A method for intelligent planning and management of a companion robot route in a scenic spot driven by reinforcement learning

Abstract

The application provides a kind of reinforcement learning driven companion robot scenic route intelligent planning management method, it is related to artificial intelligence technical field, the method includes: based on user physiological data, scenic opening time constraint and real-time flow information, generates initial tour path by reinforcement learning strategy, dynamically adjusts and optimizes the time of stay according to real-time physiological monitoring signal, outputs physiological adaptation path.The application improves the service intelligent level of companion robot and user experience.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and in particular to a reinforcement learning-driven intelligent planning and management method for tourist routes of tour guide robots. Background Technology

[0002] As the tourism industry transforms towards personalization and intelligence, companion robots are gradually becoming core equipment for scenic spots to improve service efficiency and tourist experience. The coverage rate of intelligent companion devices in high-level scenic spots in China is increasing, but existing path planning technology still faces key bottlenecks, making it difficult to meet the diverse needs of tourists and the dynamic management requirements of scenic spots. The route planning of traditional companion robots mainly relies on two types of solutions:

[0003] One type is the classic route algorithm based on static geographic data, which can only generate standardized routes based on fixed information such as the location and opening hours of attractions. It is completely unable to cope with the real-time changes in the density of people in scenic areas and sudden weather conditions, such as short-term rainfall and local high temperatures. This results in a serious lack of route flexibility and makes it easy for tourists to crowd together and be exposed to harsh environments.

[0004] Another type is a semi-dynamic planning scheme that simply integrates real-time data. Although it can adjust the route order by matching basic data, such as comparing real-time pedestrian flow data with the capacity threshold of attractions, it has two major drawbacks because it does not introduce the deep processing capabilities of neural network algorithms for multi-dimensional user data: First, it ignores the precise adaptation to individual user differences, such as the physiological tolerance data of elderly tourists (heart rate, blood oxygen, etc.) and the preference data of family tourists for children's facilities and rest stops. It can only provide a rough route through preset labels and cannot dynamically adjust based on the user's real-time physiological signals and behavioral characteristics. Second, it lacks the ability to collaboratively process multi-source dynamic data. Traditional data fusion methods cannot capture the temporal correlation and implicit features between data. For example, it cannot judge the fatigue state of users through the micro-expression changes of users over 5 consecutive minutes, nor can it conduct deep correlation analysis between real-time wind speed and attraction shading rate data, resulting in delayed route adjustments that do not meet the actual needs of users. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a reinforcement learning-driven intelligent planning and management method for tourist routes of escort robots, so as to improve the service intelligence level and user experience of escort robots.

[0006] To solve the above-mentioned technical problems, the technical solution of the present invention is as follows:

[0007] Firstly, a reinforcement learning-driven intelligent planning and management method for tourist attraction routes using a tour guide robot, the method comprising:

[0008] Based on user physiological data, attraction opening hours constraints, and real-time visitor flow information, an initial tour route is generated through reinforcement learning strategy. The stay time of the route is dynamically adjusted and optimized according to real-time physiological monitoring signals, and a physiologically adapted route is output.

[0009] The physiological adaptation path is optimized for environmental tolerance by integrating real-time meteorological data and the distribution of rain shelter facilities. Three collaborative sensing nodes are dynamically deployed along the path to form an environmental monitoring domain. An environmental coordination factor is generated by calculating the spatial configuration descriptor of the monitoring domain. Based on the environmental coordination factor, an adaptive weighted fusion of environmental assessment indicators is achieved to obtain the environmental optimization path.

[0010] Based on the environmental optimization path, the robot's perception space is discretized into a structured grid map, and the path adaptive control coefficient is solved. According to the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intention feature extraction on the user's micro-expressions and body movements. The neural network learns the deep features of the temporal data through a multilayer perceptron structure and outputs an intention encoding vector. By calculating the semantic matching degree between the user's real-time behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate a preference-driven dynamic path sequence.

[0011] Based on preference-driven dynamic path sequences, the system monitors the operational status of attractions and facility data in real time. It then performs global replanning of the sequence based on spatiotemporal correlation constraints to avoid unavailable facilities and generate the final optimized route.

[0012] Furthermore, based on user physiological data, attraction opening hours constraints, and real-time visitor flow information, an initial tour route is generated through reinforcement learning. This route is then dynamically adjusted and the dwell time optimized based on real-time physiological monitoring signals, resulting in a physiologically adapted route, including:

[0013] It receives and integrates real-time physiological data uploaded by users, pre-stored scenic spot opening time constraints, and real-time scenic spot visitor flow information. It performs time synchronization and standardization processing on multi-source data, extracts its feature vectors, and fuses them to form a reinforcement learning state space.

[0014] Based on the reinforcement learning state space, a decision-making reasoning process is performed through a pre-trained reinforcement learning policy network to obtain an initial tour path sequence that satisfies all time constraints and effectively avoids high-traffic areas.

[0015] The system acquires users' physiological monitoring signals in real time and synchronously couples these signals with the current execution status of the initial tour path. It then dynamically adjusts the path by calculating the deviation between the actual physiological load and the predicted value.

[0016] Based on the dynamic adjustment process and combined with the terrain data obtained from the geographic information database, low-lying attractions are prioritized for inclusion in the route sequence. The optimal stay duration for each attraction is calculated based on the user's real-time physiological state, resulting in a physiologically adapted route optimized in multiple dimensions such as user physiological load, time constraints, and crowd density.

[0017] Furthermore, the physiological adaptation path is optimized for environmental tolerance. Real-time meteorological data and the distribution of rain shelters are integrated, and three collaborative sensing nodes are dynamically deployed along the path to form an environmental monitoring domain, including:

[0018] Based on the physiological adaptation path, access the real-time meteorological sensor data stream; at the same time, obtain the location data of rain shelter facilities in the scenic area's geographic information system, and analyze the spatial distribution density and coverage characteristics of rain shelter facilities around the path.

[0019] Based on path information and the distribution characteristics of rain shelters, combined with real-time meteorological data, the final deployment location of three collaborative sensing nodes is dynamically calculated along the periphery of the physiological adaptation path. The deployment location must meet the requirements of maximizing the monitoring coverage of the path environment and having the highest sensitivity to meteorological changes.

[0020] Based on the final deployment location, the spatial layout of the three collaborative sensing nodes is dynamically adjusted to form a variable-scale environmental monitoring domain.

[0021] Based on a variable-scale environmental monitoring domain, multi-physics environmental data are collected in real time. By calculating the spatial configuration parameters of the monitoring domain, an environmental coordination factor is generated, resulting in an environmental monitoring domain that adapts to changes in meteorological conditions.

[0022] Furthermore, by calculating the spatial configuration descriptor of the monitoring domain to generate an environmental coordination factor, and based on the environmental coordination factor, an adaptive weighted fusion of environmental assessment indicators is achieved to obtain an environmental optimization path, including:

[0023] Based on the obtained variable-scale environmental monitoring domain, the spatial configuration descriptor of the monitoring domain is calculated, including the area parameter, azimuth angle parameter and node spacing ratio parameter of the monitoring domain, and the spatial geometric features of the monitoring domain are quantified through the parameters.

[0024] Based on the spatial configuration descriptor and combined with real-time meteorological data, an environmental coordination factor is generated. The environmental coordination factor is used to characterize the degree of matching and coordination between the current environmental monitoring domain configuration and the external environmental conditions.

[0025] Based on the environmental coordination factor, multiple environmental assessment indicators such as temperature, humidity, wind speed and rainfall probability are adaptively weighted and fused. The environmental coordination factor is used to dynamically adjust the weight allocation of each environmental indicator to obtain the weighted fusion result.

[0026] Based on the weighted fusion results, and by introducing an error compensation mechanism for the sensor system, the monitoring data is corrected to obtain the environmental risk assessment results.

[0027] Based on the environmental risk assessment results, the original route was optimized and adjusted to avoid high environmental risk areas and prioritize routes with suitable environments and rain shelter facilities, resulting in an environmentally optimized route.

[0028] Furthermore, based on environmental path optimization, the path adaptive control coefficients are solved by discretizing the robot's perception space into a structured grid map, including:

[0029] Based on the environmental optimization path, the environmental perception space around the robot is discretized according to a preset resolution to obtain a structured grid map containing environmental feature information.

[0030] Based on a structured grid map, geometric feature parameters of the environmental optimization path are extracted, and real-time user status data is obtained. The fitness evaluation value of the path is calculated through a multi-factor weighted fusion algorithm.

[0031] Based on the path fitness assessment value, combined with the environmental complexity index and the intensity of user personalized needs, the weighted average method is used to integrate the environmental complexity weight and the user personalized needs weight, and then the path adaptive control coefficient is obtained through normalization.

[0032] Based on the path adaptive control coefficient and combined with real-time collected environmental change data, the coefficient parameters are dynamically adjusted through a sliding window algorithm to obtain an adaptive control coefficient that accurately reflects the current environmental adaptability and the user's personalized needs.

[0033] Furthermore, based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intent feature extraction on user micro-expressions and body movements. The neural network learns deep features of temporal data through a multilayer perceptron structure and outputs an intent encoding vector. By calculating the semantic matching degree between real-time user behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate preference-driven dynamic path sequences, including:

[0034] Based on the path adaptive control coefficient, real-time acquisition of users' facial images and limb movement sequence data is performed, preprocessed and labeled with features to construct a user behavior time series dataset.

