Intelligent navigation method and system

Abstract

The invention discloses an intelligent navigation method and system, and relates to the field of intelligent robot control. The method comprises the following steps: acquiring multi-modal data acquired by a touch sensor, a visual sensor and a force sensor; if the matching degree of the tactile data and the visual data is lower than a preset threshold value, adjusting the tactile data by adopting a weighting coefficient to obtain a fusion feature vector; a three-dimensional point cloud map is generated according to the fusion feature vector, and if it is detected that the dynamic change rate of the point cloud map is higher than a preset threshold value, an incremental updating algorithm is triggered to correct the environment model; extracting a terrain feature set from the environment model, and generating a candidate path set by adopting a reinforcement learning algorithm; and if the passing cost value of the candidate path is lower than a preset threshold value, smoothing the candidate path to obtain an optimized navigation path. According to the invention, accurate navigation and path planning in a complex environment are realized, and the adaptability and robustness of a navigation system are improved.

Description

Technical Field

[0001] The present invention relates to the field of intelligent robot control, and in particular discloses an intelligent navigation method and system. Background Art

[0002] The field of robot perception and operation occupies a core position in modern automation technology. Its importance is reflected in significantly improving the task efficiency and intelligent level in agricultural, industrial, and service scenarios, and promoting the wide application of human-machine collaboration and intelligent equipment. However, the existing solutions still have deficiencies in complex unstructured environments, generally suffering from low perception accuracy, poor environmental adaptability, and low efficiency in integrating multi-source data.

[0003] These limitations make it difficult for robots to achieve accurate target recognition and flexible operation in dynamic scenarios, restricting their application potential in actual scenarios.

[0004] Specifically, this field faces three core challenges: multi-modal perception fusion, dynamic environment modeling, and adaptive control.

[0005] Firstly, multi-modal perception fusion involves the real-time integration of data from various sensors such as tactile, visual, and force sensors. Existing technologies are difficult to maintain perception stability in complex deformation or visionless environments.

[0006] Secondly, dynamic environment modeling requires generating a real-time environmental map through 3D reconstruction and path planning technologies, but current methods lack robustness in rapidly changing scenarios.

[0007] Finally, adaptive control requires the robot to dynamically adjust the operation force and path according to environmental feedback. However, existing algorithms have limited flexibility in dealing with diverse targets and terrains.

[0008] These unresolved technical factors make it difficult for robots to achieve efficient and stable task execution in agricultural inspections, industrial sorting, or hazardous environment operations, forming a unique technical problem.

[0009] Therefore, how to design a bionic sensor system integrating multi-modal perception, dynamic modeling, and adaptive control to achieve accurate target recognition and flexible operation in complex environments has become a key problem that needs to be solved urgently in this research. Summary of the Invention

[0010] The present invention provides an intelligent navigation method and system, aiming to solve at least one of the defects existing in the above-mentioned prior art.

[0011] One aspect of the present invention relates to an intelligent navigation method, including the following steps:

[0012] Obtain multi-modal data collected by tactile sensors, visual sensors, and force sensors, where the multi-modal data includes tactile data, visual data, and force data;

[0013] Judge whether the matching degree between the tactile data and the visual data is lower than a preset threshold. If so, adjust the force data using a weighting coefficient to obtain a fused feature vector;

[0014] Generate a three-dimensional point cloud map based on the fused feature vector. If the detected dynamic change rate of the three-dimensional point cloud map is higher than a preset threshold, trigger an incremental update algorithm to correct the environmental model;

[0015] Extract a terrain feature set from the environmental model and use a reinforcement learning algorithm to generate a set of candidate paths;

[0016] Judge whether the passing cost value of the candidate path is lower than a preset threshold. If so, smooth the candidate path to obtain an optimized navigation path;

[0017] Generate control parameters using a proportional-integral-derivative algorithm based on the optimized navigation path and force feedback data.

[0018] Furthermore, the steps of obtaining multi-modal data collected by tactile sensors, visual sensors, and force sensors include:

[0019] Obtain raw data from tactile sensors, visual sensors, and force sensors, calibrate the raw data through timestamp alignment and unit standardization, and filter out noise using a preset threshold to obtain a calibrated multi-modal data set. The raw data includes raw tactile data, raw visual data, and raw force data;

[0020] For the calibrated multi-modal data set, use a feature extraction algorithm to extract the pressure distribution feature of the tactile data, the edge contour feature of the visual data, and the moment feature of the force data, and map them to a unified feature space through data fusion to obtain a fused feature vector;

[0021] If the dimension of the fused feature vector exceeds a preset threshold, perform dimensionality reduction processing using the principal component analysis algorithm, and use a pattern recognition algorithm to judge the environmental state based on the dimensionality-reduced feature vector to determine the environmental perception result;

[0022] Match the environmental perception result with a pre-established action control mapping table to obtain an action control instruction, and combine real-time anomaly detection to judge whether the instruction triggers an anomaly threshold to obtain action control parameters.

[0023] Furthermore, the steps of judging whether the matching degree between the tactile data and the visual data is lower than a preset threshold. If so, adjusting the force data using a weighting coefficient to obtain a fused feature vector include:

[0024] Obtain tactile data and visual data from tactile sensors and visual sensors, generate a tactile feature set and a visual feature set through feature extraction, use the cosine similarity algorithm to calculate the matching degree between the tactile feature set and the visual feature set, and obtain a matching degree value;

[0025] Judge whether the matching degree value is lower than a preset threshold. If so, obtain a weighting coefficient according to a preset weighting coefficient table, and perform weighted adjustment on the force sense data through matrix multiplication to obtain the adjusted force sense data;

[0026] Stitch the adjusted force sense data with the visual feature set through a data fusion algorithm, and perform dimensionality reduction processing on the stitched data using the principal component analysis algorithm to obtain a preliminary fusion feature vector;

[0027] According to the dimensionality requirements of the preliminary fusion feature vector, normalize the elements of the preliminary fusion feature vector through normalization processing to obtain the final fusion feature vector.

[0028] Further, the steps of generating a three-dimensional point cloud map based on the fusion feature vector and triggering an incremental update algorithm to correct the environmental model if the detected dynamic change rate of the three-dimensional point cloud map is higher than a preset threshold include:

[0029] Obtain sensor data from a lidar and a camera, generate a first feature set through a feature extraction algorithm, and fuse them into a first fusion feature vector using a data integration method to obtain a three-dimensional point cloud map;

[0030] Perform change detection on the three-dimensional point cloud map and the second point cloud map at the previous moment, calculate the first dynamic change rate, and judge whether the first dynamic change rate is higher than a preset threshold;

[0031] If the first dynamic change rate is higher than a preset threshold, perform local adjustment on the three-dimensional point cloud map through an incremental update algorithm to obtain a first map correction result;

[0032] Update the environmental model according to the first map correction result, and adjust the environmental model parameters through a model optimization method to obtain an optimized first environmental model.

