A navigation method for robots in complex scenarios

Abstract

This invention relates to the field of robotics and discloses a navigation method for robots in complex scenarios. The method includes: receiving and parsing material transportation instructions to determine the location of the materials and the target point; activating a multi-source sensor system and performing self-checks on each sensor within the system to obtain the state values ​​of each sensor; determining and executing a positioning strategy based on the state values ​​of each sensor to obtain the robot's positioning result; generating a global path by calling a path planning algorithm based on the material location, the target point location, and the robot's positioning result; identifying the current terrain using a terrain recognition module to obtain the terrain type; matching the motion mode corresponding to the terrain type and navigating according to the motion mode until the global path is traversed. This invention can significantly improve robot navigation accuracy, accurately identify scene and terrain types, and adapt motion modes.

Description

Technical Field

[0001] This invention relates to the field of robotics, and more particularly to a navigation method for robots in complex scenarios. Background Technology

[0002] Industrial production scenarios such as smart manufacturing plants, chemical industrial parks, and large-scale warehousing and logistics centers are vulnerable to various types of disasters, including natural disasters, safety accidents, and extreme environmental events. The navigation systems carried by traditional mobile robots exhibit significant technical limitations under the complex characteristics of disaster scenarios, such as strong nonlinearity, high dynamism, and unstructured nature, making it difficult to meet the core needs of emergency response and operational execution.

[0003] Specifically, satellite positioning technology based on GNSS (Global Navigation Satellite System, including GPS, BeiDou, etc.) is prone to problems such as signal blockage and multipath effects in indoor spaces, enclosed areas formed by building collapses, or industrial electromagnetic interference environments. This leads to a deterioration in positioning accuracy (errors can reach the meter level) and may even cause positioning drift or complete loss of lock, making it impossible to provide a reliable position reference for robots.

[0004] LiDAR (Light Detection and Ranging) acquires 3D point cloud data of the environment by emitting laser beams. It excels at recognizing the geometric contours of rigid obstacles (such as walls and equipment wreckage), but it has shortcomings in terrain texture classification. Its ability to extract texture features from soft deposits (such as sand and gravel in ruins and collapsed building materials) and flexible obstacles (such as suspended cables and damaged protective fabrics) is insufficient, making it difficult to accurately distinguish between passable and dangerous areas, which can easily lead to path planning misjudgments.

[0005] Binocular vision relies on the parallax calculation of two cameras to achieve 3D reconstruction and texture recognition. Its performance is highly dependent on ambient lighting conditions and visual transparency. In low-visibility scenarios such as smoke from fires, industrial dust diffusion, and dense smoke obscuring the view, image texture details are severely lost, the robustness of stereo matching algorithms decreases significantly, leading to increased parallax calculation errors, significantly reduced texture recognition accuracy, and an inability to effectively perceive terrain undulations and obstacle distribution.

[0006] In addition, although existing multi-sensor integrated navigation systems have improved theoretical reliability through sensor redundancy configuration, they are prone to single-sensor or multi-sensor synchronous failure under extreme conditions in disaster scenarios (such as high temperature, severe vibration, electromagnetic pulse impact).

[0007] In summary, a single positioning method leads to inaccurate robot navigation accuracy, while multi-sensor navigation systems, when sensors fail in extreme situations, lack efficient fusion of data from multiple sensors, making it difficult to adapt to complex and ever-changing environmental conditions in disaster scenarios. Summary of the Invention

[0008] To address the aforementioned problems of traditional robot navigation systems, this invention proposes a robot navigation method in complex scenarios, which can significantly improve robot navigation accuracy, accurately identify scene and terrain types, and adapt to motion modes.

[0009] The technical solution adopted in this invention is as follows:

[0010] A navigation method for robots in complex scenarios includes:

[0011] Step 1: Receive and parse the material transportation instruction to determine the location of the material and the target point;

[0012] Step 2: Activate the multi-source sensor system and perform self-tests on each sensor within the system to obtain the status values ​​of each sensor; the sensors in the multi-source sensor system include RTK measuring instruments, lidar, binocular vision cameras, inertial measurement units, and GPS positioning units;

[0013] Step 3: Determine and execute the positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result;

[0014] Step 4: Based on the location of the materials, the location of the target point, and the robot's localization results, call the path planning algorithm to generate a global path;

[0015] Step 5: While moving along the global path, the terrain recognition module identifies the current terrain to obtain the terrain type; the movement mode corresponding to the terrain type is matched, and navigation is performed according to the movement mode until the global path is traversed.

