Dynamic article long-term repositioning method and system oriented to man-machine coexistence scene

Abstract

The invention provides a dynamic article long-term relocation method and system for a man-machine coexistence scene, and relates to the technical field of robot task planning, environmental perception and artificial intelligence, and the system comprises a probability model construction module, a dynamic article state discrimination module, a task article reasoning module, a hierarchical search planning module and a probability model updating module. And a closed-loop processing flow is formed among the modules. Specifically, the probability model construction module provides basic model support for state discrimination and task reasoning; the output of the task item reasoning module drives the execution of the hierarchical search planning module; and a sensing result in the planning execution process is fed back to the probability model updating module to form a continuously optimized closed-loop system. The system can autonomously perceive the environment, infer the potential position and generate an efficient search sequence under the scene that the position of an article changes frequently, and stable and long-term article relocation service is achieved.

Description

Technical Field

[0001] This invention relates to the fields of robot task planning, environmental perception and artificial intelligence technology, and specifically to a method and system for service robots to relocate task-related items in a long-term, dynamic manner in a human-robot coexistence environment. Background Technology

[0002] With advancements in robot perception, mapping, and navigation technologies, mobile robots are increasingly being used in human-robot coexistence environments such as homes and hospitals. A prerequisite for robots to perform various tasks is the ability to accurately and efficiently locate task-related objects. However, during long-term operation, the positions of objects often dynamically change due to human activities, making traditional one-time positioning methods unsuitable for such scenarios.

[0003] Existing methods primarily rely on semantic information guidance or learning strategies for item search. Methods guided by semantic information utilize spatial relationships between items and rooms to construct probability distributions or semantic knowledge bases; however, they depend on static common sense and struggle to handle scenarios with frequently changing item locations. Learning strategy-based methods make decisions through reinforcement learning, historical experience, or large language models (LLM), but suffer from issues such as strong correlation with historical data, mismatch between training data and the real environment, and high decision-making costs. Furthermore, existing methods often neglect the dynamic discrimination of item states and the modeling of long-term spatiotemporal patterns, resulting in low search efficiency and insufficient robustness for robots in environments with frequent human activity.

[0004] Therefore, existing dynamic relocation technologies suffer from several shortcomings when dealing with long-term relocation of dynamic objects in human-robot coexistence scenarios. These include strong reliance on static common sense, insufficient updates to the dynamic object's state, and a lack of long-term spatiotemporal pattern modeling. There is an urgent need to develop a robot system and method that can integrate spatiotemporal dynamic features, adaptively update object states, and efficiently relocate dynamic objects in long-term tasks. This would improve the robot's search efficiency and robustness in dynamic environments, meeting the practical needs of human-robot coexistence scenarios. Summary of the Invention

[0005] In view of this, the present invention provides a method and system for long-term dynamic object relocation in human-machine coexistence scenarios. The system can autonomously perceive the environment, infer potential locations, and generate efficient search sequences in scenarios where the position of an object changes frequently, thereby achieving stable and long-term object relocation services.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: On the one hand, this invention provides a method for long-term relocation of dynamic objects in human-machine coexistence scenarios, including: Based on the object perception data accumulated during the long-term operation of the robot, spatiotemporal features including the object's location and timestamp are extracted, and the spatiotemporal features are encoded. Based on the encoded spatiotemporal features, a Gaussian mixture model is used to model the distribution of item locations, construct a spatiotemporal probability model that reflects the long-term occurrence pattern of items, and build an item knowledge base that includes collinear relationships of items. Based on the probability distribution results output by the spatiotemporal probability model, the spatial encounter frequency and temporal observation frequency of the items are calculated. Combined with a preset perception interaction ratio threshold, the motion state of the items is dynamically classified, and the priority of the items in the relocation task is adjusted based on the classification. The system receives user query requests in natural language form, combines the priority of items in the relocation task, and generates a candidate item list using a pre-trained large language model and the item knowledge base; it selects the optimal task item from the candidate item list and determines its potential search location sequence. Based on the search location sequence, a hierarchical planning algorithm is used to generate a search path. The global planning layer divides the search space into task sub-environments based on the physical boundaries of the environment and generates a search sequence for each sub-environment by combining the spatiotemporal probability model. In each sub-environment, the local planning layer guides the robot to adjust its observation perspective by utilizing the co-occurrence relationship between items to achieve fine-grained search and backtrack or expand the search for cases where no target is found.

