0009] In step S1, the robot captures environmental information through its own perspective sensors: it uses a visual language model to understand the scene, uses the YOLOv8 target recognition algorithm to identify objects, and obtains the spatial location and geometric attribute description information of objects in the environment.

[0010] In step S2, the module based on the large language model includes: proposer and evaluator; the proposer generates multiple high-level, object-level potential planning schemes based on user instructions and environmental information. Subsequently, the proposer combines these object-level schemes with the robot's predefined skill library to transform them into multiple more detailed, skill-level planning schemes; each skill-level plan consists of a series of atomic skills.

[0011] In step S2, planning schemes for multiple skill levels are merged to construct a primitive skill tree. Paths with the same historical action sequence in the primitive skill tree are merged into shared nodes to form a tree structure containing multiple alternative strategies.

[0012] In step S2, the evaluator assesses an immediate reward r(n) for each node n in the skill tree; the long-term cumulative reward Q(n) for each node n is calculated using backpropagation based on the immediate reward of each node and the value of its child nodes.

[0013]

[0014] in It is a discount factor. Represents a node The set of child nodes express A certain child node, Representative node The long-term cumulative reward; starting from the root node, iteratively select the child node with the maximum cumulative reward to extract an optimal high-level original skill sequence.

[0015] In step S3, the robot calls a pre-trained atomic skill library to execute the optimal skill sequence extracted in step S2. The skill library includes low-level motor skills and high-level complex skills. Low-level motor skills include walking and climbing skills. High-level complex skills include obstacle avoidance navigation and object interaction.

[0016] In step S4, the adaptive replanning mechanism includes two roles based on a large model: the advisor and the arboreal. The advisor continuously interprets the new observations acquired by the robot and decides whether to trigger replanning. When the advisor decides to replan, it provides the arboreal with an analysis and interpretation of the observations and suggestions for replanning. The arboreal adds or removes nodes from the primitive skill tree constructed in step S2 based on the specific suggestions provided by the advisor. The conditions for triggering replanning include three types of observation interpretations: failure, new objects, and re-evaluation.

[0017] Secondly, embodiments of this application provide an adaptive interactive navigation device based on a large language model, comprising: Acquisition unit, used by the robot to capture environmental information through sensors; The generation unit is used to construct a primitive skill tree and generate the optimal task sequence based on the target indication and environmental information using modules based on a large language model. The execution unit is used by the robot to call upon the atomic skill library and execute the optimal skill sequence; The decision unit is used to receive new observation information during the robot's movement and determine whether replanning needs to be triggered through an adaptive replanning mechanism. The replanning unit is used to enable the robot to sequentially execute control commands based on the task decomposition scheme and to make replanning adjustments in real time according to the replanning task decomposition.

[0018] Thirdly, embodiments of this application provide a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0019] Fourthly, embodiments of this application provide a computer device, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps of any of the methods described above.

[0020] The adaptive interactive navigation method and apparatus based on a large language model according to the embodiments of the present invention have the following beneficial effects: 1. This invention empowers robots to actively interact with their environment, enabling them not only to remove obstacles but also to creatively utilize environmental objects as tools (such as building steps), thereby solving complex tasks that cannot be accomplished by traditional navigation and simple interaction methods.

[0021] 2. The primitive skill tree structure proposed in this invention can simultaneously explore and evaluate multiple alternative planning schemes, effectively overcoming the problem that a single planning scheme is prone to failure due to misunderstandings of large language models or environmental uncertainties, thus improving the robustness and quality of planning. The hierarchical reasoning process from the object level to the skill level also makes the generated plan more reasonable and feasible.

[0022] 3. This invention achieves efficient adaptation in unknown environments. The proposed adaptive replanning mechanism enables the robot to proactively and adaptively respond to environmental changes based on its own observations, rather than being limited to passive failure recovery. By directly modifying the skill tree instead of planning from scratch, the computational efficiency of replanning is greatly improved, enabling real-time and rapid response in unknown environments.