[0016] Further, in step 3, determining and executing the positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result includes:

[0017] When the state values ​​of all sensors in the multi-source sensor system indicate a normal state, the detection data of each sensor are fused by combining the extended Kalman filter method and the factor graph optimization method to obtain the robot localization result.

[0018] Furthermore, the method of combining extended Kalman filtering and factor graph optimization to fuse the detection data from each sensor to obtain the robot localization result includes:

[0019] Given that the current time is in the first stage, the robot's pose information is obtained by fusing the RTK measurement instrument detection data and the inertial measurement unit detection data through the Kalman filter method, and this information is used as the robot's localization result.

[0020] Given that the current time is in the second stage, based on the robot's pose information, the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are used as constraint factors. The robot's pose information is updated through the factor graph optimization method and used as the robot localization result.

[0021] The first stage and the second stage are two consecutive time periods.

[0022] Furthermore, the step of fusing RTK measurement instrument detection data and inertial measurement unit detection data using the Kalman filtering method to obtain the robot's pose information includes:

[0023] Based on the acceleration and angular velocity data output by the inertial measurement unit, the robot's pose state at the next moment is predicted by the Newtonian kinematics model.

[0024] The coordinate data output by the RTK measuring instrument is used as the measured value. It is compared with the predicted pose state, the Kalman gain is calculated and the predicted pose state is corrected to obtain the fused pose estimate, which is used as the robot's pose information.

[0025] Furthermore, based on the robot's pose information, the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are used as constraint factors to update the robot's pose information through a factor graph optimization method, including:

[0026] The robot's pose information obtained in the first stage is used as the initial node data, and the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are used as constraint factors.

[0027] A factor graph is constructed using robot pose as nodes and constraint factors as edges. The Gauss-Newton algorithm is used to minimize the sum of errors of each constraint factor, and the pose nodes are iteratively optimized to obtain the optimized factor graph.

[0028] Based on the optimized factor graph, the historical pose end data of the first stage is obtained, and the initial value is reset using the historical pose end data.

[0029] Based on the optimized factor graph, the optimized pose trajectory is determined as the reference trajectory. Based on the initial value, the reference trajectory, and the incremental optimization strategy, incremental constraints are determined. The robot's pose information is predicted and updated through the incremental constraints. The incremental constraints are used to ensure the consistency between the initial value and the subsequent updated value and the reference trajectory.

[0030] Furthermore, in step 3, determining and executing the positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result includes: when at least one sensor in the multi-source sensor system is unavailable, determining a backup positioning strategy after combining the normal state sensors based on the pre-configured sensor list.

[0031] Further, in step 5, the step of identifying the current terrain using the terrain recognition module to obtain the terrain type includes:

[0032] A three-dimensional terrain feature vector is constructed based on multi-sensor data. The three-dimensional terrain feature vector includes slope, roughness, obstacle height, and texture entropy information. The multi-sensor data includes absolute position data from RTK measuring instruments, environmental point cloud data from lidar, depth information from binocular vision cameras, attitude and motion change data from inertial measurement units, and auxiliary positioning data from GPS positioning units.

[0033] The terrain type is determined based on the three-dimensional terrain feature vector, and the terrain type includes fire scene, earthquake scene and disaster-free scene.

[0034] Further, in step 5, the motion pattern corresponding to the terrain type is matched, and navigation is performed according to the motion pattern until the global path is traversed, including:

[0035] When the terrain type is a fire scenario, the priority of the sensors is adjusted and the movement mode is set to the first type of movement mode. The first type of movement mode includes: marking areas of molten material with a temperature higher than a preset temperature as no-entry zones, identifying areas of water accumulation or water vapor accumulation and detouring around them; activating high-temperature protective gait and increasing the step frequency; and using a single-leg support jump when crossing burning debris.

[0036] Navigation is performed according to the first type of motion mode, and the local path is dynamically adjusted using the dynamic window method until the global path is traversed.