[0007] Furthermore, it also includes: dynamically updating the mean, covariance, and mixture weights of the Gaussian mixture model in the spatiotemporal neighborhood based on the latest perceived object location and time during the robot's search task, and introducing the influence of historical data to adaptively adjust the temporal distance and perception interaction ratio.

[0008] Furthermore, the mixed weight update formula is as follows: ; ; ; in, It is the mixture weight of the original Gaussian mixture model. These are the mixture weights of the updated Gaussian mixture model. These are normalized Gaussian mixture model weights. It is an incremental update. It is a scaling factor, which is inversely proportional to the distance between the current Gaussian component time center and the perception time, and is affected by the perceptual interaction rate and correlation coefficient. The impact.

[0009] Furthermore, the spatiotemporal features are encoded, including: Morton coding is used to fuse the two-dimensional position coordinates perceived by the robot with the timestamp into a unified one-dimensional spatiotemporal index.

[0010] Furthermore, the spatiotemporal probability model also explicitly models the joint distribution of location and time: ; in, It is the joint distribution density. It is component weight. It is a Gaussian distribution. It is the mean; Its mean and covariance matrices are defined as follows: ; ; in, It is the positional mean. It is the time average; It is the location variance. It is the time variance; As a spatiotemporal correlation coefficient, it effectively distinguishes between static items and weakly correlated dynamic items. Strongly related dynamic items Distribution characteristics.

[0011] Furthermore, the encounter frequency is ;in, The frequency of encounters at different locations. It is the frequency of encounters in time; The perceived interaction ratio is ;in, It is a hyperbolic function. It is a proportionality coefficient. It refers to the number of interactions. It refers to the number of times the sensory information is perceived.

[0012] Further, selecting the optimal task item from the candidate item list includes: The optimal task item is selected using a dynamic selection strategy based on cost-expectation balance; this dynamic selection strategy comprehensively considers search expectations. and path cost Construct a selection threshold: ; in, It is about selecting a threshold. It is the balance coefficient. It is a candidate item. It is an expectation. It's a probability.

[0013] Furthermore, the optimization objective function of the global planning layer is: ; in, It is the objective function. It is a cost function. It is the probability density. It is a logarithmic function. It is the summation symbol; The local planning layer utilizes the co-occurrence relationships of items in the item knowledge base. Guide the robot to gradually adjust its perspective to the target item. .

[0014] Furthermore, this invention also provides a dynamic long-term relocation system for objects in human-machine coexistence scenarios, comprising: The probability model construction module extracts spatiotemporal features containing the location and timestamp of the items based on the object perception data accumulated during the long-term operation of the robot, and encodes the spatiotemporal features; based on the encoded spatiotemporal features, it uses a Gaussian mixture model to model the distribution of the item locations, constructs a spatiotemporal probability model that reflects the long-term occurrence pattern of the items, and constructs an item knowledge base containing the collinear relationship of the items. The dynamic item state discrimination module calculates the spatial encounter frequency and temporal observation frequency of items based on the probability distribution results output by the spatiotemporal probability model, and dynamically classifies the motion state of items in combination with a preset perception interaction ratio threshold, and adjusts the priority of items in the relocation task based on the classification. The task item reasoning module receives user query requests in natural language form, combines the priority of items in the relocation task, and generates a candidate item list using a pre-trained large language model and the item knowledge base; it selects the optimal task item from the candidate item list and determines its potential search location sequence. The hierarchical search planning module generates a search path based on the search position sequence using a hierarchical planning algorithm. The global planning layer divides the search space into task sub-environments based on the physical boundaries of the environment and generates a search sequence for each sub-environment by combining the spatiotemporal probability model. The local planning layer guides the robot to adjust its observation perspective within each sub-environment by utilizing the co-occurrence relationship between items to achieve fine-grained search and backtracks or expands the search for cases where no target is found.

[0015] Furthermore, it also includes: The probability model update module dynamically updates the mean, covariance, and mixture weights of the Gaussian mixture model in the spatiotemporal neighborhood based on the latest perceived object location and time during the robot's search task, and adaptively adjusts the influence of historical data by incorporating temporal distance and perceptual interaction ratio.