[0023] 4. The framework proposed in this invention makes decisions based on the robot's own perspective, without the need to provide a global map or information in advance, which is more in line with real-world application scenarios. Attached Figure Description

[0024] Figure 1 This is a schematic diagram of the adaptive interactive navigation method based on a large language model according to an embodiment of this application; Figure 2 This is a task planning flowchart for an embodiment of this application; Figure 3 This application provides a comparison of the performance metrics of different methods in interactive navigation tasks according to embodiments of the present application. Figure 4 This is a diagram of the adaptive interactive navigation framework in an embodiment of this application; Figure 5 This is a schematic diagram of the adaptive interactive navigation device based on a large language model according to an embodiment of this application. Detailed Implementation

[0025] The present application will be further described below with reference to the accompanying drawings and embodiments.

[0026] In the following description, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance. The following description provides multiple embodiments of the invention, which can be substituted or combined with each other. Therefore, this application can also be considered to include all possible combinations of the same and / or different embodiments described. Thus, if one embodiment includes features A, B, and C, and another embodiment includes features B and D, then this application should also be considered to include embodiments containing one or more other possible combinations of features A, B, C, and D, even if such embodiments are not explicitly described in the following text.

[0027] Example 1 To address the challenges of traditional navigation methods failing to complete tasks in complex environments without pre-defined paths, existing interactive navigation methods having limited environmental utilization and being unable to solve complex problems requiring the use of obstacles as tools, and the fact that existing planning methods based on large language models mostly rely on unrealistic global environmental information and employ open-loop planning, resulting in poor adaptability in unknown or dynamically changing environments, as well as the fact that existing closed-loop replanning mechanisms are mostly passive, only triggered upon failure, lacking initiative and flexibility, and unable to efficiently utilize new observation information to proactively optimize task plans.

[0028] like Figure 1 As shown, this application provides an adaptive interactive navigation method based on a large language model, including: S1, the robot captures environmental information through sensors; S2, based on the target indication and environmental information, a primitive skill tree is constructed using modules based on a large language model and an optimal task sequence is generated; S3, the robot calls the atomic skill library and executes the optimal skill sequence; S4, the robot receives new observation information during movement and determines whether replanning needs to be triggered through an adaptive replanning mechanism; S5, based on the task decomposition scheme generated in step S2, the robot sequentially executes control commands and performs replanning adjustments in real time according to the replanning task decomposition in step S4.

[0029] This application proposes a primitive skill tree based on a large language model, decomposing tasks into a tree structure with atomic skills. It achieves efficient and robust reasoning by exploring multiple potential solutions and rapidly adapts to new environmental information by modifying the tree structure. Simultaneously, a reinforcement learning-based atomic skill library is designed. By combining it with customized reward design and task curriculum, the robot can learn to navigate and interact in complex environments. Furthermore, an adaptive replanning mechanism is proposed, which not only adaptively triggers the replanning module based on new observations but also automatically adjusts the skill tree based on observation analysis to achieve optimal action execution, enabling real-time rapid response in unknown environments.

[0030] Example 2 This application provides an adaptive interactive navigation method based on a large language model, the specific steps of which are as follows: Step 1: The robot captures environmental information through its own perspective sensors (such as cameras and LiDAR). It uses a Visual Language Model (VLM) to understand the scene and uses the YOLOv8 object recognition algorithm to identify objects, obtaining descriptive information such as the spatial location and geometric attributes of objects in the environment.

[0031] Step 2 (Task Planning): For example... Figure 2As shown, based on the target indication and environmental information, two modules based on a large language model, namely the proposer and the evaluator, are used to construct a primitive skill tree and generate the optimal task sequence.

[0032] Step 2.1 (Proposer): Based on user instructions (such as navigation goals) and environmental information, the proposer first generates multiple high-level, object-level potential planning schemes. Each scheme specifies which objects need to be interacted with and how to utilize them to achieve the navigation goal. Subsequently, the proposer combines these object-level schemes with the robot's predefined skill library to transform them into multiple more detailed, skill-level planning schemes. Each skill-level plan consists of a series of atomic skills (such as walking, climbing, and interacting with objects).