[0035] A neural network model based on a multilayer perceptron structure is used to perform deep feature learning on a time-series dataset of user behavior, extract the time-series correlation features between micro-expression change patterns and body movements, and generate a high-dimensional intent encoding vector.

[0036] The high-dimensional intent encoding vector is matched with the historical preference vector retrieved from the user's historical behavior database to calculate semantic similarity. Based on the matching results, the semantic matching degree between the user's current behavioral intent and historical preferences is output.

[0037] Based on semantic matching degree and combined with the information on the stopable areas of each attraction in the environmental optimization path, the embedding position and recommended stay duration of personalized stop locations are dynamically calculated.

[0038] Based on the embedded results of personalized stopping points, the environmental optimization path is reorganized and updated in real time to generate a dynamic path sequence that integrates the user's real-time intent and historical behavioral preferences.

[0039] Furthermore, based on preference-driven dynamic path sequences, the system monitors the operational status of attractions and facility data in real time. It then performs global replanning of the sequence based on spatiotemporal correlation constraints, avoiding unavailable facilities and generating a final optimized route, including:

[0040] Based on the preference-driven dynamic path sequence, and by accessing the scenic area management system in real time the data stream of scenic spot operation status and facility availability data, the spatiotemporal correlation constraints of each scenic spot in the path sequence are verified to obtain the verification results.

[0041] Based on the verification results, unavailable attractions or facilities are detected in the path sequence. Based on the spatiotemporal correlation constraints between multiple attractions and user preference characteristics, a candidate set of alternative attractions is generated.

[0042] Based on the candidate set of alternative attractions, combined with real-time environmental data and user status information, a new route plan that avoids unavailable facilities is generated through a route replanning algorithm, and a comprehensive evaluation index for each alternative plan is calculated.

[0043] Based on the new path scheme, the path with the final comprehensive evaluation index is selected as the final optimized route, and an emergency backup plan is also provided for the final optimized route, thus obtaining a multi-path optimization scheme.

[0044] Secondly, a reinforcement learning-driven intelligent route planning and management system for tour guide robots includes:

[0045] The acquisition module is used to generate an initial tour route based on user physiological data, attraction opening time constraints, and real-time visitor flow information through reinforcement learning strategies. It then dynamically adjusts and optimizes the stay time of the route based on real-time physiological monitoring signals and outputs a physiologically adapted route.

[0046] The computation module is used to optimize the environmental tolerance of the physiological adaptation path, integrate real-time meteorological data and the distribution of rain shelter facilities, and dynamically deploy three collaborative sensing nodes along the path to form an environmental monitoring domain; by calculating the spatial configuration descriptor of the monitoring domain, an environmental coordination factor is generated, and based on the environmental coordination factor, an adaptive weighted fusion of environmental assessment indicators is achieved to obtain the environmental optimization path;

[0047] The control module is used to optimize the path based on the environment. It solves the path adaptive control coefficient by discretizing the robot's perception space into a structured grid map. Based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intention feature extraction on the user's micro-expressions and body movements. The neural network learns the deep features of the temporal data through a multilayer perceptron structure and outputs an intention encoding vector. By calculating the semantic matching degree between the user's real-time behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate a preference-driven dynamic path sequence.

[0048] The processing module is used to monitor the operational status of attractions and facility data in real time based on preference-driven dynamic path sequences, perform global replanning on the sequence based on spatiotemporal correlation constraints, avoid unavailable facilities, and generate the final optimized route.

[0049] Thirdly, a computing device, comprising:

[0050] One or more processors;

[0051] A storage device for storing one or more programs that, when executed by one or more processors, cause the one or more processors to implement the method.

[0052] Fourthly, a computer-readable storage medium storing a program that, when executed by a processor, implements the method.

[0053] The above-described solution of the present invention has at least the following beneficial effects:

[0054] By combining user physiological data, attraction opening time constraints, and real-time visitor flow information with reinforcement learning strategies, an initial tour route is generated, and the dwell time is dynamically adjusted based on real-time physiological monitoring signals. Three collaborative sensing nodes are dynamically deployed along the physiologically adapted path to form an environmental monitoring domain, and environmental coordination factors are generated based on spatial configuration descriptors to optimize environmental tolerance. The robot's perception space is discretized into a structured grid map to solve for path adaptive control coefficients. Multilayer perceptron neural networks are used to perform temporal pattern analysis of user micro-expressions and body movements to embed personalized stopping points. Simultaneously, the operational status of attractions and facility data are monitored in real-time, and global replanning is performed based on spatiotemporal correlation constraints. Therefore, this approach overcomes the limitations of traditional static path algorithms, which cannot handle real-time visitor flow and sudden weather conditions, as well as the technical bottlenecks of semi-dynamic programming schemes that ignore user physiological tolerance and behavioral preferences, lack multi-source dynamic data collaborative processing capabilities, and fail to consider the real-time availability of attraction facilities. This achieves optimization of the tour robot's route in terms of physiological adaptability, environmental adaptability, and satisfaction of personalized user needs, avoiding high-risk areas and unavailable facilities, and improving the level of service intelligence and the visitor experience. Attached Figure Description

[0055] Figure 1 This is a flowchart illustrating a reinforcement learning-driven intelligent planning and management method for tourist routes using a tour guide robot, provided by an embodiment of the present invention.

[0056] Figure 2 This is a schematic diagram of a reinforcement learning-driven intelligent planning and management system for tourist routes of a tour guide robot, provided by an embodiment of the present invention. Detailed Implementation

[0057] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the disclosure to those skilled in the art.

[0058] like Figure 1 As shown, an embodiment of the present invention proposes a reinforcement learning-driven intelligent planning and management method for tourist route planning of a tour guide robot. The method includes the following steps:

[0059] Step 1: Based on user physiological data, attraction opening time constraints, and real-time visitor flow information, an initial tour route is generated through reinforcement learning strategy. The stay time of the route is dynamically adjusted and optimized according to real-time physiological monitoring signals, and a physiologically adapted route is output.

[0060] Step 2: Optimize the environmental tolerance of the physiological adaptation path by integrating real-time meteorological data and the distribution of rain shelter facilities, and dynamically deploy three collaborative sensing nodes along the path to form an environmental monitoring domain; generate an environmental coordination factor by calculating the spatial configuration descriptor of the monitoring domain, and achieve adaptive weighted fusion of environmental assessment indicators based on the environmental coordination factor to obtain the environmental optimization path.

[0061] Step 3: Based on the environment-optimized path, the robot's perception space is discretized into a structured grid map, and the path adaptive control coefficient is solved. Based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intention feature extraction on the user's micro-expressions and body movements. The neural network learns the deep features of the temporal data through a multilayer perceptron structure and outputs an intention encoding vector. By calculating the semantic matching degree between the user's real-time behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate a preference-driven dynamic path sequence.

[0062] Step 4: Based on the preference-driven dynamic path sequence, monitor the operational status of attractions and facility data in real time, and perform global replanning of the sequence based on spatiotemporal correlation constraints to avoid unavailable facilities and generate the final optimized route.

[0063] In this embodiment of the invention, an initial path is generated by integrating user physiological data, scenic spot opening time constraints, and real-time crowd flow information through reinforcement learning. This path is dynamically adjusted and prioritized for low-lying scenic spots based on real-time physiological signals. Real-time meteorological data and the Poisson distribution of rain shelters are integrated, and three collaborative sensing nodes are deployed along the path to construct a facility monitoring domain. Environmental coordination factors are calculated to optimize the environmental path. The robot's perception space is discretized into a grid map to calculate control coefficients. User behavior characteristics are analyzed, and historical preferences are matched to dynamically embed stopping points. Scenic spot facility data is monitored, and the route is replanned based on spatiotemporal constraints, providing backup solutions. Therefore, this overcomes the problems of traditional algorithms, such as difficulty in handling real-time crowd flow and weather, neglecting user physiological differences, lack of dynamic risk assessment in environmental perception, lack of personalized adaptation, and difficulty in avoiding route interruptions due to unavailable facilities. This achieves the effect of real-time adaptation to user physiological and environmental interference, accurate fulfillment of personalized needs, and ensuring continuous and stable routes, thereby improving the intelligence level of the companion robot and enhancing user comfort, safety, and customized experience.

[0064] In a preferred embodiment of the present invention, step 1 above may include:

[0065] Step 1.1: Receive and integrate real-time physiological data uploaded by the user, pre-stored attraction opening time constraints, and acquired real-time attraction visitor flow information. Perform time synchronization and standardization processing on the multi-source data, and extract and fuse their feature vectors to form a reinforcement learning state space. Specifically, this includes: the companion robot receives real-time physiological data uploaded by the user via wearable devices through a preset data source access channel, simultaneously calls its own local data storage unit to extract pre-stored attraction opening time constraint data, and obtains real-time attraction visitor flow information through a communication link established with the scenic area management system. This achieves the initial collection and fusion of three types of data: user physiological data, attraction opening time constraint data, and real-time visitor flow information. Addressing potential time-related discrepancies in the aforementioned multi-source data... To address the synchronization issue, timestamp alignment was used to synchronize the three types of data, ensuring consistency across all data points at the same time. Simultaneously, considering the potential interference from differences in data magnitude for subsequent analysis, data normalization was applied to standardize physiological data, pedestrian flow data, and time-constrained data to eliminate the impact of these differences. Feature extraction algorithms were then used to extract key feature vectors from the synchronized and standardized multi-source data. For example, heart rate trend features were extracted from physiological data, regional density features from pedestrian flow data, and scenic spot opening hours features from time-constrained data. These extracted feature vectors were then fused according to predefined fusion rules to ultimately form the state space required for reinforcement learning.