[0033] Further, the steps of extracting a terrain feature set from the environmental model and generating a candidate path set using a reinforcement learning algorithm include:

[0034] Obtain spatial information from environmental data, and decompose it into height, slope, and curvature through rasterization processing to obtain a terrain feature set;

[0035] According to the terrain feature set, use a reinforcement learning algorithm, combined with a preset reward function and path constraints, to generate a candidate path set;

[0036] If there are paths in the candidate path set that do not meet the path length or obstacle avoidance requirements, recalculate the path constraints based on the spatial information to obtain an optimized path set;

[0037] From the optimized path set, obtain the height, slope, and curvature information of each path, and determine whether it meets the preset threshold to obtain the final path set.

[0038] Further, the steps of determining whether the passage cost value of the candidate path is lower than the preset threshold, and if so, smoothing the candidate path to obtain an optimized navigation path include:

[0039] Obtain a candidate path set from the navigation target, divide the path segments using node coordinates, calculate the passage cost value of each candidate path based on the path length and traffic conditions to obtain a passage cost data set;

[0040] Determine whether the passage cost value in the passage cost data set is lower than the preset threshold. If so, extract the corresponding candidate path and verify the effectiveness of the candidate path using data accuracy to obtain a path set to be smoothed;

[0041] Adjust the node coordinates of the path set to be smoothed through smoothing processing to obtain an optimized path segment and obtain a smoothed path data set;

[0042] Based on the smoothed path data set and combined with the real-time updated traffic conditions, determine whether the path meets the navigation target to determine the optimized navigation path.

[0043] Further, the steps of generating control parameters using a proportional-integral-derivative algorithm based on the optimized navigation path and force feedback data include:

[0044] Obtain preset navigation path data from the path planning module, obtain real-time environment data from the environment perception module, and integrate the navigation path data and environment data using a data fusion method to obtain fused path feature data;

[0045] Based on the path feature data and the real-time force feedback data collected by the force feedback module, calculate the current error value and the cumulative historical error value using a proportional-integral-derivative algorithm to determine the adjusted control parameters;

[0046] If the control parameter exceeds the preset threshold, smooth the control parameter through the dynamic response module to obtain optimized control instruction data;

[0047] Convert the control instruction data into an execution signal of the target device through the instruction output module to obtain the final navigation control signal.

[0048] Another aspect of the present invention relates to an intelligent navigation system for implementing the above intelligent navigation method. The intelligent navigation system includes:

[0049] A first acquisition module, configured to acquire multi-modal data collected by a tactile sensor, a visual sensor, and a force sensor, where the multi-modal data includes tactile data, visual data, and force data;

[0050] A second acquisition module, configured to determine that the matching degree between the tactile data and the visual data is lower than a preset threshold. If so, adjust the force data by using a weighting coefficient to obtain a fused feature vector;

[0051] A correction module, configured to generate a three-dimensional point cloud map according to the fused feature vector. If it is detected that the dynamic change rate of the point cloud map is higher than a preset threshold, trigger an incremental update algorithm to correct the environmental model;

[0052] A first generation module, configured to extract a terrain feature set from the environmental model and generate a candidate path set by using a reinforcement learning algorithm;

[0053] A third acquisition module, configured to determine whether the passing cost value of the candidate path is lower than a preset threshold. If so, perform smoothing processing on the candidate path to obtain an optimized navigation path;

[0054] A second generation module, configured to generate control parameters by using a proportional-integral-derivative algorithm according to the optimized navigation path and force feedback data.

[0055] Further, the first acquisition module includes:

[0056] A first acquisition unit, configured to acquire raw data from the tactile sensor, the visual sensor, and the force sensor, calibrate the raw data through timestamp alignment and unit standardization, and filter noise by using a preset threshold to obtain a calibrated multi-modal data set, where the raw data includes raw tactile data, raw visual data, and raw force data;

[0057] A second acquisition unit, configured to, for the calibrated multi-modal data set, extract the pressure distribution feature of the tactile data, the edge contour feature of the visual data, and the torque feature of the force data by using a feature extraction algorithm, and map them to a unified feature space through data fusion to obtain a fused feature vector;

[0058] A determination unit, configured to determine whether the dimension of the fused feature vector exceeds a preset threshold. If so, perform dimensionality reduction processing by using a principal component analysis algorithm, and determine the environmental state according to the dimensionality-reduced feature vector by using a pattern recognition algorithm to determine the environmental perception result;

[0059] A third acquisition unit, configured to match the environmental perception result with a pre-established action control mapping table to obtain an action control instruction, and combine real-time anomaly detection to determine whether the instruction triggers an anomaly threshold to obtain action control parameters.

[0060] Further, the second acquisition module includes:

[0061] A fourth acquisition unit, configured to acquire tactile data and visual data from a tactile sensor and a visual sensor, generate a tactile feature set and a visual feature set through feature extraction, and calculate the matching degree between the tactile feature set and the visual feature set by using a cosine similarity algorithm to obtain a matching degree value;

[0062] A fifth acquisition unit, configured to determine whether the matching degree value is lower than a preset threshold. If so, obtain a weighting coefficient according to a preset weighting coefficient table, and perform weighted adjustment on the force sense data through matrix multiplication to obtain adjusted force sense data;

[0063] A sixth acquisition unit, configured to splice the adjusted force sense data and the visual feature set through a data fusion algorithm, and perform dimensionality reduction processing on the spliced data by using a principal component analysis algorithm to obtain a preliminary fusion feature vector;

[0064] A seventh acquisition unit, configured to normalize the elements of the preliminary fusion feature vector through normalization processing according to the dimensionality requirement of the preliminary fusion feature vector to obtain a final fusion feature vector.

[0065] The beneficial effects achieved by the present invention are as follows:

[0066] The present invention provides an intelligent navigation method and system. By acquiring multi-modal data of tactile, visual, and force sense sensors, performing matching degree analysis and weighted adjustment on the data to generate a fusion feature vector; constructing a three-dimensional point cloud map according to the feature vector, and dynamically correcting the environment model by using an incremental update algorithm; extracting terrain features from the environment model, generating a low-cost candidate path by using a reinforcement learning algorithm, and performing smoothing processing to obtain an optimized navigation path; finally, combining force sense feedback and using a PID algorithm to generate control parameters. The intelligent navigation method and system provided by the present invention achieve precise navigation and path planning in a complex environment, improve the adaptability and robustness of the navigation system, and can be widely applied to fields such as autonomous mobile robots and driverless vehicles. BRIEF DESCRIPTION OF THE DRAWINGS

[0067] Figure 1 It is a schematic flowchart of an embodiment of an intelligent navigation method of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0068] In order to better understand the above technical solutions, the above technical solutions will be described in detail below in conjunction with the accompanying drawings of the specification and specific embodiments.

[0069] As Figure 1 shown, a first embodiment of the present invention proposes an intelligent navigation method, including the following steps:

[0070] Step S100: Obtain multi-modal data collected by tactile sensors, visual sensors, and force sensors. The multi-modal data includes tactile data, visual data, and force data.

[0071] Obtain the tactile data collected by the tactile sensor, the visual data collected by the visual sensor, and the force data collected by the force sensor.

[0072] Step S200: Determine whether the matching degree between the tactile data and the visual data is lower than a preset threshold. If so, adjust the force data using a weighting coefficient to obtain a fused feature vector.