[0033] Step 2.2 (Skill Tree Construction): Merge planning schemes from multiple skill levels to construct a primitive skill tree. In this tree, paths with the same historical action sequence are merged into shared nodes, thus forming a tree structure containing multiple alternative strategies. This structure is more robust than a single linear programming approach.

[0034] Step 2.3 (Evaluator): The evaluator evaluates an immediate reward r(n) for each node n (i.e. each atomic skill) in the skill tree. This reward takes into account both the contribution of the atomic skill to completing the final task and the difficulty of performing the skill.

[0035] Step 2.4 (Optimal Task Sequence Selection): The backpropagation method is used to calculate the long-term cumulative reward Q(n) of each node n based on the immediate reward of each node and the value of its child nodes.

[0036]

[0037] in It is a discount factor. Represents a node The set of child nodes express A certain child node, Representative node The long-term cumulative reward is then calculated. Finally, starting from the root node, the child node with the maximum cumulative reward is iteratively selected to extract an optimal high-level original skill sequence from the tree, which serves as the task decomposition scheme to be executed.

[0038] Step 3 (Motion Planning): The robot invokes a pre-trained atomic skill library to execute the optimal skill sequence output in Step 2. This skill library, trained through reinforcement learning, contains low-level motor skills and high-level complex skills. Low-level motor skills include walking and climbing, enabling strategies for walking on complex terrain based on speed commands and climbing obstacles of varying heights. High-level complex skills include obstacle avoidance and navigation, where obstacle avoidance navigates to a target point while avoiding obstacles, and object interaction involves actively pushing objects to designated positions. Training employs curriculum learning and customized reward functions to enhance the robustness and naturalness of the skills. Execution of the skill library translates the task planning scheme into specific control commands for the robot.

[0039] Step 4 (Adaptive Replanning): During its movement, the robot receives new observation information through its own sensors. This invention proposes an adaptive replanning mechanism, which includes two roles based on a large model: Advisor and Arborist, to adaptively determine when and how to replan.

[0040] Step 4.1 (Proposer): This module continuously interprets new observations acquired by the robot and determines whether replanning needs to be triggered. The conditions for triggering replanning include three types of observation interpretations: failure, new object, and reevaluation. Failure refers to a situation where the current skill execution fails; a new object means the robot has discovered a new, potentially useful interactive object that could generate a better path; and reevaluation refers to a new understanding of the properties of a known object (e.g., after attempting to push a box, it is found to be too heavy to move), which necessitates a re-evaluation of the original task planning strategy.

[0041] Step 4.2 (Arboretum): If the advisor decides to replan, they will provide the arboretum with an analysis and interpretation of the observations and suggestions for replanning. The arboretum will then add or remove nodes from the primitive skill tree constructed in Step 2 based on the specific suggestions provided by the advisor. This includes adding new branch nodes to the tree for newly discovered objects or new strategies, or removing infeasible paths through node pruning.

[0042] Step 5: After the task decomposition plan is generated in step 2, the robot executes control commands sequentially and adjusts the task decomposition in real time according to the replanning module in step 4 until the navigation task is completed.

[0043] like Figure 3As shown, this invention will demonstrate the simulation and real-world verification results of the adaptive interactive navigation scheme in navigation tasks across various scenarios. First, comparative experiments were conducted in multiple simulation scenarios. We selected sampling-based methods (Art-planner), hierarchical reinforcement learning methods (Hierarchical RL), and two large-model-based methods (RoboTool and Tree-planner) as comparison methods. Comparative experiments were carried out in low-difficulty, medium-difficulty, and high-difficulty scenarios involving box obstruction, box usage, and stair building, as well as in Matterport dataset scenarios, urban ruins, and field ruins. Comparison metrics included success rate (SR), total completion time (OT), completion time on success (OTS), planning time (PT), and execution time (ET). Specific data are shown in Table 1.