[0066] Step 1.2: Based on the reinforcement learning state space, a pre-trained reinforcement learning policy network is used for decision-making reasoning to obtain an initial tour path sequence that satisfies all time constraints and effectively avoids high-traffic areas. Specifically, this includes: Based on the constructed reinforcement learning state space, the tour guide robot calls its built-in pre-trained reinforcement learning policy network. The policy network is pre-trained using a large amount of historical tour data of scenic spots, user physiological feedback data, and path optimization case data, enabling it to make path decisions under multiple constraints. During the decision-making reasoning process, the reinforcement learning policy network takes the feature vectors in the reinforcement learning state space as input and uses the pre-set decision logic within the network to determine the appropriate path. The feasibility and optimization level of each potential tour route are evaluated one by one. The evaluation includes, but is not limited to, whether each route meets the opening time constraints of all attractions, whether it can effectively avoid the current high-traffic areas, and whether it meets the preset total route duration requirements. At the same time, combined with the preset route optimization objectives, such as the rationality of tour coverage and the efficiency of route movement, each potential route is comprehensively judged and ranked. After the above complete decision-making and reasoning process, the reinforcement learning policy network finally outputs an initial tour route sequence that meets all time constraints and can effectively avoid high-traffic areas. The initial tour route sequence specifies the order in which the user visits attractions and the movement path between attractions.

[0067] Step 1.3: Acquire the user's physiological monitoring signals in real time and synchronously couple these signals with the current execution status of the initial tour path. Dynamically adjust the path by calculating the deviation between the actual physiological load and the predicted value. Specifically, this includes: continuously acquiring the user's physiological monitoring signals through a link maintaining real-time communication with the user's wearable device, based on the initial tour path process. These physiological monitoring signals include, but are not limited to, parameters that directly reflect the user's current physiological load, such as heart rate, respiratory rate, and cadence. Simultaneously, the robot, relying on its own path execution monitoring function, records the current execution status of the initial tour path in real time, including the current location of the attraction, the completed tour duration, the attractions to be visited, and the estimated travel time between attractions. The real-time acquired physiological monitoring signals are then used to... The system synchronously couples the current path execution status with the data to ensure complete matching of the two types of data in the time dimension. Based on this, it combines the user's historical physiological data, the current path execution status, and the characteristics of the attractions to be visited to calculate the expected physiological load value that the user should have at the current stage of the visit. The expected physiological load value is then compared with the actual physiological load value obtained in real time to obtain the deviation value between the two. When the calculated deviation value exceeds a preset reasonable threshold, it is determined that the currently executed path may cause the user's physiological load to be too high. Then, based on the specific magnitude and direction of the deviation value, the path to be executed is dynamically adjusted. The adjustment methods include, but are not limited to, shortening the stay time at the current attraction, advancing or delaying the visit to some subsequent attractions, and changing the movement route between some attractions.

[0068] Step 1.4: Based on the dynamic adjustment process and combined with terrain data obtained from the geographic information database, low-lying scenic spots are prioritized for inclusion in the route sequence. The optimal stay duration for each scenic spot is calculated based on the user's real-time physiological state, resulting in a physiologically adapted route optimized across multiple dimensions, including user physiological load, time constraints, and crowd density. Specifically, during the dynamic route adjustment process, the tour robot accesses the preset geographic information database to obtain terrain data for all scenic spots within the area. This terrain data includes information that directly affects the user's physiological load, such as the altitude of each scenic spot, the slope of the internal tour path, and the undulation of the connecting roads between scenic spots. This data is used to assess the impact of different scenic spots and routes on the user's physiological load. Based on the obtained terrain data, low-lying scenic spots with lower altitudes, gentler internal path slopes, and smoother connecting roads are prioritized for selection. Points were incorporated into the adjusted route sequence to avoid further increasing the user's physiological load due to visiting high-altitude attractions. Simultaneously, combining continuously acquired real-time physiological parameters such as current heart rate, respiratory rate, and fatigue assessment results derived from historical data, a weighted analysis method was used to calculate the optimal stay duration for each selected low-altitude attraction. This weighted analysis fully considered the user's current physiological tolerance to the stay duration, ensuring that the calculated stay duration met the user's need for ample sightseeing without causing excessive physiological load beyond a reasonable range. Through the above-mentioned attraction selection combining terrain data, stay duration optimization based on real-time physiological status, and coordinated with dynamic route adjustments, a physiologically adapted route was ultimately obtained that optimized across three dimensions: user physiological load, time constraints, and crowd density.

[0069] In this embodiment of the invention, by integrating real-time user physiological data, pre-stored attraction opening time constraints, and real-time crowd flow information, and extracting feature vectors after time synchronization and standardization processing to form a reinforcement learning state space, an initial path is generated using a pre-trained reinforcement learning strategy network. Real-time physiological monitoring signals are then synchronously coupled with the current execution state of the path to dynamically adjust the path. Simultaneously, terrain data is used to prioritize low-lying attractions and calculate the optimal stay duration for each attraction. Therefore, this overcomes the technical problems of traditional path planning, which relies solely on static data and cannot cope with real-time crowd flow changes, ignores differences in user physiological tolerance, lacks multi-source data collaborative processing, and lacks dynamic path adjustment mechanisms. This achieves the technical effect of generating physiologically adapted paths optimized in multiple dimensions of user physiological load, time constraints, and crowd density, reducing user physiological burden, avoiding high-traffic areas, and improving visitor comfort and safety.

[0070] In a preferred embodiment of the present invention, step 2 above may include:

[0071] Step 2.1: Based on the physiological adaptation path, access the real-time meteorological sensor data stream; simultaneously, acquire the location data of rain shelter facilities in the scenic area's geographic information system, and analyze the spatial distribution density and coverage characteristics of the rain shelter facilities around the path. Specifically, this includes: based on the generated physiological adaptation path, accessing the real-time meteorological sensor data stream through a preset meteorological data communication link. The data stream contains dynamic environmental parameters such as temperature, humidity, wind speed, and rainfall probability. At the same time, the robot retrieves the location coordinate data of the rain shelter facilities through the data interface of the scenic area's geographic information system, and uses spatial statistical methods to analyze this data to clarify the distribution density pattern of the rain shelter facilities around the physiological adaptation path, i.e., the number of facilities per unit area, and the actual coverage of each facility in the path area, including the path length and radiation radius that the facility can shelter.

[0072] Step 2.2: Based on path information and the distribution characteristics of rain shelters, combined with real-time meteorological data, the final deployment positions of the three collaborative sensing nodes are dynamically calculated along the periphery of the physiological adaptation path. The deployment positions must meet the requirements of maximizing the monitoring coverage of the path environment and having the highest sensitivity to meteorological changes. Specifically, based on the physiological adaptation path information, the distribution characteristics of rain shelters, and real-time meteorological data, the companion robot starts the node position calculation program. The program first divides the monitoring sections according to the total path length and terrain undulation characteristics, and then, combined with the coverage of rain shelters and the risk warning parameters in the real-time meteorological data, evaluates the monitoring effectiveness of potential deployment points one by one along the periphery of the path. During the evaluation process, two core indicators are considered: first, the spatial coverage ratio of the monitoring points to the path must reach a preset threshold or higher; second, the response sensitivity to changes in meteorological parameters must meet the real-time monitoring requirements. Through multiple rounds of iterative calculations, the final deployment coordinates of the three collaborative sensing nodes are finally determined, ensuring that each node can form a monitoring overlap area to ensure data redundancy, and can quickly capture key environmental parameters in the event of sudden weather changes.

[0073] Step 2.3: Based on the final deployment location, dynamically adjust the spatial layout of the three collaborative sensing nodes to form a variable-scale environmental monitoring domain. Specifically, this includes: based on the obtained final deployment location, the companion robot controls the three collaborative sensing nodes to initiate a spatial layout adjustment mechanism. Each node obtains real-time location information through its built-in positioning and achieves mutual position calibration through short-range wireless communication. The relative distance and arrangement angle between nodes are dynamically adjusted according to the path direction. When the path changes direction or the environmental characteristics change significantly, the nodes automatically adjust their spatial distribution shape, expanding or shrinking the monitoring range to adapt to changes in path length, ultimately forming an environmental monitoring domain with dynamically expandable boundaries.