[0016] The technological advancements achieved by this invention due to the adoption of the above technical solutions are as follows: This invention constructs a spatiotemporal probabilistic model based on a Gaussian mixture model, enabling service robots to accurately characterize the distribution patterns of dynamic objects over long-term operation, overcoming the limitations of traditional semantic methods that rely on static common sense. The proposed dual discrimination mechanism of encounter frequency and perceptual interaction ratio effectively distinguishes objects with different dynamic characteristics, enhancing the system's robustness to occasional human activities. The cost-expectation balance-based task item selection strategy and hierarchical search planning mechanism significantly improve the robot's search efficiency in complex human-robot coexistence environments. In particular, the spatiotemporal neighborhood probabilistic model update method enables continuous adaptation to dynamic environments, solving the model drift problem in long-term deployment. Experimental results show that the system outperforms existing benchmark methods in both search success rate and efficiency in real-world environments, providing reliable technical support for the long-term autonomous operation of service robots in dynamic environments. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a system diagram of the robot dynamic object long-term repositioning system in an embodiment of the present invention; Figure 2 This is a flowchart illustrating the long-term repositioning system for dynamic objects using a robot, as described in an embodiment of the present invention. Figure 3 This is a schematic diagram illustrating the characteristic distribution of the motion state of an item in an embodiment of the present invention; Figure 4 This is a schematic diagram of hierarchical search planning in an embodiment of the present invention; Figure 5 This is a schematic diagram of probability model updating in an embodiment of the present invention; Figure 6 This is a visualization of the robot's search trajectory at different starting points and time periods in an embodiment of the present invention; Figure 7 This is a quantitative result of the robot's search trajectory length and search time at different seven points and at different times in an embodiment of the present invention; Figure 8 These are the task success rates of the robot in this embodiment of the invention over one week and one month; Figure 9 This is a schematic diagram of the scene layout for the generalization experiment in an embodiment of the present invention; Figure 10 This represents the task success rate results of the generalization experiment in this embodiment of the invention over 10 days. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0020] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0021] like Figure 1 As shown, a long-term dynamic object relocation system for human-machine coexistence scenarios includes a probabilistic model construction module, a dynamic object state discrimination module, a task object reasoning module, a hierarchical search planning module, and a probabilistic model update module. These modules form a closed-loop processing flow. Specifically, the probabilistic model construction module provides basic model support for state discrimination and task reasoning; the output of the task object reasoning module drives the execution of the hierarchical search planning module; and the perception results during the planning execution process are fed back to the probabilistic model update module, forming a continuously optimized closed-loop system. Wherein: The probability model construction module extracts spatiotemporal features containing the location and timestamp of the items based on the object perception data accumulated during the long-term operation of the robot, and encodes the spatiotemporal features. Based on the encoded spatiotemporal features, a Gaussian mixture model is used to model the distribution of the item locations, construct a spatiotemporal probability model that reflects the long-term occurrence pattern of the items, and construct an item knowledge base containing the collinear relationship of the items.

[0022] The dynamic item state discrimination module calculates the spatial encounter frequency and temporal observation frequency of items based on the probability distribution results output by the spatiotemporal probability model. Combined with a preset perception-interaction ratio threshold, it dynamically classifies the item's motion state and adjusts the item's priority in the relocation task based on this classification. Resources are allocated and relocation costs are calculated according to the item's motion state classification (static, weakly correlated dynamic, or strongly spatiotemporally correlated dynamic). Static items are typically not task objects and can have their ranking in the candidate list reduced. For dynamic items, the expected success rate and expected relocation cost can be calculated based on their perception rate and interaction rate. Furthermore, the candidate items are ranked to determine their priority.

[0023] The task item reasoning module receives user query requests in natural language form, combines the priority of items in the relocation task, and generates a candidate item list using a pre-trained Large Language Model (LLM, a large-scale natural language processing model based on deep learning that can understand and generate human language). It then selects the optimal task item from the candidate item list and determines its potential search location sequence.

[0024] The hierarchical search planning module generates a search path based on the search position sequence using a hierarchical planning algorithm. The global planning layer divides the search space into task sub-environments based on the physical boundaries of the environment and generates a search sequence for each sub-environment by combining the spatiotemporal probability model. The local planning layer guides the robot to adjust its observation perspective within each sub-environment by utilizing the co-occurrence relationship between items to achieve fine-grained search and backtracks or expands the search for cases where no target is found.

[0025] The probability model update module dynamically updates the mean, covariance, and mixture weights of the Gaussian mixture model in the spatiotemporal neighborhood based on the latest perceived object location and time during the robot's search task, and adaptively adjusts the influence of historical data by incorporating temporal distance and perceptual interaction ratio.