[0044] Table 1: Comparison of results from different navigation methods

[0045] As can be seen, the adaptive interactive navigation method proposed in this project has a higher success rate and lower task completion time compared to other methods in various scenarios. This is attributed to the improved task planning quality achieved by the primitive skill tree design of this invention, while the adaptive replanning mechanism flexibly adapts to new environmental information. Furthermore, this invention can efficiently plan and respond in real time in unknown environments without relying on prior global information, thus achieving a high success rate in various indoor and outdoor scenarios. Although node evaluation slightly increases planning time, the high-quality planning and flexible replanning mechanism result in a higher overall task success rate and shorter completion time.

[0046] The first key technical point of this application is a task planning method based on primitive skill trees, which uses a large language model to generate multiple planning schemes in two stages (from object level to skill level), merges these schemes into a skill tree containing alternative paths, and then selects the optimal task sequence through evaluation and search.

[0047] Key technical point two: An adaptive replanning method, which designs a consultant module that can proactively determine the timing of replanning based on three different interpretations of new observations (failure, new object, re-evaluation); and sets up a tree arborist module that modifies the skill tree structure (adding or pruning nodes) based on the consultant's suggestions to achieve efficient plan adjustment.

[0048] The third key technical point is the closed-loop navigation system that combines skill tree planning with adaptive replanning. The key point is that the structure of the skill tree provides a foundation for efficient replanning, allowing adjustments to the plan to be made by locally modifying the tree structure, thereby enabling a rapid and flexible response to unknown environmental information.

[0049] Example 3 like Figure 4 As shown, this application proposes an adaptive interactive navigation method based on a large language model. This method is a closed-loop hierarchical planning system that actively interacts with the environment to create paths through task and motion planning, and can adaptively respond to new observations to solve navigation problems in complex environments. The method includes: First, a primitive skill tree based on a large language model is proposed, which decomposes the task into a tree structure with atomic skills. Efficient and robust reasoning is achieved by exploring multiple potential solutions, and the tree structure is modified to quickly adapt to new environmental information. Second, an atomic skill library based on reinforcement learning is designed. By combining it with customized reward design and task curriculum, the robot can learn to navigate and interact in complex environments. Third, an adaptive replanning mechanism is proposed, which not only adaptively triggers the replanning module based on new observations, but also automatically adjusts the skill tree based on observation analysis to achieve optimal action execution.

[0050] This application addresses complex tasks that traditional navigation and simple interaction methods cannot accomplish. The primitive skill tree structure can simultaneously explore and evaluate multiple alternative planning schemes, effectively overcoming the problem that a single planning scheme is prone to failure due to misunderstandings by large language models or environmental uncertainties. The proposed adaptive replanning mechanism enables the robot to proactively and adaptively respond to environmental changes based on its own observations, making the generated plans more reasonable and feasible. By directly modifying the skill tree instead of planning from scratch, the computational efficiency of replanning is greatly improved, enabling real-time and rapid response in unknown environments.

[0051] Example 4 like Figure 5 As shown, this application embodiment provides an adaptive interactive navigation device based on a large language model, including: an acquisition unit 201, used by the robot to capture environmental information through sensors; a generation unit 202, used to construct a primitive skill tree and generate an optimal task sequence based on the target indication and environmental information using modules based on the large language model; an execution unit 203, used by the robot to call the atomic skill library and execute the optimal skill sequence; a judgment unit 204, used by the robot to receive new observation information during movement and determine whether replanning needs to be triggered through an adaptive replanning mechanism; and a replanning unit 205, used by the robot to sequentially execute control commands based on a task decomposition scheme and to perform replanning adjustments in real time according to the replanning task decomposition.

[0052] In this application, the embodiments of the adaptive interactive navigation device based on the large language model are basically similar to the embodiments of the adaptive interactive navigation method based on the large language model. For relevant details, please refer to the description of the embodiments of the adaptive interactive navigation method based on the large language model.

[0053] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the aforementioned adaptive interactive navigation method based on a large language model. The computer-readable storage medium may include, but is not limited to, any type of disk, including floppy disks, optical disks, DVDs, CD-ROMs, microdrives, as well as magneto-optical disks, ROMs, RAMs, EPROMs, EEPROMs, DRAMs, VRAMs, flash memory devices, magnetic cards or optical cards, nanosystems (including molecular memory ICs), or any type of medium or device suitable for storing instructions and / or data.