[0074] Step 2.4: Based on the variable-scale environmental monitoring domain, multi-physics environmental data is collected in real time. An environmental coordination factor is generated by calculating the spatial configuration parameters of the monitoring domain, resulting in an environmental monitoring domain that adapts to changes in meteorological conditions. Specifically, this includes: based on the variable-scale environmental monitoring domain, three collaborative sensing nodes simultaneously activate data acquisition functions to obtain real-time multi-physics environmental data such as temperature, humidity, and airflow around the path. The robot analyzes the spatial geometric features of the monitoring domain, extracting the distance ratio, coverage angle, and spatial distribution entropy parameters between nodes as spatial configuration parameters. These spatial configuration parameters are then correlated with the real-time meteorological data. First, the corresponding correlation between the spatial configuration parameters and real-time meteorological data, as well as the weights of each correlation dimension, are determined. Specifically, the distance ratio between nodes corresponds to real-time wind speed data, with higher wind speeds resulting in higher weights for this dimension; the coverage angle corresponds to real-time rainfall probability data, with higher rainfall probabilities resulting in higher weights for this dimension; and the spatial distribution entropy corresponds to real-time temperature and humidity comprehensive data, with greater temperature and humidity fluctuations resulting in higher weights for this dimension. The weights of each dimension are determined based on the historical meteorological data of the scenic area and the monitoring results. The data is pre-set with a total weight of 1. Next, the distance ratio between nodes, coverage angle, and spatial distribution entropy are standardized, mapping each parameter value to the 0-1 range to eliminate magnitude differences. The same standardization process is applied to real-time wind speed, rainfall probability, and temperature and humidity composite data. Then, for each correlation dimension, the standardized spatial configuration parameter is multiplied by the corresponding standardized meteorological data, and then multiplied by the pre-set weight for that dimension to obtain the weighted correlation value for each dimension. Specifically, the first dimension is standardized distance ratio × standardized wind speed × distance-wind speed correlation weight; the second dimension is standardized coverage angle × standardized rainfall probability × angle-rainfall correlation weight; and the third dimension is standardized distribution entropy × standardized temperature and humidity data × entropy value-temperature and humidity correlation weight. Finally, the weighted correlation values ​​of the three dimensions are summed to obtain the initial result, which is then mapped to the 0-1 range for normalization. The final value is the environmental coordination factor. Through dynamic updates of the environmental coordination factor, the environmental monitoring domain can automatically optimize data acquisition frequency and monitoring accuracy, achieving adaptive adjustment to changes in meteorological conditions.

[0075] In this embodiment of the invention, by combining physiological adaptation path access to real-time meteorological sensor data stream, obtaining location data of rain shelter facilities in scenic areas and analyzing their surrounding distribution characteristics, and then co-calculating with real-time meteorological data to determine the final deployment positions of the three sensor nodes that maximize monitoring coverage and have the highest meteorological sensitivity, the node layout is adjusted to form a variable-scale monitoring domain. Finally, multi-physics field data is collected through this monitoring domain and spatial configuration parameters are calculated to generate environmental coordination factors. Therefore, this invention overcomes the technical problems of traditional monitoring that do not combine path and rain shelter facilities, fixed sensor nodes leading to incomplete coverage or low sensitivity, unadjustable monitoring domain, and lack of environmental matching parameters. As a result, it achieves accurate monitoring of the surrounding environment of the path, dynamic adaptation of the monitoring domain with meteorological conditions, and environmental coordination factors, thereby improving the robot's environmental perception and adaptability.

[0076] In a preferred embodiment of the present invention, step 2 above may include:

[0077] Step 2.5: Based on the obtained variable-scale environmental monitoring domain, calculate the spatial configuration descriptor of the monitoring domain, including the area parameter, azimuth angle parameter, and node spacing ratio parameter. Quantify the spatial geometric features of the monitoring domain through these parameters. Specifically, based on the formed variable-scale environmental monitoring domain, the companion robot first initiates the spatial parameter calculation process to conduct a comprehensive quantification of the geometric features of the monitoring domain. The area parameter of the monitoring domain is obtained by on-site measurement of the spatial range enclosed by the boundary of the monitoring domain or by calculating the coordinate data from the scenic area's geographic information system. The azimuth angle parameter is obtained by determining the angle between the overall orientation of the monitoring domain and the direction of the physiological adaptation path. The node spacing ratio parameter is obtained by measuring the actual distance between each pair of the three collaborative sensing nodes and calculating the proportional relationship between each distance value. Through the precise calculation of the above three parameters, the originally abstract spatial geometric features of the monitoring domain are transformed into specific and quantifiable numerical information.

[0078] Step 2.6: Based on the spatial configuration descriptor and combined with real-time meteorological data, an environmental coordination factor is generated. The environmental coordination factor is used to characterize the degree of matching and coordination between the current environmental monitoring domain configuration and the external environmental conditions. Specifically, based on the obtained spatial configuration descriptor, the companion robot performs multi-dimensional correlation analysis on the area, azimuth angle, and node spacing ratio parameters contained in the descriptor with the real-time collected meteorological data. The real-time meteorological data here covers key information such as temperature change trends, humidity fluctuations, wind speed and direction, and rainfall probability and intensity. In the correlation analysis process, the focus is on whether the spatial configuration of the monitoring domain matches the monitoring needs under the current meteorological conditions. For example, when the wind speed is high, it is necessary to determine whether the azimuth angle of the monitoring domain can effectively capture wind direction changes and whether the node spacing can ensure the continuity of wind speed data collection. Through correlation analysis calculation, the environmental coordination factor is finally generated. The environmental coordination factor intuitively characterizes the degree of matching and coordination between the current environmental monitoring domain spatial configuration and the external real-time meteorological conditions in a specific numerical form.

[0079] Step 2.7: Based on the environmental coordination factor, an adaptive weighted fusion of multiple environmental assessment indicators, including temperature, humidity, wind speed, and rainfall probability, is performed. The environmental coordination factor is used to dynamically adjust the weight allocation of each environmental indicator to obtain the weighted fusion result. Specifically, this involves: The tour guide robot constructing a dynamic weighting mechanism for environmental assessment indicators based on the environmental coordination factor. First, initial basic weights are set for the four environmental assessment indicators: temperature, humidity, wind speed, and rainfall probability. These initial weights are determined based on historical environmental risk data of the scenic area and conventional tour comfort requirements. Then, the initial weights of each indicator are adaptively adjusted according to the real-time changes in the environmental coordination factor. If the environmental coordination factor indicates a significant risk of rainfall under current meteorological conditions, the weight of the rainfall probability indicator is increased accordingly. If the factor indicates that temperature has a greater impact on the tour experience, the weight of the temperature indicator is increased. Through dynamic weighted calculation, the monitoring data of the four environmental assessment indicators are fused and calculated according to the adjusted weights to obtain a weighted fusion result that reflects the actual risk and comfort level of the current environment.

[0080] Step 2.8: Based on the weighted fusion results, an error compensation mechanism for the sensor system is introduced to correct the monitoring data and obtain the environmental risk assessment results. Specifically, based on the weighted fusion results, the companion robot further introduces a sensor system error compensation mechanism to eliminate deviations in the monitoring data and improve the accuracy of the environmental risk assessment. The error compensation mechanism first retrieves the historical calibration records of the sensors and compares the deviation values ​​between the current sensor-collected data and the historical standard calibration data, thereby correcting the systematic errors in the monitoring data. At the same time, combined with the spatiotemporal variation patterns of current environmental parameters, such as the normal fluctuation range of temperature at different times and the intermittent variation characteristics of wind speed, the random errors that may exist in the monitoring data are smoothed. Through the dual correction process of systematic error correction and random error smoothing, the originally biased monitoring data is optimized, and finally, an environmental risk assessment result that can truly reflect the current environmental conditions is generated.

[0081] Step 2.9: Based on the environmental risk assessment results, optimize and adjust the original route to avoid high-environmental-risk areas. Prioritize routes with suitable environments and rain shelters to obtain an optimized environmental route. Specifically, this includes: Based on the environmental risk assessment results, the tour guide robot initiates a route optimization and adjustment program. First, it analyzes the assessment results one by one, screening out areas where the risk value exceeds a preset safety threshold. These high-environmental-risk areas are marked as route avoidance targets to ensure that the adjusted route does not pass through such areas. Simultaneously, the tour guide robot accesses the scenic area's geographic information system to retrieve route segments whose environmental parameters meet the tour comfort standards and are covered by rain shelters. These route segments must meet temperature and humidity requirements. Within a suitable range, with low wind speeds and a low probability of rainfall, and with adequate coverage of rain shelters to ensure emergency shelter for tourists in the event of sudden downpours, the attributes of each suitable route segment are extracted. This identifies the starting and ending coordinates, intermediate nodes, estimated travel time, and associated attraction information for each route segment, establishing a route segment attribute database. Subsequently, based on the original route's attraction visit order and the logical relationships between attractions, such as the spatial relationships between adjacent attractions and the order of visits, the route segments in the attribute database are initially sorted. This ensures that the sorting results conform to the logical habits of normal user visits, avoiding situations where the order of attraction visits is reversed or there are excessive spatial jumps. Next, for the sorted adjacent path segments, the algorithm performs spatial matching between their endpoints and starting points, calculating the shortest connecting path between the two points. The connecting path must avoid marked high-risk environmental areas and prioritize passages with shelter or suitable environmental conditions. Simultaneously, the estimated travel time for the connecting path is calculated and incorporated into the overall time calculation system. Based on this, the algorithm sums the total travel time of all path segments and connecting paths with the preset stay time at each attraction, comparing this sum with the user's total available tour time and the opening hours of each attraction. If the total time exceeds the constraints, the algorithm fine-tunes the travel speed parameters or attraction stay times for some path segments, or replaces some path segments with more suitable ones. The algorithm selects suitable alternative path segments until the time constraint is met. It then optimizes the coherence of the overall path by smoothing the turning angles of adjacent path segments, reducing sharp bends or turns, and improving the smoothness of the path. At the same time, it checks for repeated areas in the path and eliminates them by adjusting the order of intermediate path segments or replacing connecting paths. Finally, the algorithm performs a second verification on the optimized overall path to confirm that it completely avoids high-environmental-risk areas, ensures that all path segments meet the requirements for environmental suitability and rain shelter coverage, and that the overall tour order and time constraint meet the preset conditions. After successful verification, the algorithm outputs the completed environmentally optimized path.