[0026] like Figure 2 As shown, the working process of the above system is as follows: The robot first constructs an item knowledge base containing item co-occurrence relationships and a spatiotemporal probability model based on a Gaussian mixture model through long-term observation. When a user request is received, the system uses LLM to parse the requirement and generate a candidate item list. By comprehensively considering the expected benefits provided by the probability model and the search cost derived from path planning, the final task items and their potential location distribution are determined. Based on this, the system implements a hierarchical search strategy consisting of global planning and local decision-making, making full use of the environmental physical boundaries and the spatiotemporal patterns of item changes to adaptively infer the most likely item locations and determine the priority relocation sequence.

[0027] The robot dynamic object long-term relocation system of this invention is based on the Fetch mobile robot, which is equipped with a SICK TIM571 LiDAR and a PrimeSense Carmine 1.09 short-range RGB-D sensor. The computing platform uses a laptop equipped with an NVIDIA RTX4060, running Mask R-CNN for object recognition. The long-term experimental environment is a real human-robot coexistence environment of approximately 50 square meters, including four functional areas: a living room, study, dining room, and bedroom. Figure 7 The environmental layout shown.

[0028] Specifically, a long-term dynamic object relocation method for human-machine coexistence scenarios includes the following steps: Step 1: Spatiotemporal feature extraction and encoding.

[0029] The robot collects environmental information using LiDAR and RGB-D sensors, detecting task-related items such as books, milk, and cups, and extracting the item's location coordinates and perception timestamps. Morton coding is used to represent the two-dimensional location (…). ) and timestamp The samples are merged into a one-dimensional spatiotemporal index to construct a sample set for probabilistic modeling. This encoding process is achieved through bit-interleaving operations, which preserves the spatial topology and enables efficient fusion of spatiotemporal information, providing a compact and spatiotemporally consistent feature representation for subsequent probabilistic modeling.

[0030] Step 2: Constructing the probability model.

[0031] The object location distribution is modeled using a Gaussian mixture model, where each Gaussian component includes the location mean, time mean, covariance matrix, and mixture weights. The number of Gaussian components is set. Initialize the mean vector of each component. Covariance matrix and mixed weights .like Figure 3 The diagram illustrating the spatiotemporal distribution characteristics of static and dynamic objects shows, from left to right, the distribution characteristics of different types of objects in a spatiotemporal coordinate system: (The standard Gaussian distribution is used as a reference.) and Static items have a concentrated spatial distribution ( The characteristics of weak spatiotemporal correlation dynamic items are ( The distribution characteristics of ); dynamically related items with strong spatiotemporal correlation exhibit an inclined distribution pattern. This intuitively reflects the correlation between location and time. Figure 3 The spatiotemporal distribution characteristics of various items shown provide prior guidance for model parameter initialization.

[0032] Step 3: Dynamic item status determination.

[0033] By statistically analyzing the frequency of encounters between objects in space Frequency of observation over time Combined with the perceived interaction ratio The system determines whether an item is static, weakly correlated dynamic, or strongly spatiotemporally correlated dynamic, enabling fine-grained classification of the item's motion state. This discrimination mechanism allows the system to identify sporadic events such as "a robot encountering an item held by a user" and provides a basis for adjusting subsequent search priorities.

[0034] Step 4: Matching quest items.

[0035] When a user requests "I'm thirsty," the system uses an LLM (Limited Module Management) to parse the request and retrieves a list of candidate items (cups, milk) from the item knowledge base. By calculating the search expectation and path cost of each candidate item, a cost-expectation balance strategy is applied. Select the optimal task object, where .

[0036] Step 5: Hierarchical search planning.

[0037] like Figure 4 As shown, the system first divides the search space into multiple task sub-environments based on the physical boundaries of the environment, and then uses the probability distribution of items in each sub-environment as a basis for further analysis. and connectivity constraints Generate a global search sequence Ultimately, this guides detailed search planning within each region based on the probability distribution of specific locations. At the local decision-making level, it utilizes the co-occurrence relationships between items. Guide the robot to perform fine-grained searches within each sub-environment.

[0038] The specific implementation process of global planning is shown in Table 1.

[0039] Table 1

[0040] Step 6: Update the probability model.