[0054] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A large language model-based adaptive interactive navigation method, characterized in that, include: S1, the robot captures environmental information through sensors; S2, based on target indications and environmental information, utilizes modules based on a large language model to construct a primitive skill tree and generate the optimal task sequence; S3, the robot calls upon the atomic skill library and executes the optimal skill sequence; S4, the robot receives new observation information during its movement and determines whether to trigger replanning through an adaptive replanning mechanism; S5. Based on the task decomposition scheme generated in step S2, the robot executes control commands and performs real-time replanning adjustments according to the replanning task decomposition in step S4.

2. The adaptive interactive navigation method based on a large language model according to claim 1, characterized in that, In step S1, the robot captures environmental information through its own perspective sensors: it uses a visual language model to understand the scene, uses the YOLOv8 target recognition algorithm to identify objects, and obtains information on the spatial location and geometric attributes of objects in the environment.

3. The adaptive interactive navigation method based on a large language model according to claim 1 or 2, characterized in that, In step S2, the modules of the large language model include: proposer and evaluator; the proposer generates multiple high-level, object-level potential planning schemes based on user instructions and environmental information. Subsequently, the proposer combines these object-level schemes with the robot's predefined skill library to transform them into multiple more detailed, skill-level planning schemes; each skill-level plan consists of a series of atomic skills.

4. The adaptive interactive navigation method based on a large language model according to claim 1 or 2, characterized in that, In step S2, planning schemes for multiple skill levels are merged to construct a primitive skill tree. Paths with the same historical action sequence in the primitive skill tree are merged into shared nodes, forming a tree structure containing multiple alternative strategies.

5. The adaptive interactive navigation method based on a large language model according to claim 1 or 2, characterized in that, In step S2, the evaluator evaluates an immediate reward r(n) for each node n in the skill tree; using backpropagation, the long-term cumulative reward Q(n) for each node n is calculated based on the immediate reward of each node and the value of its child nodes. in It is a discount factor. Represents a node The set of child nodes express A certain child node, Representative node The long-term cumulative reward; starting from the root node, iteratively select the child node with the maximum cumulative reward to extract an optimal high-level original skill sequence.

6. The adaptive interactive navigation method based on a large language model according to claim 1 or 2, characterized in that, In step S3, the robot calls the pre-trained atomic skill library to execute the optimal skill sequence extracted in step S2; The skill set includes basic motor skills and high-level complex skills; Basic motor skills include walking and climbing. Advanced and complex skills include obstacle avoidance navigation and object interaction.

7. The adaptive interactive navigation method based on a large language model according to claim 1 or 2, characterized in that, In step S4, the adaptive replanning mechanism includes two roles based on a large model: the advisor and the arboreal. The advisor continuously interprets the new observations acquired by the robot and decides whether to trigger replanning. When the advisor decides to replan, it provides the arboreal with an analysis and interpretation of the observations and suggestions for replanning. The arboreal adds or removes nodes from the primitive skill tree constructed in step S2 based on the specific suggestions provided by the advisor. The conditions for triggering replanning include three types of observation interpretations: failure, new objects, and re-evaluation.

8. An adaptive interactive navigation device based on a large language model, characterized in that, include: The acquisition unit utilizes the robot to capture environmental information through sensors; The generation unit is used to construct a primitive skill tree and generate the optimal task sequence based on the target indication and environmental information using modules based on a large language model. The execution unit utilizes the robot to access the atomic skill library and execute the optimal skill sequence. The decision unit is used to receive new observation information during the robot's movement and determine whether replanning needs to be triggered through an adaptive replanning mechanism. The replanning unit is used to execute control commands for the robot based on the task decomposition scheme and to make replanning adjustments in real time according to the replanning task decomposition.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method described in any one of claims 1-7.

10. A computer device, comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method according to any one of claims 1-7.

Images