[0082] In this embodiment of the invention, a spatial configuration descriptor, including area parameters, azimuth angle parameters, and node spacing ratio parameters, is calculated through a variable-scale environmental monitoring domain to quantify the spatial geometric characteristics of the monitoring domain. Real-time meteorological data is combined to generate an environmental coordination factor characterizing the degree of matching and coordination between the current environmental monitoring domain configuration and external environmental conditions. Based on this environmental coordination factor, multiple environmental assessment indicators, including temperature, humidity, wind speed, and rainfall probability, are adaptively weighted and fused to dynamically adjust the weight allocation of each indicator. Simultaneously, an error compensation mechanism for the sensor system is introduced to correct the monitoring data to obtain an environmental risk assessment result. Finally, the original path is optimized based on this environmental risk assessment result. This approach, which integrates and avoids high-risk environmental areas while prioritizing routes with suitable environments and rain shelter coverage, effectively overcomes the technical problems of traditional environmental optimization processes. These problems include difficulties in determining environmental configuration matching due to the lack of quantification of the spatial geometric characteristics of the monitoring domain, fixed environmental indicator weights that cannot adapt to dynamic meteorological conditions, insufficient monitoring data accuracy due to lack of error compensation, and poor environmental adaptability due to the failure to combine precise environmental risk and rain shelter distribution in route optimization. As a result, it can accurately characterize the matching status between the monitoring domain and the environment, achieve dynamic adaptation of environmental assessment indicator weights with meteorological conditions, improve the accuracy of monitoring data, and ultimately generate environmentally optimized routes that take into account low risk, environmental suitability, and rain shelter protection.

[0083] In a preferred embodiment of the present invention, step 3 above may include:

[0084] Step 3.1: Based on the environmental optimization path, the environmental perception space around the robot is discretized according to a preset resolution to obtain a structured grid map containing environmental feature information. Specifically, this includes: Based on the obtained environmental optimization path, the tour robot first determines its current tour location and uses this as the center to delineate the environmental perception space range to be monitored. This range must cover the preset distance of the subsequent extension of the environmental optimization path to ensure complete capture of the environmental information around the path; Based on the complexity of the scenic area terrain and the robot's perception requirements, the preset resolution for discretization is determined. For example, a larger resolution is used in areas with gentle terrain and simple environment to improve processing efficiency, while a smaller resolution is used in areas with complex terrain and many obstacles to ensure perception accuracy; According to the preset resolution, the delineated environmental perception space is evenly divided into several independent grids, and each grid is labeled with corresponding environmental feature information, including but not limited to whether there are obstacles in the grid, whether it belongs to a scenic area, whether it is a passageway, and the width of the passageway. Through the above discretization processing and feature labeling, a structured grid map containing complete environmental feature information is finally formed.

[0085] Step 3.2: Based on the structured grid map, extract the geometric feature parameters of the environmental optimization path, and simultaneously acquire the user's real-time status data. Calculate the path's fitness evaluation value using a multi-factor weighted fusion algorithm. Specifically, based on the constructed structured grid map, the companion robot initiates a geometric feature extraction program, identifying the grid sequence traversed by the environmental optimization path in the grid map one by one. Based on this, it calculates and extracts the path's geometric feature parameters, including the overall path length, the curvature of each segment reflecting the path's smoothness, the slope of the segment calculated using grid elevation data, and the minimum distance between the path and surrounding obstacles. Simultaneously, the robot, through collaboration with the user's wearable device and its own perception, acquires the user's real-time status data. The algorithm incorporates the user's current physiological state, such as heart rate and fatigue level, and behavioral feedback, such as movement speed and dwell frequency. A multi-factor weighted fusion algorithm is then employed. First, the set of factors participating in the fusion calculation is defined. This set contains two types of factors: one is the environmental optimization path geometric feature parameters extracted from the structured raster map, specifically the overall path length, curvature of each segment, segment slope, and minimum distance between the path and surrounding obstacles; the other is the acquired real-time user status data, specifically the user's heart rate, fatigue level, movement speed, and dwell frequency. The two types of factors are merged to determine the total number of factors as n, which are sequentially denoted as factors 1 to n. Next, an initial weight is assigned to each factor, where the initial weight of the i-th factor, with i ranging from 1 to n, is denoted as w. i Each w i It needs to be determined based on the scenic area's historical visitor data, such as the correlation data between different path factors and user experience ratings, as well as past user feedback results, and all w i The sum of all additions must equal 1 to ensure the rationality of the weight allocation; then, the actual value of each factor is preprocessed, and the value of the i-th factor obtained through actual measurement or calculation is denoted as x. i Because the numerical magnitudes of different factors differ, in order to eliminate the impact of this difference in magnitude on the fusion result, it is necessary to first analyze each x... i After standardization, the value of the i-th factor is denoted as x. i ' Next, the weighted contribution value of each factor is calculated. The calculation method is to take the x-th factor as the weighted contribution value. i ' With the corresponding w i Multiplying these factors yields the weighted contribution value of the i-th factor. Finally, the fitness evaluation value of the path is calculated by summing the weighted contribution values ​​of all n factors. The summation result is the fitness evaluation value F of the path, which is equal to the sum of the weighted contribution values ​​of the 1st factor, the 2nd factor, and so on up to the nth factor.

[0086] Step 3.3: Based on the path fitness assessment value, combined with the environmental complexity index and the intensity of user personalized needs, a weighted average method is used to integrate the environmental complexity weight and the user personalized needs weight, and then normalization is applied to obtain the path adaptive adjustment coefficient. Specifically, this includes: analyzing the obtained path fitness assessment value; if the assessment value is high, it indicates that the current path is well adapted to the environment and the user, and the weight of the more adapted factor between the environmental complexity index and the intensity of user personalized needs will be appropriately increased; if the assessment value is low, the weight of factors that have a greater impact on adaptability needs to be adjusted; next, the specific quantitative basis of the environmental complexity index is clarified. This index is composed of the distribution density of obstacles in the structured grid map, the road... The number of path branches and the degree of terrain undulation are jointly determined; the intensity of user personalized needs is quantified by analyzing users' historical travel preference records, such as the degree of preference for natural or cultural landscapes, and active feedback during the current tour, such as the duration of attention to a certain attraction; then, a weighted average method is used to integrate the environmental complexity weight and the user personalized need weight, and the path adaptability evaluation value is used as the adjustment benchmark during the integration process to ensure that the integration result can reflect the relationship between path adaptability and environment and user needs; finally, the integrated result is normalized to limit its numerical range to a preset standard range to avoid the impact of excessively large or small values ​​on the accuracy of subsequent calculations. Through the above operations, the path adaptive control coefficient is finally obtained.

[0087] Step 3.4: Based on the path adaptive control coefficient and combined with real-time collected environmental change data, the coefficient parameters are dynamically adjusted using a sliding window algorithm to ultimately obtain an adaptive control coefficient that accurately reflects the current environmental adaptability and the user's personalized needs. Specifically, based on the obtained initial path adaptive control coefficient, the tour guide robot first sets the sliding window duration and data collection interval according to the common frequencies of environmental changes in the scenic area, such as the intervals during peak hours and the probability of temporary obstacles appearing, ensuring that the window can cover a sufficient environmental change cycle to accurately capture data fluctuations. Subsequently, the robot continuously collects environmental change data through real-time communication between its own perception module and the scenic area management system. This data includes path cycles... The system monitors real-time pedestrian density, information on temporarily appearing obstacles, and minor fluctuations in weather conditions. During this process, a sliding window moves forward sequentially at preset intervals. After each slide, the collected environmental change data within the window is statistically analyzed to determine if any changes have occurred, such as a sudden increase in pedestrian density or the addition of new obstacles. If changes occur, the initial path adaptive control coefficient is corrected based on the magnitude and direction of the change. For example, when pedestrian density increases, the coefficient corresponding to environmental complexity is appropriately increased. If the data remains unchanged, the coefficient is kept relatively stable. Through the dynamic monitoring and coefficient correction operations of the sliding window, an adaptive control coefficient that accurately reflects the current environmental adaptability and the user's personalized needs in real time is ultimately obtained.