[0073] The fused feature vector refers to the process of integrating feature vectors from multiple sources or modalities into a unified representation through a specific method, aiming to improve the data representation ability and model performance. In this embodiment, the force data is adjusted using a weighting coefficient to obtain a fused feature vector.

[0074] The weighting coefficient is used in multi-sensor data fusion to dynamically adjust the contribution degree of the force data, so as to optimize the perception accuracy and robustness of the system for physical interactions.

[0075] Step S300: Generate a three-dimensional point cloud map based on the fused feature vector. If it is detected that the dynamic change rate of the three-dimensional point cloud map is higher than a preset threshold, trigger an incremental update algorithm to correct the environmental model.

[0076] The dynamic change rate of the three-dimensional point cloud map is used to describe the degree of dynamic attribute changes of the point cloud data in the time or space dimension.

[0077] The incremental update algorithm is a calculation method that only performs local updates on the changed parts of the environment or data, avoiding global repeated calculations or storage, so as to improve efficiency and real-time performance.

[0078] The environmental model is a mathematical or physical representation tool that abstractly and formally describes the structure, function, or dynamic changes of the environmental system.

[0079] Step S400: Extract a terrain feature set from the environmental model and generate a candidate path set using a reinforcement learning algorithm.

[0080] The terrain feature set is a comprehensive description of the terrain elements and their spatial relationships in a certain area, covering core attributes such as terrain type, terrain undulation, landform structure, and distribution law.

[0081] The reinforcement learning algorithm is a class of machine learning methods that aim to maximize the cumulative reward through the interaction and trial-and-error of an agent with the environment, and its core revolves around the state, action, policy, and reward mechanism.

[0082] The candidate path set is a set of feasible paths that meet the constraint conditions generated by an algorithm during the path planning process, and is used for subsequent optimal path screening and trajectory optimization. The candidate path set is a set of multiple potential feasible paths generated by a path planning algorithm when a robot or an unmanned aerial vehicle performs a task in a specific environment.

[0083] Step S500: Determine whether the passing cost value of the candidate path is lower than a preset threshold. If so, perform smoothing processing on the candidate path to obtain an optimized navigation path.

[0084] The smoothing processing of the candidate path aims to optimize the discrete path points generated by path planning, eliminate sharp turns or discontinuities, and make it meet the kinematic constraints and trajectory continuity requirements of the robot or the unmanned aerial vehicle. The smoothing processing method can adopt the interpolation method and the filtering and averaging method.

[0085] The optimized navigation path refers to the process of adjusting or re-planning the initial path through an algorithm on the premise of meeting the kinematic constraints and environmental constraints of the robot, the unmanned aerial vehicle or the autonomous vehicle, so that it reaches a better state in performance indicators such as path length, smoothness, and safety. Its core goal is to generate an efficient, safe and executable motion trajectory through mathematical modeling and algorithm design.

[0086] Step S600: Generate control parameters according to the optimized navigation path and the force feedback data by using the proportional-integral-derivative algorithm.

[0087] The proportional-integral-derivative algorithm (PID for short) is a closed-loop control algorithm based on error feedback. By dynamically adjusting the system input quantity, the deviation between the actual output and the expected target is minimized. Its core idea is to realize fast response, suppress steady-state error and enhance system stability through the linear combination of the proportional term, the integral term and the differential term.

[0088] Control parameters are adjustable variables that affect the dynamic behavior of the system or the performance of the algorithm. By adjusting these parameters, the stability, convergence speed and accuracy of the system output or the learning process can be optimized. The core role of the control parameters is to balance the system response characteristics and the constraint conditions to achieve precise control of the target behavior. In this embodiment, the role of the control parameters is to optimize the navigation of the robot. By adjusting the robot motion response characteristics, the environmental perception strategy and the decision-making logic, the comprehensive optimization of the navigation efficiency, the obstacle avoidance accuracy and the dynamic adaptability is realized.

[0089] Furthermore, for the intelligent navigation method provided in this embodiment, step S100 includes:

[0090] Step S110: Obtain the raw data from the tactile sensor, visual sensor, and force sensor. Calibrate the raw data through timestamp alignment and unit standardization, and filter out the noise using a preset threshold to obtain the calibrated multi-modal data set. The raw data includes raw tactile data, raw visual data, and raw force data.

[0091] For the multi-modal perception and motion control of tactile, visual, and force data, the core lies in achieving precise environment perception and generating reliable control instructions through multi-sensor data fusion.

[0092] The following analyzes and gives examples around the technical theme, with the scenario limited to a robot grasping an object.

[0093] Exemplarily, the tactile sensor obtains the raw data, reflecting the pressure distribution on the object's surface. Suppose the robot grasps an apple, and the tactile sensor records the pressure values at the contact points, with the unit of Pascal.

[0094] During calibration, ensure the synchronization of tactile data with visual and force data through timestamp alignment, and normalize the pressure values to the range of 0 to 1 through unit standardization.

[0095] Filter out the noise using the preset threshold, and remove the outliers below 0.01 to obtain the calibrated pressure data set. This step ensures data consistency and improves the reliability of subsequent feature extraction.

[0096] In a possible implementation, the visual sensor captures the edge information of the object.

[0097] The robot obtains the image data of the apple through the camera, and after calibration, uses the edge detection algorithm to extract the contour features, such as the circular boundary of the apple.

[0098] Noise filtering removes the light interference and retains the clear edge point set. The edge contour features reflect the shape of the object and provide a geometric basis for fusion. It should be noted that the accuracy of visual data directly affects the integrity of the fusion features.

[0099] Specifically, the force sensor measures the torque data during grasping, with the unit of Newton-meter. After calibration, extract the torque features, such as the direction and magnitude of the torque when grasping the apple, which reflect the grasping force distribution. Use the preset threshold to remove the abnormal torques caused by jitter to obtain the stable features. This step provides dynamic information for fusion and ensures the stability of motion control.

[0100] Step S120: For the calibrated multi-modal data set, use the feature extraction algorithm to extract the pressure distribution features of the tactile data, the edge contour features of the visual data, and the torque features of the force data, and map them to a unified feature space through data fusion to obtain the fusion feature vector.

[0101] Preferably, after feature extraction, the pressure distribution features of touch, the edge contour features of vision, and the moment features of force sense are mapped to a unified feature space through data fusion.

[0102] Assume that the pressure distribution is a two-dimensional vector, the contour is a set of point coordinates, and the moment is a three-dimensional vector. After fusion, a high-dimensional feature vector is formed. The fusion algorithm can be based on weighted average, combining the three types of features with weights of 0.4, 0.3, and 0.3 to generate a fused feature vector. This step integrates multi-modal information and improves the comprehensiveness of environmental perception.

[0103] Step S130: If the dimension of the fused feature vector exceeds a preset threshold, the principal component analysis algorithm is used for dimensionality reduction processing. Based on the dimensionality-reduced feature vector, the environmental state is judged using a pattern recognition algorithm to determine the environmental perception result.

[0104] In one embodiment, if the dimension of the fused feature vector exceeds the threshold, such as 50 dimensions, principal component analysis dimensionality reduction is used. The first 10 principal components are retained, covering 90% of the variance, reducing the computational complexity.