[0037] Further, in step 5, the motion pattern corresponding to the terrain type is matched, and navigation is performed according to the motion pattern until the global path is traversed, including:

[0038] When the terrain type is an earthquake scenario, the sensor priority is adjusted and the motion mode is set to the second type of motion mode. The second type of motion mode includes: identifying concrete cracks with an actual width greater than a preset width and marking them as needing to be detoured; identifying sloping ground with an actual slope greater than a preset slope and marking it as dangerous; judging the stability of stacked gravel by the variance of the point cloud data of the lidar, and marking areas with a point cloud data variance greater than a preset variance as unstable areas; adopting a trial-and-support gait in the gravel area, and enabling bipedal coordination when crossing cracks;

[0039] Navigation is performed according to the second type of motion mode, and the local path is dynamically adjusted using the dynamic window method until the global path is traversed.

[0040] Furthermore, after completing the traversal of the global path, the process also includes:

[0041] The robot uses a fixed device and pressure sensors to monitor the grabbed materials in real time. If the pressure sensor detects abnormal pressure fluctuations that last for more than a preset time, it is determined that the materials are at risk of falling, and the robot's movement posture is adjusted or its movement speed is reduced.

[0042] Upon reaching the target area, images are captured using a binocular vision camera, and acoustic signals are transmitted and received using a sonic positioning device to analyze the shape and posture characteristics of the trapped personnel. The robot's posture is then adjusted, and the robotic arm unloads the supplies into the reach of the trapped personnel.

[0043] Wait for confirmation signal, record task data based on confirmation signal result, return to preset standby point or retry delivery; the task data includes path information, terrain recognition results, motion mode switching status and material transportation status.

[0044] The beneficial effects of this invention are as follows:

[0045] This invention is based on a multi-source sensor data fusion strategy, which integrates detection data from RTK measuring instruments, inertial measurement units, lidar, binocular vision cameras, etc., and configures a backup positioning strategy when sensors fail. This enables accurate and stable output of robot pose information, ensuring the robot can move safely and stably on various complex terrains, thereby achieving precise navigation of the robot in complex scenarios. Attached Figure Description

[0046] Figure 1 This is a flowchart of a robot navigation method in a complex scene according to Embodiment 1 of the present invention.

[0047] Figure 2 This is a flowchart of a robot navigation method in a complex scene according to Embodiment 2 of the present invention. Detailed Implementation

[0048] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments are now described. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention; that is, the described embodiments are only a part of the embodiments of the invention, not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0049] Example 1

[0050] like Figure 1 As shown, this embodiment provides a navigation method for a robot in complex scenarios, including:

[0051] Step 1: Receive and parse the material transportation instruction to determine the location of the material and the target point;

[0052] Step 2: Activate the multi-source sensor system, perform self-tests on each sensor within the system, and obtain the status values ​​of each sensor;

[0053] Step 3: Determine and execute the positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result;

[0054] Step 4: Based on the location of the materials, the location of the target point, and the robot's localization results, call the path planning algorithm to generate a global path;

[0055] Step 5: While moving along the global path, the terrain recognition module identifies the current terrain to obtain the terrain type; matches the movement mode corresponding to the terrain type, and navigates according to the movement mode until the global path is traversed.

[0056] The sensors in the multi-source sensor system can be RTK (Real-Time Kinematic) measuring instruments, LiDAR, binocular vision cameras, IMU (Inertial Measurement Unit), and GPS positioning units, etc., used to acquire data on the robot's own state and the surrounding environment. The terrain recognition module refers to the functional module that uses multi-sensor data fusion analysis to determine the type of terrain currently in the robot's location.

[0057] Specifically, the navigation method of this embodiment can be implemented using the following process:

[0058] Step 1: The robot receives material transportation instructions sent by the host computer or remote control terminal through the communication module, parses the key information in the instructions, clarifies the specific storage location coordinates of the materials and the location coordinates of the target point, and establishes the task coordinate system.

[0059] Step 2: Start the multi-source sensor system, send self-test commands to each sensor, and each sensor will report its own power supply status, data transmission status, and functional operation status. The status values ​​of each sensor are then summarized.

[0060] Step 3: Determine the status values ​​of each sensor, distinguish between cases where all are normal or partially abnormal, and execute different positioning strategies accordingly. Calculate the robot's real-time positioning result through data fusion or backup schemes.

[0061] Step 4: Input the parsed material location, target point location, and robot localization results into the path planning algorithm. The algorithm combines preset passage rules (such as avoiding known dangerous areas) to generate a global path connecting the starting point, material location, and target point.