[0088] In this embodiment of the invention, the environmental perception space around the robot is discretized at a preset resolution according to the environmental optimization path to obtain a structured grid map containing environmental feature information. Then, based on this map, the geometric feature parameters of the environmental optimization path are extracted and combined with the user's real-time status data to calculate the path fitness evaluation value through a multi-factor weighted fusion algorithm. Subsequently, based on the fitness evaluation value, combined with the environmental complexity index and the intensity of the user's personalized needs, the weights of the two are fused using a weighted average method and normalized to obtain the path adaptive control coefficient. Finally, based on the control coefficient and combined with the real-time collected environmental change data, the coefficient parameters are dynamically adjusted using a sliding window algorithm. Therefore, this method overcomes the technical problems in traditional path control, such as inaccurate environmental feature extraction due to the unstructured environmental perception space, one-sided evaluation due to the lack of consideration of multiple factors in the path fitness evaluation, poor adaptability due to the lack of integration of environmental complexity and user needs in the control coefficient, and insufficient real-time adaptability due to the fixed coefficient that cannot be dynamically adjusted with the environment. Thus, the control coefficient can balance environmental adaptability and user personalized needs, and can be dynamically optimized with environmental changes, ultimately obtaining a control coefficient that reflects the current environmental adaptability and user needs.

[0089] In a preferred embodiment of the present invention, step 3 above may include:

[0090] Step 3.5: Based on the path adaptive control coefficient, real-time acquisition of user facial images and limb movement sequence data is performed, preprocessed, and feature-annotated to construct a user behavior time-series dataset. Specifically, this includes: based on the data acquisition frequency and accuracy requirements determined by the path adaptive control coefficient, real-time capture of user facial image sequences using a high-definition camera mounted on the companion robot, and simultaneous acquisition of user limb movement sequence data such as acceleration and angular velocity using the robot's built-in motion sensors. The acquired facial images are preprocessed with illumination compensation and face alignment to eliminate environmental interference. The limb movement data is filtered, denoised, and standardized to unify the data scale. Subsequently, the preprocessed facial images are labeled with micro-expression types such as smiling and frowning, and the limb movement sequences are labeled with action types such as walking and pausing. The labeled feature data are arranged in time stamp order to construct a time-series dataset containing continuous user behavior features.

[0091] Step 3.6: Based on a multilayer perceptron structure neural network model, deep feature learning is performed on the user behavior time-series dataset to extract the temporal correlation features between micro-expression change patterns and body movements, generating a high-dimensional intent encoding vector. Specifically, this includes: First, determining the overall model architecture as a multilayer perceptron structure, containing an input layer, hidden layers, and an output layer. Information transmission between neurons is achieved through fully connected layers, thus constructing a complete neural network model. The number of neurons in the input layer matches the dimension of the feature vectors in the user behavior time-series dataset. Each neuron receives one type of feature data, such as the intensity value of micro-expressions or the acceleration value of body movements. The input layer converts the raw feature data into numerical signals that the neural network can process. The hidden layer is set to multiple levels, with the first hidden layer having more neurons than the input layer, used for preliminary nonlinear transformation of the input features. The model enhances its ability to capture non-linear features by employing the ReLU activation function, extracting fundamental correlation information such as the time interval features of micro-expression changes and the speed changes of body movements. The features output from the previous layer are further combined and abstracted using a fully connected weight matrix, focusing on learning the synchronous correlation pattern between micro-expression intensity changes and body posture transitions, such as the linkage feature of increased gait frequency when a user frowns and the deep correlation information of sustained pausing motion when smiling. The number of neurons in the output layer is set according to the dimensional requirements of the intent features, with each neuron corresponding to one dimension of the high-dimensional intent encoding vector. A linear activation function is used to output quantized feature values, ensuring that the vector fully covers the key information of the user's real-time behavioral intent. During model construction, a training strategy is determined simultaneously, using time-series user behavior data labeled with behavioral intent tags as training samples, and a loss function is selected. ,in, It is the cross-entropy loss value. It refers to the first Each user behavior intent category It is the total number of categories. It is the one-hot encoded value of the real label. The model is for the first The predicted probability of class intent measures the difference between the model output and the real intent label. The connection weights and bias parameters between each layer are iteratively adjusted through the gradient descent algorithm. After multiple rounds of training, the model error is reduced to below the preset threshold, so that the model can accurately capture the deep correlation between behavioral features. Finally, a high-dimensional intent encoding vector containing key information of the user's real-time behavioral intent is generated through the output layer. Each dimension of the vector corresponds to a quantitative expression of a behavioral feature.

[0092] Step 3.7: Perform semantic similarity matching calculation between the high-dimensional intent encoding vector and the historical preference vector retrieved from the user's historical behavior database. Output the semantic matching degree between the user's current behavioral intent and historical preferences based on the matching results. Specifically, this includes: extracting the user's past visit preferences for attraction types, average stay duration, and attention actions towards specific landscapes from the user's historical behavior database; converting the data into a historical preference vector with the same dimension as the high-dimensional intent encoding vector; comparing the high-dimensional intent encoding vector and the historical preference vector using a semantic similarity calculation method; analyzing the distance and directional consistency between the two in the feature space; and outputting a semantic matching degree value between 0 and 1 based on the calculation results. The higher the value, the higher the fit between the user's current behavioral intent and historical preferences.

[0093] Step 3.8: Based on the semantic matching degree and the information on the stopable areas of each attraction in the environmental optimization path, dynamically calculate the embedding position and recommended stay duration of the personalized stop point. Specifically, this includes: judging the correlation between the user's current behavioral intent and historical preferences based on the semantic matching degree value; when the matching degree is high, prioritizing the types of attractions that the user is interested in in historical preferences; combining the specific location coordinates, surrounding facility distribution, and space capacity of the stopable areas of each attraction in the environmental optimization path, calculating the distance between each stopable area and the user's current location, and the smoothness of the path to that area; and simultaneously referring to the user's historical stay duration at similar attractions and the fatigue level signal in the current behavioral characteristics to dynamically determine the specific embedding position and recommended stay duration of the personalized stop point.

[0094] Step 3.9: Based on the embedding results of personalized stop points, the environmental optimization path is reorganized and updated in real time to generate a dynamic path sequence that integrates the user's real-time intent and historical behavioral preferences. Specifically, this includes: rearranging the order of attractions on the original environmental optimization path according to the determined embedding positions and recommended stay durations of personalized stop points; reasonably inserting stop points into the path sequence to ensure that the total length and tour time of the path after insertion are within the user's acceptable range; deleting redundant attractions in the path that have extremely low matching degree with the user's current intent and historical preferences; and generating a dynamic path sequence that includes both the user's real-time behavioral intent and historical preferences by updating the spatial coordinate sequence and time node arrangement of the path in real time. The sequence can be continuously adjusted as user behavior changes and environmental data is updated.

[0095] In this embodiment of the invention, by collecting user facial images and limb movement sequence data in real time and constructing a behavioral time-series dataset, and combining this with a multilayer perceptron neural network to learn deep features and generate a high-dimensional intent encoding vector, and then dynamically determining personalized stopping points and reorganizing paths through semantic similarity matching calculation with historical preference vectors, this technique overcomes the technical problems in the prior art that cannot accurately capture the correlation between users' real-time behavioral intent and historical preferences, are difficult to dynamically embed stopping points that meet user needs, and result in a lack of personalization and real-time adaptability of paths. Thus, it achieves a deep fusion of users' real-time intent and historical behavioral preferences, and the generated dynamic path sequence can accurately match users' personalized needs, significantly improving the targeting and flexibility of the tour experience.

[0096] In a preferred embodiment of the present invention, step 4 above may include:

[0097] Step 4.1: Based on the preference-driven dynamic path sequence, and by real-time access to the scenic area management system's data stream on scenic spot operation status and facility availability, verify the spatiotemporal correlation constraints of each scenic spot in the path sequence to obtain the verification results. Specifically, this includes: based on the preference-driven dynamic path sequence, accessing the scenic area management system's continuously pushed data stream on scenic spot operation status and facility availability through a real-time communication link. The scenic spot operation status data stream includes the current open or closed status of each scenic spot, temporary maintenance information, and visitor capacity warnings. The facility availability data covers rest facilities, access facilities, and rain shelters involved in the path sequence. The robot checks whether the functions are normal and whether there is any temporary occupancy. Then, it verifies the spatiotemporal constraints of each attraction in the path sequence. In the time dimension, it focuses on verifying whether the planned visit time of each attraction matches the current opening hours and whether the estimated travel time between adjacent attractions is sufficient to support arrival within the opening hours. In the spatial dimension, it focuses on verifying whether the facilities on which the path connecting each attraction depends are available and whether there is a risk of passage interruption due to facility failure. Through the above dual verification in the time and spatial dimensions, the final verification result includes the availability status of each attraction and related facilities in the path sequence and the spatiotemporal constraint matching.