[0105] The dimensionality-reduced feature vector is used for pattern recognition to judge the environmental state. For example, to identify whether an apple is soft or hard, and output the state of "soft object". Dimensionality reduction and recognition ensure high computational efficiency and accurate results.

[0106] Step S140: By matching the environmental perception result with a pre-established action control mapping table, an action control instruction is obtained, and combined with real-time anomaly detection to judge whether the instruction triggers an anomaly threshold, and an action control parameter is obtained.

[0107] It can be understood that the environmental perception result is matched with the action control mapping table to generate a control instruction.

[0108] Assume that the mapping table stipulates that "soft object" corresponds to a grasping force of 5 Newtons, and the corresponding instruction is output after the perception result is matched.

[0109] Real-time anomaly detection checks whether the instruction exceeds a safety threshold, such as 10 Newtons. If the anomaly is not triggered, the instruction is converted into an action parameter to control the robot to grasp with a force of 5 Newtons. This step ensures the safety of the action and its adaptation to the environment. For example, if the anomaly detection finds that the grasping force suddenly increases to 12 Newtons and triggers the threshold, the system adjusts to pause the grasping and re-perceive to avoid damaging the apple or the robot. Anomaly detection improves the robustness of the system and ensures the safety of the operation.

[0110] Through the above implementation, multi-modal data fusion and dimensionality reduction optimize the computational efficiency, pattern recognition improves the perception accuracy, and anomaly detection ensures the reliability of the action.

[0111] The overall scheme achieves precise control in the robot grasping scenario, reduces the risk of misoperation, and improves the task success rate.

[0112] Preferably, for the intelligent navigation method provided in this embodiment, step S200 includes:

[0113] Step S210: Obtain tactile data and visual data from a tactile sensor and a visual sensor, generate a tactile feature set and a visual feature set through feature extraction, and calculate the matching degree between the tactile feature set and the visual feature set using the cosine similarity algorithm to obtain a matching degree value.

[0114] In a possible implementation manner, the process of obtaining data from the tactile sensor and the visual sensor can be understood as the starting point of multimodal information acquisition.

[0115] Tactile sensors are usually used to detect the pressure distribution on the surface of an object, while visual sensors are responsible for capturing image information in the environment. For example, in a scenario where a robot grasps an object, the tactile sensor can sense the pressure change when the finger touches the object, while the visual sensor records the shape and position of the object.

[0116] Preferably, the tactile sensor may collect 100 groups of pressure data per second, and each group of data contains the pressure values of multiple points; the visual sensor may obtain images with a resolution of 1280x720 at a speed of 30 frames per second. This acquisition method can provide rich basic data for subsequent feature extraction.

[0117] Specifically, feature extraction is the process of converting raw data into analyzable features. For tactile data, the peak region and change trend of the pressure distribution can be extracted. For example, when grasping an object with uneven hardness, the tactile feature set may include a pressure peak of 5 Newtons in a certain area and a change frequency of 2 times per second. For visual data, the edge contour and key point positions of the object can be extracted. For example, the visual feature set may include an aspect ratio of the object of 2:1 and an edge curvature radius of 10 centimeters. This feature extraction method can simplify complex raw data into more easily processable information.

[0118] Exemplarily, the cosine similarity algorithm is used to evaluate the correlation between the tactile feature set and the visual feature set. It should be noted that the cosine similarity judges the similarity degree by calculating the cosine value of the included angle between two feature vectors. For example, the tactile feature set may be a vector containing pressure peaks and frequencies, while the visual feature set contains edge curvatures and aspect ratios.

[0119] If the calculation result shows that the cosine value is 0.6, which is lower than the preset threshold of 0.8, it indicates that the correlation between the two groups of features is insufficient. This evaluation method can quickly judge whether multimodal data is consistent.

[0120] Step S220: Determine whether the matching degree value is lower than a preset threshold. If so, obtain a weighting coefficient according to a preset weighting coefficient table, and perform weighted adjustment on the force perception data through matrix multiplication to obtain adjusted force perception data.

[0121] In one embodiment, if the matching degree is lower than the threshold, the force perception data is adjusted through a weighting coefficient table.

[0122] It can be understood that force perception data usually reflects the force feedback when a robot executes an action. For example, a force perception sensor may detect that the force during grasping is 10 Newtons and the torque is 0.5 Newton-meter. The weighting coefficient table may stipulate that the weighting coefficient for force is 0.7 and the coefficient for torque is 0.3. After matrix multiplication, the adjusted force perception data may become 7 Newtons and 0.15 Newton-meter. This adjustment method can balance the influence of different modality data.

[0123] Step S230: Concatenate the adjusted force perception data with the visual feature set through a data fusion algorithm, and perform dimensionality reduction processing on the concatenated data using a principal component analysis algorithm to obtain a preliminary fusion feature vector.

[0124] For example, the data fusion algorithm concatenates the adjusted force perception data with the visual feature set into a unified vector.

[0125] Preferably, during the concatenation process, the 7 Newtons of the force perception data may be combined with the aspect ratio of 2:1 of the visual feature into a multi-dimensional vector. This concatenation can integrate multi-source information.

[0126] Immediately afterwards, the principal component analysis algorithm performs dimensionality reduction processing on the concatenated data. For example, the original vector may contain 20 dimensions, and after dimensionality reduction, 5 main dimensions are retained. This dimensionality reduction method can reduce the computational complexity.

[0127] Step S240: According to the dimensionality requirements of the preliminary fusion feature vector, normalize the elements of the preliminary fusion feature vector through normalization processing to obtain the final fusion feature vector.

[0128] In one embodiment, normalization processing normalizes the elements of the preliminary fusion feature vector. For example, the element value range of the preliminary feature vector may be from 0 to 100, and after normalization, it is adjusted to the range from 0 to 1. For example, 7 Newtons is normalized to 0.07. This normalization method can ensure that features with different dimensions are compared on a unified scale, and finally obtain a fusion feature vector suitable for subsequent processing. This processing method provides a basis for the efficient utilization of multi-modal data.

[0129] Furthermore, for the intelligent navigation method provided in this embodiment, step S300 includes:

[0130] Step S310: Obtain sensor data from the lidar and camera, generate a first feature set through a feature extraction algorithm, and fuse it into a first fused feature vector using a data integration method to obtain a 3D point cloud map.

[0131] Exemplarily, when obtaining sensor data from the lidar and camera, it usually involves multi-modal data acquisition. For example, the lidar generates high-precision 3D point cloud data by emitting laser pulses and receiving reflected signals, while the camera captures 2D RGB image data. In a possible implementation, the lidar may generate point cloud data containing 100,000 points at a frequency of 20 frames per second, and the camera synchronously acquires images at 1080p resolution. This method ensures the alignment of the two types of sensor data in time and space, laying the foundation for subsequent feature extraction.

[0132] Specifically, when the feature extraction algorithm generates the first feature set, it can be achieved through point cloud segmentation and image edge detection. For example, the lidar point cloud data can extract geometric features of the object surface, such as planes or curved surfaces, through the Euclidean clustering algorithm; the camera image data extracts contour features of the object through the Canny edge detection algorithm. It should be noted that the point cloud features may contain the depth information of the object, while the image features provide color and texture information. These two types of feature sets are complementary at the semantic level, providing a rich information source for subsequent fusion.