[0062] Step 5: As the robot moves along the global path, it continuously collects and analyzes multi-sensor data through the terrain recognition module to identify the current terrain type; based on the terrain type, it matches the preset motion mode and adjusts the robot's driving parameters and avoidance strategies until it has completely traversed the global path and completed the transportation task.

[0063] It should be noted that the navigation method in this embodiment, through the above-mentioned instruction parsing, sensor activation, positioning strategy selection, path planning and terrain adaptation process, can enable the robot to navigate autonomously in complex disaster scenarios, ensuring that the robot can accurately and safely complete material transportation tasks and adapt to the complex needs of different scenarios.

[0064] Preferably, in step 3, when all sensors in the multi-source sensor system are in normal condition, a fusion strategy combining the extended Kalman filter (EKF) method and the factor graph optimization method is used to fuse the detection data of each sensor to finally obtain the robot localization result.

[0065] In this context, "sensor in normal condition" refers to a state where the sensor can stably output valid data without faults, data loss, or abnormal fluctuations. The Extended Kalman Filter (EKF) method is a data analysis method based on recursive Bayesian estimation, used for state estimation of nonlinear systems to achieve preliminary fusion of multi-source data. The Factor Graph Optimization method is an optimization algorithm based on a graph model, constructing a factor graph with state variables as nodes and constraints as edges, achieving state optimization by minimizing errors. Detection data fusion processing integrates complementary data collected by different sensors, eliminating the limitations of a single sensor and obtaining more accurate state estimation results.

[0066] Specifically, when the status values ​​of all sensors in the multi-source sensor system are normal, the data fusion and positioning process is initiated: First, the detection data of each sensor is preprocessed to unify the data format and coordinate system; then, the extended Kalman filter method is used to perform preliminary fusion of the multi-sensor data to reduce the impact of data noise; then, the factor graph optimization method is introduced to perform secondary optimization of the preliminary fusion results, correct system errors, and finally obtain high-precision robot positioning results.

[0067] It should be noted that by combining extended Kalman filtering with factor graph optimization, we can achieve high-frequency pose output using extended Kalman filtering to meet the real-time requirements of motion control, and correct accumulated errors through factor graph optimization to achieve high-precision and high-stability positioning results.

[0068] This embodiment divides the localization process into two adjacent time periods, namely the first stage and the second stage. Preferably, when the current time belongs to the first stage, the robot's pose information is obtained by fusing the RTK measurement instrument detection data and the inertial measurement unit detection data through the Kalman filter method, and this is used as the robot localization result. When the current time belongs to the second stage, based on the robot's pose information, the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are used as constraint factors, and the robot's pose information is updated through the factor graph optimization method, which is also used as the robot localization result.

[0069] The first stage refers to the initial phase of the localization process, a period focused on quickly acquiring initial pose information. The second stage, adjacent to the first, is the subsequent phase of the localization process, involving precise optimization based on the initial pose information. Pose information refers to comprehensive state data including the robot's position coordinates and attitude angles (such as pitch, roll, and yaw). Point cloud matching error refers to the deviation when matching point cloud data acquired by the LiDAR with preset maps or historical point cloud data. Reprojection error refers to the deviation between the projected image feature points acquired by the binocular vision camera and their actual scene positions. Constraint factors are error parameters used to limit the optimization range of pose nodes, ensuring that the optimization results conform to the actual scene.

[0070] Specifically, in the first stage, the position detection data of the RTK measuring instrument and the motion state data of the inertial measurement unit are received. The two types of data are fused and calculated using the Kalman filter method to quickly obtain the robot's initial pose information, which is then output as the positioning result. In the second stage, based on the pose information obtained in the first stage, the point cloud matching error of the lidar and the reprojection error of the binocular vision camera are extracted and set as constraint factors.

[0071] It should be noted that by adopting a phased positioning strategy, a balance between real-time positioning and accuracy is achieved. The first phase ensures the immediate response of motion control, while the second phase corrects accumulated errors, ensuring that the robot maintains high-precision positioning continuously during long-term and long-distance tasks.

[0072] More preferably, in the first stage, the robot's pose state at the next moment is predicted based on the acceleration and angular velocity data output by the inertial measurement unit, using the Newtonian kinematics model. Then, the coordinate data output by the RTK measuring instrument is used as the measurement value to calculate the Kalman gain and correct the predicted pose, finally obtaining the fused pose information.