[0098] Step 4.2: Based on the verification results, detect unavailable attractions or facilities in the route sequence. Based on the spatiotemporal correlation constraints between multiple attractions and user preference characteristics, generate a candidate set of alternative attractions. Specifically, based on the obtained verification results, the tour guide robot first systematically reviews the results, filtering out attractions explicitly marked as unavailable in the route sequence, such as temporarily closed attractions and facilities, or damaged access facilities. It also records the specific location of unavailable elements in the route sequence and their impact on the tour; for example, the unavailability of an intermediate attraction may cause subsequent road sections relying on that attraction to become impassable. The robot then develops alternative attraction plans based on the spatiotemporal correlation constraints between multiple attractions. The selection process involves several steps. First, the selected alternative attractions must meet certain constraints: their opening hours must match the current tour schedule; their spatial distance from preceding and following attractions must be appropriate, and their travel time must be within a reasonable range; they must also form a coherent tour logic with other surrounding attractions. Second, established user preference characteristics are retrieved, such as user preferences for natural or cultural landscapes, past preferred tour durations, and the types of attractions they are interested in. Candidate attractions that initially meet the spatiotemporal constraints are then filtered based on preference matching, prioritizing attractions with a high degree of alignment with user preferences. Through this dual filtering process of spatiotemporal constraint screening and preference matching, a final set of alternative attraction candidates containing multiple alternative attractions is generated.

[0099] Step 4.3: Based on the candidate set of alternative attractions, and combining real-time environmental data and user status information, a new route plan to avoid unavailable facilities is generated through a path replanning algorithm. The comprehensive evaluation index of each alternative plan is calculated. Specifically, based on the candidate set of alternative attractions, the tour robot first obtains real-time environmental data through its own perception module in collaboration with the scenic area's meteorological system. This data includes current temperature, humidity, wind speed, rainfall probability, and crowd density information in each area. Simultaneously, through real-time communication with the user's wearable device, it obtains the user's current status information, covering the user's physiological state, such as heart rate and fatigue level, and behavioral state, such as current movement speed and willingness to stay. Subsequently, the path replanning algorithm is activated, and different attractions in the candidate set are replanned separately. Substitute the original route sequence, replace unavailable attractions, and adjust the route based on real-time environmental data to avoid unavailable facilities. For example, during rainfall, prioritize routes with rain shelters. Optimize the planned stay time at each attraction based on user status information. If user fatigue is high, appropriately extend the stay time at rest facilities near alternative attractions. This generates multiple new route plans to avoid unavailable facilities. Based on this, the robot calculates a comprehensive evaluation index for each new route plan. The index includes whether the total route time matches the user's available time, the overall environmental suitability of the route, the matching degree between alternative attractions and user preferences, and the smoothness of route passage. By quantifying and scoring each dimension and accumulating them according to preset weights, a comprehensive evaluation score is obtained for each new route plan.

[0100] Step 4.4: Based on the new path scheme, select the path with the highest comprehensive evaluation index as the final optimized route. Simultaneously, equip the final optimized route with an emergency backup plan, ultimately obtaining a multi-path optimization scheme. Specifically, based on multiple sets of new path schemes and their corresponding comprehensive evaluation indexes, the tour robot first ranks the comprehensive evaluation scores of each scheme, prioritizing the scheme with the highest score as the final optimized route. The route must meet the core requirements of reasonable total path duration, excellent environmental suitability, high user preference matching, smooth passage, and complete avoidance of unavailable facilities. Simultaneously, it must re-verify the completeness and feasibility of its spatiotemporal constraints to ensure no unavailable elements are overlooked. Subsequently, to ensure the continuity and safety of the tour process, the robot... The final optimized route is equipped with an emergency backup plan. The emergency backup plan needs to be formulated to address unforeseen circumstances that may occur during the tour, such as the temporary closure of a scenic spot on the final optimized route, sudden severe weather causing some sections of the road to be impassable, or sudden changes in the user's physiological state requiring route adjustments. Corresponding emergency backup scenic spots are preset, and highly adaptable alternative scenic spots and emergency adjustment strategies are selected from the candidate set of backup scenic spots, such as switching to rain shelter routes in case of sudden rainfall, and embedding nearby rest points when the user is fatigued. The emergency backup plan must clearly define the triggering conditions, adjustment steps, and connection methods with the original route. By determining the final optimized route and equipping it with an emergency backup plan, a multi-path optimization scheme that includes the main route and the emergency backup plan is finally formed.

[0101] In this embodiment of the invention, based on a preference-driven dynamic path sequence, real-time access is provided to the scenic area management system to analyze the operational status data stream and facility availability data of attractions within the path sequence. The spatiotemporal correlation constraints of each attraction in the path sequence are then verified. After detecting unavailable attractions or facilities based on the verification results, a candidate set of alternative attractions is generated according to the spatiotemporal correlation constraints between multiple attractions and user preference characteristics. Combining real-time environmental data and user status information, a new path scheme to avoid unavailable facilities is generated using a path replanning algorithm, and a comprehensive evaluation index for each alternative scheme is calculated. Finally, the path with the highest comprehensive evaluation index is selected as the final optimized route, and an emergency backup plan is provided to ensure optimal routing. The technology of multi-path optimization overcomes the technical problems of traditional path planning, such as the inability to identify unavailable attractions in advance due to the lack of real-time connection with scenic area operation data, the failure to combine time and space constraints and user preferences in the selection of alternative attractions, resulting in failure to meet the needs of the tour, the poor adaptability of the generated new plan due to the lack of comprehensive environmental and user status, and the easy interruption of the tour process due to the lack of emergency backup plans. It can avoid unavailable attractions and facilities in real time, ensure that alternative attractions match user preferences and conform to the tour sequence logic, and the new path plan takes into account environmental suitability and user status adaptability. At the same time, the emergency backup plan ensures that the tour process is not interrupted, thus improving the reliability and flexibility of the guide robot's path planning.

[0102] like Figure 2As shown, embodiments of the present invention also provide a reinforcement learning-driven intelligent route planning and management system for tourist attractions using a tour guide robot, comprising:

[0103] The acquisition module is used to generate an initial tour route based on user physiological data, attraction opening time constraints, and real-time visitor flow information through reinforcement learning strategies. It then dynamically adjusts and optimizes the stay time of the route based on real-time physiological monitoring signals and outputs a physiologically adapted route.

[0104] The computation module is used to optimize the environmental tolerance of the physiological adaptation path, integrate real-time meteorological data and the distribution of rain shelter facilities, and dynamically deploy three collaborative sensing nodes along the path to form an environmental monitoring domain; by calculating the spatial configuration descriptor of the monitoring domain, an environmental coordination factor is generated, and based on the environmental coordination factor, an adaptive weighted fusion of environmental assessment indicators is achieved to obtain the environmental optimization path;

[0105] The control module is used to optimize the path based on the environment. It solves the path adaptive control coefficient by discretizing the robot's perception space into a structured grid map. Based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intention feature extraction on the user's micro-expressions and body movements. The neural network learns the deep features of the temporal data through a multilayer perceptron structure and outputs an intention encoding vector. By calculating the semantic matching degree between the user's real-time behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate a preference-driven dynamic path sequence.

[0106] The processing module is used to monitor the operational status of attractions and facility data in real time based on preference-driven dynamic path sequences, perform global replanning on the sequence based on spatiotemporal correlation constraints, avoid unavailable facilities, and generate the final optimized route.

[0107] The above description represents the preferred embodiments of the present invention. It should be noted that those skilled in the art can make various improvements and modifications without departing from the principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

Claims

1. A reinforcement learning-driven intelligent route planning and management method for tourist attractions using a tour guide robot, characterized in that, The method includes: Based on user physiological data, attraction opening hours constraints, and real-time visitor flow information, an initial tour route is generated through reinforcement learning strategy. The stay time of the route is dynamically adjusted and optimized according to real-time physiological monitoring signals, and a physiologically adapted route is output. Environmental tolerance optimization is performed on the physiological adaptation path. Real-time meteorological data and the distribution of rain shelters are integrated, and three collaborative sensing nodes are dynamically deployed along the path to form an environmental monitoring domain. An environmental coordination factor is generated by calculating the spatial configuration descriptor of the monitoring domain. The spatial configuration descriptor includes the area parameter, azimuth angle parameter, and node spacing ratio parameter of the monitoring domain, quantifying the spatial geometric characteristics of the monitoring domain. The environmental coordination factor characterizes the degree of matching and coordination between the current environmental monitoring domain configuration and external environmental conditions, and is generated based on the spatial configuration descriptor and real-time meteorological data. Based on the environmental coordination factor, multiple environmental assessment indicators, including temperature, humidity, wind speed, and rainfall probability, are adaptively weighted and fused. Based on the weighted fusion result, an error compensation mechanism for the sensor system is introduced to correct the monitoring data, resulting in an environmental risk assessment result. The original path is then optimized and adjusted based on the environmental risk assessment result. Based on the environmental optimization path, the robot's perception space is discretized into a structured grid map, and the path adaptive control coefficient is solved. Based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intent feature extraction on user micro-expressions and body movements. The neural network learns deep features of temporal data through a multilayer perceptron structure and outputs an intent encoding vector. By calculating the semantic matching degree between real-time user behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate a preference-driven dynamic path sequence. Based on the path fitness evaluation value, combined with the environmental complexity index and the intensity of user personalized needs, a weighted average method is used to fuse the environmental complexity weight and the user personalized need weight, and then normalization is applied to obtain the path adaptive control coefficient. The path fitness evaluation value is based on the structured grid map, extracting geometric feature parameters of the environmental optimization path, and simultaneously acquiring real-time user state data, calculated using a multi-factor weighted fusion algorithm. Based on preference-driven dynamic path sequences, the system monitors the operational status of attractions and facility data in real time. It then performs global replanning of the sequence based on spatiotemporal correlation constraints to avoid unavailable facilities and generate the final optimized route.