[0133] In one embodiment, when the data integration method fuses the first feature set into the first fused feature vector, it can adopt a deep learning-based feature alignment technique. For example, through a pre-trained neural network, the point cloud features and image features are mapped to a unified feature space, and then a fused feature vector is generated through a concatenation operation. This fused vector can express both the geometric shape and visual appearance of the object.

[0134] Preferably, the fused feature vector can retain the main information through dimensionality reduction processing, thereby generating a 3D point cloud map. For example, a scene containing 100,000 point cloud points may be compressed into a map containing 1,000 key feature points.

[0135] Step S320: Perform change detection on the 3D point cloud map and the second point cloud map at the previous moment, calculate the first dynamic change rate, and determine whether the first dynamic change rate is higher than a preset threshold.

[0136] It can be understood that when performing change detection on the 3D point cloud map and the second point cloud map at the previous moment, it is mainly to identify dynamic objects in the scene. For example, in an indoor navigation scenario, the point cloud map at the previous moment shows a stationary table, while the current map detects a new moving object on the table, such as a robot.

[0137] In one embodiment, change detection can be achieved through point cloud difference analysis. Calculate the point cloud overlap rate between two frames of maps to obtain the first dynamic change rate. For example, the change rate may be 20%, indicating significant dynamic changes in the scene.

[0138] Step S330: If the first dynamic change rate is higher than the preset threshold, locally adjust the three-dimensional point cloud map through an incremental update algorithm to obtain the first map correction result.

[0139] For example, if the first dynamic change rate is higher than the preset threshold, such as 10%, locally adjust the three-dimensional point cloud map through an incremental update algorithm.

[0140] In a possible implementation, incremental update only recalculates the point cloud data in the changed area instead of reconstructing the map as a whole. For example, only update the point cloud data around the table and retain the original map information of other static areas. This method significantly improves the efficiency of map update.

[0141] Step S340: Update the environmental model according to the first map correction result, and adjust the environmental model parameters through a model optimization method to obtain the optimized first environmental model.

[0142] Specifically, when the first map correction result is used to update the environmental model, it can be achieved through a probability grid model. For example, convert the corrected point cloud data into an occupancy grid and mark each grid as occupied or free. This model can clearly express the dynamic changes of the environment.

[0143] Preferably, when the model optimization method adjusts the environmental model parameters, it can smooth the map boundary through Bayesian filtering. For example, the optimized model may adjust the edge of the table from a blurred point cloud cluster to a clear geometric boundary, thereby improving the accuracy of the environmental model.

[0144] In one embodiment, the optimized first environmental model provides a reliable basis for subsequent navigation tasks. For example, a robot can plan a path to avoid dynamic objects based on this model. The dynamic update ability of this model ensures the adaptability of the robot in a complex environment.

[0145] It should be noted that the above method significantly improves the robustness of environmental perception through multi-modal data fusion and dynamic adjustment, providing support for real-time decision-making.

[0146] Preferably, for the intelligent navigation method provided in this embodiment, step S400 includes:

[0147] Step S410: Obtain spatial information from environmental data, and decompose it into height, slope, and curvature through rasterization processing to obtain a terrain feature set.

[0148] Exemplarily, obtaining spatial information from environmental data usually relies on the collaborative work of multiple sensors.

[0149] LiDAR can scan the surrounding environment to generate high-precision three-dimensional point cloud data, while the camera provides texture and color information. After preprocessing, these data can reflect key information such as the height and slope of the terrain. For example, in a field exploration scenario, LiDAR generates 100,000 point cloud data points per second, covering the terrain features within a range of 100 meters, and the camera assists in capturing details of the terrain surface, such as rock distribution or vegetation coverage.

[0150] It should be noted that the quality of spatial information directly affects the accuracy of subsequent analysis, so multi-source data is usually combined for calibration.

[0151] In a possible implementation, rasterization is a core step in decomposing spatial information. The spatial data is divided into grids of a fixed size, for example, each grid is 1 meter × 1 meter, and the average height, slope angle, and curvature change within each grid are calculated.

[0152] Height reflects the terrain undulation, slope determines the difficulty of passage, and curvature indicates the terrain smoothness. For example, the height difference within a grid is 0.5 meters, the slope is 15 degrees, and the curvature is low, indicating that the terrain here is relatively flat and suitable for path planning.

[0153] Preferably, rasterization can also dynamically adjust the grid size to adapt to the complexity of different terrains.

[0154] Step S420: According to the terrain feature set, use a reinforcement learning algorithm, combined with a preset reward function and path constraints, to generate a set of candidate paths.

[0155] Specifically, using a reinforcement learning algorithm to generate a set of candidate paths based on the terrain feature set, the core lies in guiding the algorithm to select better paths through the reward function.

[0156] The reward function may include factors such as path length, slope size, and obstacle distance. For example, the algorithm sets that paths with a slope less than 10 degrees in the reward function get positive scores, and paths with an obstacle distance less than 1 meter are deducted points.

[0157] In a scenario, after the reinforcement learning model iterates 1000 times, 5 candidate paths are generated, and the path lengths are between 50 meters and 70 meters.

[0158] It should be noted that reinforcement learning can dynamically adapt to environmental changes and continuously optimize path selection.

[0159] Step S430: If there are paths in the candidate path set that do not meet the path length or obstacle avoidance requirements, recalculate the path constraints based on the spatial information to obtain an optimized path set.

[0160] In one embodiment, if there are paths in the candidate path set that do not meet the conditions, such as a path length exceeding 80 meters or being close to an obstacle, the system will recalculate the path constraints according to the spatial information. For example, re-analyze the lidar data, identify the obstacle boundaries, adjust the path to avoid dangerous areas, and finally generate an optimized path set. The number of paths may be reduced to 3, but each path meets the safety constraints.

[0161] It can be understood that this recalculation can effectively improve the practicality of the path.

[0162] Step S440: Through the optimized path set, obtain the height, slope, and curvature information of each path, and determine whether it meets the preset thresholds to obtain the final path set.

[0163] For example, after obtaining the optimized path set, the system will analyze the height, slope, and curvature information of each path one by one to determine whether it meets the preset thresholds.

[0164] Assume that the thresholds are set as the slope is less than 20 degrees, the height change is less than 1 meter, and the curvature change is smooth. The slope of one path is 18 degrees, the height change is 0.8 meters, and the curvature is stable, meeting the conditions; the slope of another path is 25 degrees, so it is excluded. The final path set may only retain 2 paths.

[0165] In a possible implementation, the system will also record the specific parameters of each path for subsequent dynamic adjustment.

[0166] Preferably, the generation of the final path set is not an isolated step, but is closely related to the continuous update of environmental data. For example, during the movement of the detection device, new spatial information is collected in real time, the terrain feature set is updated, and the path set will also be optimized accordingly. This method can ensure that the path always adapts to environmental changes and improve the reliability of the planning.

[0167] Furthermore, for the intelligent navigation method provided in this embodiment, step S500 includes:

[0168] Step S510: Obtain a candidate path set from the navigation target, divide the path segments by node coordinates, and calculate the passing cost value of each candidate path according to the path length and traffic conditions to obtain a passing cost data set.