[0041] When the robot is in position ,time When a target object is detected, the system [does something] in the spatiotemporal neighborhood. Internally update the parameters of the Gaussian mixture model. For example... Figure 5 As shown, considering the time interval between two perceptions and the short-term unpredictability of human activities, the model update occurs at the perception moment. This is performed within the neighborhood of the target area, by introducing a time distance weighting factor. and perceived interaction ratio The model adaptively adjusts the mixing weights of each Gaussian component to ensure real-time response to dynamic environmental changes.

[0042] Step 7: Long-term repositioning implementation and evaluation.

[0043] The robot performs search tasks, records search trajectories and times, and updates the knowledge base and probability model. For example... Figure 6 The comparison of relocation trajectories under four different strategies shown demonstrates the search path of the relocation method of this invention under different starting points and time periods. Figure 7 The trajectory length and search time statistics further quantify the performance advantages of this method.

[0044] The long-term performance of the system is evaluated using metrics such as success rate (SR), average trajectory length (ATL), and average search time (AST). Figure 8 and Figure 9 The long-term performance evaluation results shown verify the effectiveness and robustness of the system in a real-world environment.

[0045] To verify the effectiveness of this invention, the inventors conducted long-term experiments in a real human-robot coexistence environment. Experimental results show that this invention performs excellently in dynamic object relocation tasks.

[0046] like Figure 6 The comparison of relocation trajectories under the four different strategies clearly shows that the search path of our invention is closest to that of the human participant (HS) and scene graph (SG) methods, and significantly outperforms benchmark methods based on uniform search (US), landmark (LS), greedy algorithm (GS), and impression (IS). Especially in nighttime scenes, our system effectively utilizes spatiotemporal patterns, avoiding ineffective searches in low-probability areas.

[0047] Figure 7 Quantitative statistical results further confirm the advantages of this method. Data from the starting point SP1 show that the average trajectory length (ATL) of the system of this invention is only 10.92% longer than that of HS during the midday period, but significantly better than other benchmark methods; during the nighttime period, the AST of the system of this invention is 22.77% longer than that of HS, but 4.02% shorter than that of the SG method, demonstrating good time efficiency.

[0048] The long-term performance evaluation results are even more encouraging. For example... Figure 8 The trend of search success rate over the past month shows that although the success rate dropped briefly in the second week due to the introduction of an unexpected event (the robot encountering a user holding milk), the system recovered quickly through a dynamic state discrimination mechanism, and the overall success rate maintained an upward trend, demonstrating its strong robustness to unexpected events.

[0049] Generalization ability test, such as Figure 9 and 10 As shown. Figure 9 The scenario layout of the generalization experiment is shown. Figure 10 The success rate of the task over 10 days in the generalization experiment scenario is demonstrated. The system of this invention maintained stable search performance in four different environmental scenarios. Particularly in the structurally complex scenarios 3 and 4, the system exhibited a more stable success rate due to the increased spatial constraints provided by the environmental structure. This verifies the good adaptability of the search strategy based on the spatiotemporal probability model to different environmental structures.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A dynamic long-term relocation method for coexistence scenarios, characterized in that, The method comprises the following steps: Based on the accumulated item perception data in the long-term operation of the robot, the spatiotemporal features containing the item position and the timestamp are extracted, and the spatiotemporal features are encoded; Based on the encoded spatiotemporal features, the Gaussian mixture model is used to model the item position distribution, the spatiotemporal probability model reflecting the long-term appearance regularity of the item is constructed, and the item knowledge base containing the co-occurrence relationship of the items is constructed; According to the probability distribution result output by the spatiotemporal probability model, the spatial encounter frequency and the time observation frequency of the item are calculated, the preset perception interaction ratio threshold is combined, the motion state of the item is dynamically classified, and the priority of the item in the relocation task is adjusted based on the classification; A user query request in a natural language form is received, the priority of the item in the relocation task is combined, a pre-trained large language model and the item knowledge base are used to generate a candidate item list; the optimal task item is selected from the candidate item list, and a potential search position sequence is determined; Based on the search position sequence, a hierarchical planning algorithm is used to generate a search path, wherein a global planning layer divides the search space into task sub-environments based on the physical boundaries of the environment, combines the spatiotemporal probability model to generate a search sequence of the sub-environment, and a local planning layer uses the co-occurrence relationship between the items to guide the robot to adjust the observation angle in each sub-environment, realizes fine-grained search, and performs backtracking or expanded search in the case that the target is not found.