2. The reinforcement learning-driven intelligent planning and management method for tourist routes of a tour guide robot according to claim 1, characterized in that, Based on user physiological data, attraction opening hours constraints, and real-time visitor flow information, an initial tour route is generated using a reinforcement learning strategy. This route is then dynamically adjusted and its dwell time optimized based on real-time physiological monitoring signals, resulting in a physiologically adapted route, including: It receives and integrates real-time physiological data uploaded by users, pre-stored scenic spot opening time constraints, and real-time scenic spot visitor flow information. It performs time synchronization and standardization processing on multi-source data, extracts its feature vectors, and fuses them to form a reinforcement learning state space. Based on the reinforcement learning state space, a decision-making reasoning process is performed through a pre-trained reinforcement learning policy network to obtain an initial tour path sequence that satisfies all time constraints and effectively avoids high-traffic areas. The system acquires users' physiological monitoring signals in real time and synchronously couples these signals with the current execution status of the initial tour path. It then dynamically adjusts the path by calculating the deviation between the actual physiological load and the predicted value. Based on the dynamic adjustment process and combined with the terrain data obtained from the geographic information database, low-lying attractions are prioritized for inclusion in the route sequence. The optimal stay duration for each attraction is calculated based on the user's real-time physiological state, resulting in a physiologically adapted route optimized in multiple dimensions such as user physiological load, time constraints, and crowd density.

3. The reinforcement learning-driven intelligent planning and management method for tourist routes of escort robots according to claim 2, characterized in that, Environmental tolerance optimization of the physiological adaptation path is performed by integrating real-time meteorological data and the distribution of rain shelters. Three collaborative sensing nodes are dynamically deployed along the path to form an environmental monitoring domain, including: Based on the physiological adaptation path, access the real-time meteorological sensor data stream; at the same time, obtain the location data of rain shelter facilities in the scenic area's geographic information system, and analyze the spatial distribution density and coverage characteristics of rain shelter facilities around the path. Based on path information and the distribution characteristics of rain shelters, combined with real-time meteorological data, the final deployment location of three collaborative sensing nodes is dynamically calculated along the periphery of the physiological adaptation path. The deployment location must meet the requirements of maximizing the monitoring coverage of the path environment and having the highest sensitivity to meteorological changes. Based on the final deployment location, the spatial layout of the three collaborative sensing nodes is dynamically adjusted to form a variable-scale environmental monitoring domain. Based on a variable-scale environmental monitoring domain, multi-physics environmental data are collected in real time. By calculating the spatial configuration parameters of the monitoring domain, an environmental coordination factor is generated, resulting in an environmental monitoring domain that adapts to changes in meteorological conditions.

4. The reinforcement learning-driven intelligent planning and management method for tourist routes of a tour guide robot according to claim 3, characterized in that, By calculating the spatial configuration descriptor of the monitoring domain to generate an environmental coordination factor, and based on the environmental coordination factor, an adaptive weighted fusion of environmental assessment indicators is achieved to obtain an environmental optimization path, including: Based on the obtained variable-scale environmental monitoring domain, the spatial configuration descriptor of the monitoring domain is calculated.

5. The reinforcement learning-driven intelligent planning and management method for tourist routes of escort robots according to claim 4, characterized in that, Based on environmental path optimization, the path adaptive control coefficients are solved by discretizing the robot's perception space into a structured grid map, including: Based on the environmental optimization path, the environmental perception space around the robot is discretized according to a preset resolution to obtain a structured grid map containing environmental feature information. Based on the path adaptive control coefficient and combined with real-time collected environmental change data, the coefficient parameters are dynamically adjusted through a sliding window algorithm to obtain an adaptive control coefficient that accurately reflects the current environmental adaptability and the user's personalized needs.

6. The reinforcement learning-driven intelligent planning and management method for tourist routes of a tour guide robot according to claim 5, characterized in that, Based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intent feature extraction on user micro-expressions and body movements. The neural network learns the deep features of temporal data through a multilayer perceptron structure and outputs an intent encoding vector. By calculating the semantic matching degree between real-time user behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate preference-driven dynamic path sequences, including: Based on the path adaptive control coefficient, real-time acquisition of users' facial images and limb movement sequence data is performed, preprocessed and labeled with features to construct a user behavior time series dataset. A neural network model based on a multilayer perceptron structure is used to perform deep feature learning on a time-series dataset of user behavior, extract the time-series correlation features between micro-expression change patterns and body movements, and generate a high-dimensional intent encoding vector. The high-dimensional intent encoding vector is matched with the historical preference vector retrieved from the user's historical behavior database to calculate semantic similarity. Based on the matching results, the semantic matching degree between the user's current behavioral intent and historical preferences is output. Based on semantic matching degree and combined with the information on the stopable areas of each attraction in the environmental optimization path, the embedding position and recommended stay duration of personalized stop locations are dynamically calculated. Based on the embedded results of personalized stopping points, the environmental optimization path is reorganized and updated in real time to generate a dynamic path sequence that integrates the user's real-time intent and historical behavioral preferences.

7. The reinforcement learning-driven intelligent planning and management method for tourist routes of a tour guide robot according to claim 6, characterized in that, Based on preference-driven dynamic path sequences, the system monitors the operational status of attractions and facility data in real time. It then performs global replanning of the sequence based on spatiotemporal correlation constraints, avoiding unavailable facilities and generating a final optimized route, including: Based on the preference-driven dynamic path sequence, and by accessing the scenic area management system in real time the data stream of scenic spot operation status and facility availability data, the spatiotemporal correlation constraints of each scenic spot in the path sequence are verified to obtain the verification results. Based on the verification results, unavailable attractions or facilities are detected in the path sequence. Based on the spatiotemporal correlation constraints between multiple attractions and user preference characteristics, a candidate set of alternative attractions is generated. Based on the candidate set of alternative attractions, combined with real-time environmental data and user status information, a new route plan that avoids unavailable facilities is generated through a route replanning algorithm, and a comprehensive evaluation index for each alternative plan is calculated. Based on the new path scheme, the path with the final comprehensive evaluation index is selected as the final optimized route, and an emergency backup plan is also provided for the final optimized route, thus obtaining a multi-path optimization scheme.

8. A reinforcement learning-driven intelligent route planning and management system for tourist attractions using a tour guide robot, wherein the system implements the method as described in any one of claims 1 to 7, characterized in that, include: The acquisition module is used to generate an initial tour route based on user physiological data, attraction opening time constraints, and real-time visitor flow information through reinforcement learning strategies. It then dynamically adjusts and optimizes the stay time of the route based on real-time physiological monitoring signals and outputs a physiologically adapted route. The computation module optimizes the environmental tolerance of the physiological adaptation path, integrates real-time meteorological data and the distribution of rain shelters, and dynamically deploys three collaborative sensing nodes along the path to form an environmental monitoring domain. It generates an environmental coordination factor by calculating the spatial configuration descriptor of the monitoring domain. The spatial configuration descriptor includes the area parameter, azimuth angle parameter, and node spacing ratio parameter of the monitoring domain, quantifying the spatial geometric characteristics of the monitoring domain. The environmental coordination factor characterizes the degree of matching and coordination between the current environmental monitoring domain configuration and external environmental conditions, and is generated based on the spatial configuration descriptor and real-time meteorological data. Based on the environmental coordination factor, it adaptively weights and fuses multiple environmental assessment indicators such as temperature, humidity, wind speed, and rainfall probability. Based on the weighted fusion result, it introduces an error compensation mechanism from the sensor system to correct the monitoring data, obtaining an environmental risk assessment result. Based on the environmental risk assessment result, it optimizes and adjusts the original path. The control module is used to optimize the path based on the environment. It discretizes the robot's perception space into a structured grid map and solves for the path adaptive control coefficient. Based on the adaptive control coefficient, a neural network algorithm is used to perform temporal pattern analysis and intent feature extraction on the user's micro-expressions and body movements. The neural network learns deep features of the temporal data through a multilayer perceptron structure and outputs an intent encoding vector. By calculating the semantic matching degree between the user's real-time behavior data and historical preference vectors, personalized stopping points are dynamically embedded online to generate a preference-driven dynamic path sequence. Specifically, based on the path fitness evaluation value, combined with the environmental complexity index and the intensity of the user's personalized needs, a weighted average method is used to fuse the environmental complexity weight and the user's personalized needs weight, and then normalization is applied to obtain the path adaptive control coefficient. The path fitness evaluation value is based on the structured grid map, extracting the geometric feature parameters of the environmental optimization path, and simultaneously acquiring the user's real-time state data, calculated using a multi-factor weighted fusion algorithm. The processing module is used to monitor the operational status of attractions and facility data in real time based on preference-driven dynamic path sequences, perform global replanning on the sequence based on spatiotemporal correlation constraints, avoid unavailable facilities, and generate the final optimized route.

9. A computing device, characterized in that, include: One or more processors; A storage device for storing one or more programs, which, when executed by one or more processors, cause the one or more processors to implement the method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program that, when executed by a processor, implements the method as described in any one of claims 1 to 7.

Images