[0169] In a possible implementation, the process of obtaining a set of candidate paths from a navigation target mainly lies in generating multiple feasible paths based on the spatial relationship between the target point and the starting point, combined with map data. For example, assuming the navigation target is from the city center point A to the suburban point B, the system extracts the road network through map data and generates three candidate paths: the main road straight path, the detour through secondary roads path, and the mixed path.

[0170] The main road path is shorter in distance but may be congested. The detour path is slightly longer in distance but has a lower traffic flow. The mixed path takes both into account. This method covers the possibilities of different traffic conditions through multi-path generation.

[0171] Specifically, when dividing path segments using node coordinates, each path can be decomposed into segments connected by multiple nodes. For example, the main road path may contain 10 key nodes, each node corresponding to an intersection or a turning point, and the path segment is the connection line between adjacent nodes. After division, the system calculates the travel cost based on the length of each segment and the real-time traffic conditions.

[0172] Suppose a certain segment of the main road is 2 kilometers long and the traffic congestion index is 0.8. The cost value can be calculated through weighted calculation, considering both time and distance comprehensively. A certain segment of the detour path is 3 kilometers long, but the congestion index is only 0.2, and the cost value may be lower. This calculation method based on node division can refine path analysis.

[0173] Step S520: Determine whether the travel cost value in the travel cost data set is lower than a preset threshold. If so, extract the corresponding candidate path, and verify the validity of the candidate path using data accuracy to obtain a set of paths to be smoothed.

[0174] It should be noted that the threshold screening of the travel cost data set aims to eliminate inefficient paths. For example, the threshold is set to 100, and only paths with cost values lower than this enter the subsequent verification. Suppose the cost of the main road path is 120, the detour path is 80, and the mixed path is 90, then the main road path is eliminated.

[0175] Data accuracy verification further ensures the validity of the path. For example, verify whether the node coordinates of the detour path are consistent with the map data. If it is found that the deviation of a certain node exceeds 50 meters, readjust it to ensure the feasibility of the path.

[0176] Step S530: Adjust the node coordinates of the set of paths to be smoothed through smoothing processing to obtain optimized path segments and obtain a smoothed path data set.

[0177] Preferably, when adjusting the node coordinates through smoothing processing, the path segment can be optimized through curve fitting. For example, if there is a sharp turn in a certain segment of the detour path, after smoothing, the straight line between nodes is adjusted to a smooth curve to reduce the abruptness during navigation.

[0178] After the smooth path dataset is generated, it is combined with the real-time traffic conditions to determine whether the navigation goal is met. For example, if sudden congestion is detected on a certain section of the mixed path, the system can dynamically adjust and preferentially select a detour path as the optimized navigation path. This dynamic adjustment mechanism can effectively cope with complex traffic environments.

[0179] Step S540: According to the smooth path dataset, combined with the real-time updated traffic conditions, determine whether the path meets the navigation goal and determine the optimized navigation path.

[0180] For example, the update of real-time traffic conditions may be based on sensor or user feedback data.

[0181] Suppose during the navigation process, an accident occurs in front of the main road, and the system receives the information that the congestion index has risen to 0.9. It immediately re-evaluates the path and confirms that the detour path is better. This method is supported by multi-dimensional data, ensuring the timeliness and accuracy of path selection.

[0182] It can be understood that each link of the above method is closely connected. From path generation to cost calculation, screening and verification to smoothing optimization, a complete navigation path optimization process is formed.

[0183] For example, node division not only facilitates cost calculation but also provides basic data for subsequent smoothing processing, while real-time traffic update further enhances the dynamic adaptability of the path.

[0184] This design with multi-faceted mutual support can provide efficient navigation support in complex urban environments.

[0185] Preferably, for the intelligent navigation method provided in this embodiment, step S600 includes:

[0186] Step S610: Obtain preset navigation path data from the path planning module, obtain real-time environment data from the environment perception module, and use a data fusion method to integrate the navigation path data and the environment data to obtain the fused path feature data.

[0187] Exemplarily, when obtaining preset navigation path data from the path planning module, it usually involves a set of multiple candidate paths from the starting point to the ending point.

[0188] Taking an autonomous vehicle as an example, the path planning module may generate a recommended path from point A to point B in the city, including information such as road type and number of lanes.

[0189] Suppose the path data shows that the whole journey is 5 kilometers, including 2 kilometers of urban roads and 3 kilometers of highways.

[0190] The environmental perception module obtains real-time environmental data through sensors such as lidar and cameras. For example, there are pedestrians or obstacles 100 meters ahead. This data provides a dynamic basis for subsequent fusion.

[0191] In a possible implementation, when integrating navigation path data and environmental data using a data fusion method, techniques such as weighted average or Kalman filtering can be used.

[0192] Suppose the path data indicates going straight, and the environmental data detects a temporary construction area 50 meters ahead. After fusion, path feature data is generated, which may be adjusted to decelerate or change lanes 50 meters in advance. The fused data is closer to the actual scenario and can provide reliable input for subsequent decisions.

[0193] Step S620: According to the path feature data and the real-time force feedback data collected by the force feedback module, use the proportional-integral-derivative algorithm to calculate the current error value and the cumulative historical error value, and determine the adjusted control parameters.

[0194] Specifically, according to the path feature data and the real-time force feedback data of the force feedback module, the error value is calculated using the proportional-integral-derivative algorithm. For example, the force feedback module detects a deflection force of 0.5 Newton on the steering wheel, indicating that the vehicle may deviate from the preset path.

[0195] The proportional part adjusts the control force according to the current error, the integral part accumulates the deviation in the past 10 seconds, and the derivative part predicts the future trend. Finally, the adjusted control parameters are generated, such as correcting the steering wheel angle by 2 degrees. This method analyzes the error through multiple dimensions to ensure control accuracy.

[0196] Step S630: If the control parameter exceeds the preset threshold, smooth the control parameter through the dynamic response module to obtain the optimized control instruction data.

[0197] Preferably, if the control parameter exceeds the preset threshold, smooth it through the dynamic response module.

[0198] Suppose the threshold is 5 degrees and the calculated parameter is 7 degrees. The dynamic response module may adjust the angle to 4.8 degrees through a smooth curve to avoid discomfort caused by sudden changes. For example, in a sharp turn scenario, the smoothing process enables the vehicle to turn with a more natural posture, improving the stability of execution.

[0199] Step S640: Convert the control instruction data into an execution signal of the target device through the instruction output module to obtain the final navigation control signal.

[0200] For example, when converting the control instruction data into an execution signal of the target device through the instruction output module, it may be converting the adjusted steering wheel angle of 4.8 degrees into a specific pulse signal of the motor.

[0201] Assume that the motor requires 100 pulses per degree, and the command output module generates a signal of 480 pulses to drive the steering wheel to rotate precisely.

[0202] This conversion ensures seamless connection between the commands and the hardware, and finally forms a navigation control signal to drive the vehicle to travel along the optimized path.