2. The dynamic long-term relocation method for coexistence of man and machine according to claim 1, characterized in that, Further comprising: According to the latest perceived item position and time during the search task process of the robot, the mean, covariance and mixing weight of the Gaussian mixture model are dynamically updated in the spatiotemporal neighborhood, and the influence of the historical data is adaptively adjusted by introducing the time distance and the perception interaction ratio.

3. The dynamic long-term relocation method for coexistence scenarios between humans and machines according to claim 2, characterized in that, The mixing weight update formula is: ； ； ； wherein, is the mixing weight of the original Gaussian mixture model, is the mixing weight of the updated Gaussian mixture model, is the normalized Gaussian mixture model weight, is the update increment, is a scaling factor inversely proportional to the distance between the time center of the current Gaussian component and the perception time, influenced by the perception interaction rate and the correlation coefficient .

4. The dynamic long-term relocation method for coexistence of human and machine according to claim 1, wherein, The encoding of the spatiotemporal features comprises: The two-dimensional position coordinates and the timestamp perceived by the robot are fused into a unified one-dimensional spatiotemporal index by using Morton encoding.

5. The dynamic long-term relocation method for coexistence of human and machine according to claim 1, wherein, The spatiotemporal probability model also explicitly models the joint distribution of position and time: ； wherein, is a joint distribution density, is a component weight, is a Gaussian distribution, is a mean; The mean and the covariance matrix are defined as: ； ； wherein, is the position mean, is the time mean; is the position variance, is the time variance; as the spatio-temporal correlation coefficient, effectively distinguishes the distribution characteristics of static items, weakly correlated dynamic items and strongly correlated dynamic items .

6. The dynamic long-term relocation method for coexistence of human and machine according to claim 1, wherein, The encounter frequency is ; wherein, is the encounter frequency of positions, is the time encounter frequency; The perceived interaction ratio is ; wherein, is a hyperbolic function, is a proportional coefficient, is the number of interactions, is the number of perceptions.

7. The dynamic long-term relocation method for coexistence of human and machine according to claim 1, wherein, Selecting the optimal task item from the candidate item list comprises: An optimal task item is selected by using a dynamic selection strategy based on cost-expectation balance; the dynamic selection strategy builds a selection threshold by comprehensively considering search expectation and path cost ​ ； wherein, is a selection threshold, is a balancing coefficient, is a candidate item, is a desire, is a probability.

8. The dynamic long-term relocation method for coexistence of human and machine according to claim 1, wherein, The optimization objective function of the global planning layer is: ； wherein, is an objective function, is a cost function, is a probability density, is a logarithm function, is a summation symbol; The local planning layer utilizes item co-occurrence relations in the item knowledge base Guiding the robot to adjust the view step by step to the target item .

9. A dynamic item long-term relocation system for a human-machine coexistence scenario, characterized in that, Comprising: A probability model construction module, based on the accumulated item perception data in the long-term operation of the robot, extracts spatiotemporal features containing item positions and timestamps, and encodes the spatiotemporal features; based on the encoded spatiotemporal features, the Gaussian mixture model is used to model the item position distribution, the spatiotemporal probability model reflecting the long-term appearance regularity of the item is constructed, and the item knowledge base containing the co-occurrence relationship of the items is constructed; A dynamic item state discrimination module, according to the probability distribution result output by the spatiotemporal probability model, calculates the spatial encounter frequency and the time observation frequency of the item, combines the preset perception interaction ratio threshold, dynamically classifies the motion state of the item, and adjusts the priority of the item in the relocation task based on the classification; A task item reasoning module receives a user query request in natural language form, generates a candidate item list using a pre-trained large language model and the item knowledge base in combination with the priority of the item in the relocation task; selects the optimal task item from the candidate item list and determines its potential search location sequence; A hierarchical search planning module generates a search path based on the search location sequence using a hierarchical planning algorithm, where the global planning layer divides the search space into task sub-environments based on environmental physical boundaries, generates a search sequence for each sub-environment in combination with the spatio-temporal probability model, and the local planning layer adjusts the observation angle within each sub-environment using the co-occurrence relationship between items to achieve fine-grained search and backtrack or expand the search in the case of not finding the target.

10. The dynamic long-term relocation system for coexistence scenario between human and machine according to claim 9, characterized in that, Further comprising: A probability model updating module dynamically updates the mean, covariance and mixing weight of the Gaussian mixture model within the spatio-temporal neighborhood according to the latest perceived item location and time during the robot's execution of the search task, and adaptively adjusts the influence of historical data by introducing time distance and perception interaction ratio.

Images