[0203] It can be understood that the implementation methods of each of the above topics are closely linked. From path planning to final signal output, a complete navigation control link is formed. For example, the fused data provides a basis for error calculation, and the smoothing process optimizes the hardware execution effect. This interlocking logic ensures the reliability and efficiency of the navigation system.

[0204] The present invention relates to an intelligent navigation system for implementing the above intelligent navigation method. The intelligent navigation system includes a first acquisition module, a second acquisition module, a correction module, a first generation module, a third acquisition module, and a second generation module. Among them, the first acquisition module is used to acquire multi-modal data collected by a tactile sensor, a vision sensor, and a force sensor. The multi-modal data includes tactile data, vision data, and force data. The second acquisition module is used to adjust the force data with a weighting coefficient to obtain a fused feature vector if the matching degree between the tactile data and the vision data is lower than a preset threshold. The correction module is used to generate a three-dimensional point cloud map according to the fused feature vector, and trigger an incremental update algorithm to correct the environmental model if the detected dynamic change rate of the point cloud map is higher than a preset threshold. The first generation module is used to extract a terrain feature set from the environmental model and generate a candidate path set by using a reinforcement learning algorithm. The third acquisition module is used to smooth the candidate path to obtain an optimized navigation path if the passing cost value of the candidate path is lower than a preset threshold. The second generation module is used to generate control parameters according to the optimized navigation path and the force feedback data by using a proportional-integral-derivative algorithm.

[0205] Furthermore, for the intelligent navigation system provided in this embodiment, the first acquisition module includes a first acquisition unit, a second acquisition unit, a determination unit, and a third acquisition unit. Among them, the first acquisition unit is configured to acquire raw data from a tactile sensor, a vision sensor, and a force sensor, calibrate the raw data through timestamp alignment and unit standardization, filter out noise using a preset threshold, and obtain a calibrated multi-modal data set. The raw data includes raw tactile data, raw vision data, and raw force data. The second acquisition unit is configured to extract the pressure distribution feature of the tactile data, the edge contour feature of the vision data, and the torque feature of the force data from the calibrated multi-modal data set using a feature extraction algorithm, and map them to a unified feature space through data fusion to obtain a fused feature vector. The determination unit is configured to determine whether the dimension of the fused feature vector exceeds a preset threshold. If so, perform dimensionality reduction processing using the principal component analysis algorithm, and use a pattern recognition algorithm to determine the environmental state based on the dimensionality-reduced feature vector to obtain an environmental perception result. The third acquisition unit is configured to match the environmental perception result with a pre-established action control mapping table to obtain an action control instruction, and combine real-time anomaly detection to determine whether the instruction triggers an anomaly threshold to obtain action control parameters.

[0206] Preferably, for the intelligent navigation system provided in this embodiment, the second acquisition module includes a fourth acquisition unit, a fifth acquisition unit, a sixth acquisition unit, and a seventh acquisition unit. Among them, the fourth acquisition unit is configured to acquire tactile data and vision data from a tactile sensor and a vision sensor, generate a tactile feature set and a vision feature set through feature extraction, and calculate the matching degree between the tactile feature set and the vision feature set using the cosine similarity algorithm to obtain a matching degree value. The fifth acquisition unit is configured to determine whether the matching degree value is lower than a preset threshold. If so, obtain a weighting coefficient according to a preset weighting coefficient table, and perform weighted adjustment on the force data through matrix multiplication to obtain adjusted force data. The sixth acquisition unit is configured to splice the adjusted force data with the vision feature set through a data fusion algorithm, and perform dimensionality reduction processing on the spliced data using the principal component analysis algorithm to obtain a preliminary fused feature vector. The seventh acquisition unit is configured to normalize the elements of the preliminary fused feature vector through normalization processing according to the dimensionality requirement of the preliminary fused feature vector to obtain a final fused feature vector.

[0207] This embodiment provides an intelligent navigation method and system. Compared with the prior art, by acquiring multi-modal data from tactile, visual, and force sensors, performing matching degree analysis and weighted adjustment on the data to generate a fused feature vector; constructing a three-dimensional point cloud map based on the feature vector and dynamically correcting the environmental model using an incremental update algorithm; extracting terrain features from the environmental model, generating a low-cost candidate path using a reinforcement learning algorithm, and performing smoothing processing to obtain an optimized navigation path; finally, combining force feedback and using a PID algorithm to generate control parameters. The intelligent navigation method and system provided by this embodiment achieve precise navigation and path planning in complex environments, improve the adaptability and robustness of the navigation system, and can be widely applied to fields such as autonomous mobile robots and driverless vehicles.

[0208] Although the preferred embodiments of the present invention have been described, those skilled in the art can make additional changes and modifications to these embodiments once they know the basic creative concept. Therefore, the appended claims are intended to be construed to include the preferred embodiments as well as all changes and modifications falling within the scope of the present invention. Obviously, those skilled in the art can make various changes and variations to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalent technologies, the present invention is also intended to include these modifications and variations.

Claims

1. An intelligent navigation method, characterized in that, Including the following steps: Obtain multimodal data collected by a tactile sensor, a visual sensor, and a force sensor, where the multimodal data includes tactile data, visual data, and force data; Judge whether the matching degree between the tactile data and the visual data is lower than a preset threshold. If so, adjust the force data using a weighting coefficient to obtain a fused feature vector; Generate a three-dimensional point cloud map according to the fused feature vector. If it is detected that the dynamic change rate of the three-dimensional point cloud map is higher than a preset threshold, trigger an incremental update algorithm to correct the environmental model; Extract a terrain feature set from the environmental model and generate a candidate path set using a reinforcement learning algorithm; Judge whether the passing cost value of the candidate path is lower than a preset threshold. If so, smooth the candidate path to obtain an optimized navigation path; Generate control parameters using a proportional-integral-derivative algorithm according to the optimized navigation path and the force feedback data.

2. The intelligent navigation method according to claim 1, wherein, The step of obtaining multimodal data collected by a tactile sensor, a visual sensor, and a force sensor includes: Obtain raw data from a tactile sensor, a visual sensor, and a force sensor, calibrate the raw data through timestamp alignment and unit standardization, and filter noise using a preset threshold to obtain a calibrated multimodal data set, where the raw data includes raw tactile data, raw visual data, and raw force data; For the calibrated multimodal data set, use a feature extraction algorithm to extract the pressure distribution feature of the tactile data, the edge contour feature of the visual data, and the torque feature of the force data, and map them to a unified feature space through data fusion to obtain a fused feature vector; If the dimension of the fused feature vector exceeds a preset threshold, perform dimensionality reduction processing using a principal component analysis algorithm, and judge the environmental state according to the dimensionality-reduced feature vector using a pattern recognition algorithm to determine the environmental perception result; Match the environmental perception result with a pre-established action control mapping table to obtain an action control instruction, and combine real-time anomaly detection to judge whether the instruction triggers an anomaly threshold to obtain action control parameters.

3. The intelligent navigation method according to claim 1, wherein The step of judging whether the matching degree between the tactile data and the visual data is lower than a preset threshold. If so, adjusting the force data using a weighting coefficient to obtain a fused feature vector includes: Obtain tactile data and visual data from a tactile sensor and a visual sensor, generate a tactile feature set and a visual feature set through feature extraction, and calculate the matching degree between the tactile feature set and the visual feature set using a cosine similarity algorithm to obtain a matching degree value; Judge whether the matching degree value is lower than a preset threshold. If so, obtain a weighting coefficient according to a preset weighting coefficient table, and perform weighted adjustment on the force data through matrix multiplication to obtain adjusted force data; Stitch the adjusted force data with the visual feature set through a data fusion algorithm, and perform dimensionality reduction processing on the stitched data using a principal component analysis algorithm to obtain a preliminary fused feature vector; According to the dimensionality requirements of the preliminary fusion feature vector, normalize the elements of the preliminary fusion feature vector through standardization processing to obtain the final fusion feature vector.

4. The intelligent navigation method according to claim 1, characterized in that, The steps of generating a three-dimensional point cloud map based on the fusion feature vector and triggering an incremental update algorithm to correct the environmental model if the detected dynamic change rate of the three-dimensional point cloud map is higher than a preset threshold include: Obtain sensor data from a lidar and a camera, generate a first feature set through a feature extraction algorithm, and fuse them into a first fusion feature vector using a data integration method based on the first feature set to obtain a three-dimensional point cloud map; Perform change detection on the three-dimensional point cloud map and the second point cloud map at the previous moment, calculate the first dynamic change rate, and determine whether the first dynamic change rate is higher than a preset threshold; If the first dynamic change rate is higher than the preset threshold, locally adjust the three-dimensional point cloud map through an incremental update algorithm to obtain a first map correction result; Update the environmental model according to the first map correction result, and adjust the parameters of the environmental model through a model optimization method to obtain an optimized first environmental model.

5. The intelligent navigation method according to claim 1, wherein The steps of extracting a terrain feature set from the environmental model and generating a candidate path set using a reinforcement learning algorithm include: Obtain spatial information from environmental data, and decompose it into height, slope, and curvature through rasterization processing to obtain a terrain feature set; According to the terrain feature set, use a reinforcement learning algorithm, combined with a preset reward function and path constraints, to generate a candidate path set; If there are paths in the candidate path set that do not meet the path length or obstacle avoidance requirements, recalculate the path constraints through the spatial information to obtain an optimized path set; Through the optimized path set, obtain the height, slope, and curvature information of each path, and determine whether it meets the preset threshold to obtain the final path set.

6. The intelligent navigation method according to claim 1, wherein The steps of determining whether the passing cost value of the candidate path is lower than a preset threshold, and if so, smoothing the candidate path to obtain an optimized navigation path include: Obtain a candidate path set from a navigation target, divide the path segments using node coordinates, and calculate the passing cost value of each candidate path according to the path length and traffic conditions to obtain a passing cost data set; Determine whether the passing cost value in the passing cost data set is lower than a preset threshold. If so, extract the corresponding candidate path, and verify the effectiveness of the candidate path using data accuracy to obtain a path set to be smoothed; Adjust the node coordinates of the path set to be smoothed through smoothing processing to obtain an optimized path segment, and obtain a smoothed path data set; According to the smoothed path data set, combined with the real-time updated traffic conditions, determine whether the path meets the navigation target to determine the optimized navigation path.

7. The intelligent navigation method according to claim 1, characterized in that, The steps of generating control parameters using a proportional-integral-derivative algorithm according to the optimized navigation path and the force feedback data include: Obtain preset navigation path data from a path planning module, obtain real-time environmental data from an environmental perception module, and integrate the navigation path data and the environmental data using a data fusion method to obtain fused path feature data; According to the path feature data and the real-time force sense data collected by the force sense feedback module, the proportional-integral-differential algorithm is used to calculate the current error value and the cumulative value of historical errors, and the adjusted control parameters are determined; If the control parameter exceeds the preset threshold, the dynamic response module is used to smooth the control parameter to obtain the optimized control instruction data; The control instruction data is converted into an execution signal of the target device through the instruction output module to obtain the final navigation control signal.

8. An intelligent navigation system for implementing the intelligent navigation method according to any one of claims 1 to 7, characterized in that, The intelligent navigation system includes: The first acquisition module is used to acquire multi-modal data collected by a tactile sensor, a vision sensor, and a force sense sensor, and the multi-modal data includes tactile data, vision data, and force sense data; The second acquisition module is used to determine that the matching degree between the tactile data and the vision data is lower than the preset threshold. If so, a weighted coefficient is used to adjust the force sense data to obtain a fused feature vector; The correction module is used to generate a three-dimensional point cloud map according to the fused feature vector. If it is detected that the dynamic change rate of the point cloud map is higher than the preset threshold, an incremental update algorithm is triggered to correct the environment model; The first generation module is used to extract a terrain feature set from the environment model and generate a candidate path set by using a reinforcement learning algorithm; The third acquisition module is used to determine whether the passing cost value of the candidate path is lower than the preset threshold. If so, the candidate path is smoothed to obtain an optimized navigation path; The second generation module is used to generate control parameters by using the proportional-integral-differential algorithm according to the optimized navigation path and the force sense feedback data.

9. The intelligent navigation system according to claim 8, characterized in that, The first acquisition module includes: The first acquisition unit is used to obtain raw data from a tactile sensor, a vision sensor, and a force sense sensor, calibrate the raw data through timestamp alignment and unit standardization, and filter noise by using a preset threshold to obtain a calibrated multi-modal data set. The raw data includes raw tactile data, raw vision data, and raw force sense data; The second acquisition unit is used to extract the pressure distribution feature of the tactile data, the edge contour feature of the vision data, and the torque feature of the force sense data from the calibrated multi-modal data set by using a feature extraction algorithm, and map them to a unified feature space through data fusion to obtain a fused feature vector; The determination unit is used to determine that the dimension of the fused feature vector exceeds the preset threshold. If so, a principal component analysis algorithm is used for dimensionality reduction processing, and the environmental state is judged by using a pattern recognition algorithm according to the dimensionality-reduced feature vector to determine the environmental perception result; The third acquisition unit is used to match the environmental perception result with a pre-established action control mapping table to obtain an action control instruction, and combine real-time anomaly detection to judge whether the instruction triggers an anomaly threshold to obtain action control parameters.

10. The intelligent navigation system according to claim 8, characterized in that, The second acquisition module includes: The fourth acquisition unit is used to obtain tactile data and vision data from a tactile sensor and a vision sensor, generate a tactile feature set and a vision feature set through feature extraction, and calculate the matching degree between the tactile feature set and the vision feature set by using a cosine similarity algorithm to obtain a matching degree value; A fifth acquisition unit, configured to determine whether the matching degree value is lower than a preset threshold. If so, a weighting coefficient is obtained according to a preset weighting coefficient table, and the haptic data is weighted and adjusted through matrix multiplication to obtain adjusted haptic data; A sixth acquisition unit, configured to splice the adjusted haptic data with the visual feature set through a data fusion algorithm, and perform dimensionality reduction processing on the spliced data by using a principal component analysis algorithm to obtain a preliminary fusion feature vector; A seventh acquisition unit, configured to normalize the elements of the preliminary fusion feature vector through normalization processing according to the dimensionality requirement of the preliminary fusion feature vector to obtain a final fusion feature vector.

Images