[0073] The Newtonian kinematics model, based on Newton's laws of motion, is a mathematical model used to describe the changes in the robot's motion state over time. The Kalman gain is a parameter used to adjust the weights of the predicted and measured values; it is dynamically adjusted based on data errors to optimize the fusion result.

[0074] Specifically, in the first stage, the real-time acceleration and angular velocity data output by the inertial measurement unit are extracted and input into a preset Newtonian kinematics model. Then, based on the Newtonian kinematics model, the robot's pose state (including position and attitude) at the next moment is predicted according to the current motion state. The absolute coordinate data output by the RTK measuring instrument is obtained and compared with the predicted pose state as the measured value, and the deviation between the two is calculated. The Kalman gain is calculated according to the magnitude of the deviation. The weights of the predicted value and the measured value are dynamically adjusted through the Kalman gain to correct the predicted pose state. Finally, the fused pose estimate is obtained as the robot's pose information.

[0075] It should be noted that the above steps can fully leverage the advantages of high-frequency data output from the inertial measurement unit and the high-precision positioning of the RTK measurement instrument. Through a prediction-correction closed-loop process, high-frequency and precise pose output can be achieved, ensuring the real-time performance and accuracy of robot motion control.

[0076] More preferably, in the second stage, the pose information from the first stage is used as the initial node, and a factor map is constructed using the correlation error between LiDAR and binocular vision as constraint factors. The pose node is optimized using the Gauss-Newton algorithm, the initial value of the Kalman filter is reset using the optimized historical pose data, and the subsequent pose is updated based on the reference trajectory and incremental optimization strategy.

[0077] The factor graph is a graphical model constructed with robot poses as nodes and constraint factors as edges, used to visually represent the relationship between pose and constraints. The Gauss-Newton algorithm is an optimization algorithm for solving nonlinear least squares problems, iteratively approximating the optimal solution and minimizing the total error. The reference trajectory is the robot pose change trajectory presented in the optimized factor graph, serving as the benchmark for subsequent pose updates. The incremental optimization strategy is a constraint that ensures subsequent pose updates remain consistent with the reference trajectory, preventing the optimization results from deviating from reality.

[0078] Specifically, in the second stage, the robot pose information obtained in the first stage is used as the initial node data of the factor graph, while the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are set as constraint factors. A complete factor graph is constructed using the robot's pose at each moment as nodes and the constraint factors as edges connecting the nodes. The Gauss-Newton algorithm is used to iteratively optimize the factor graph, minimizing the sum of errors of all constraint factors, and continuously adjusting the values ​​of the pose nodes to obtain the optimized factor graph. The end data of the historical poses from the first stage are extracted from the optimized factor graph, and this data is used to reset the initial values ​​for subsequent optimization. The pose trajectory corresponding to the optimized factor graph is used as the reference trajectory, and the incremental constraints are determined by combining the reset initial values ​​with the incremental optimization strategy. Based on the incremental constraints, the robot's pose information at subsequent moments is predicted and continuously updated to ensure that the updated values ​​remain consistent with the reference trajectory.

[0079] It should be noted that factor graph optimization can effectively correct inertial measurement unit drift and short-term fusion errors, resetting the initial value of the Kalman filter can achieve cyclic improvement of positioning accuracy, and incremental optimization can ensure the consistency between the current pose and the historical trajectory, significantly reducing the cumulative error of long-term positioning.

[0080] Preferably, in step 3, when at least one sensor in the multi-source sensor system is unavailable, sensors in normal condition are selected and combined according to a pre-configured sensor list to determine and execute a corresponding backup positioning strategy. Here, "unavailable" refers to a working state where the sensor cannot output valid data, experiences abnormal data fluctuations, or malfunctions. The pre-configured sensor list refers to a pre-stored list of sensor combination schemes, containing backup sensing combinations for different sensor failure scenarios.

[0081] Specifically, the status values ​​of each sensor in the multi-source sensor system are monitored in real time to determine if at least one sensor is unavailable. If an unavailable sensor is detected, all sensors in normal condition are selected and their types and functions are recorded. The pre-configured sensor list is queried, and the corresponding sensor combination scheme is matched according to the type and function of the sensors in normal condition. Based on the matched sensor combination scheme, an alternative positioning strategy is determined, and the strategy is activated to collect and fuse data to obtain the robot positioning result.

[0082] It should be noted that in complex scenarios where some sensors fail, the backup positioning strategy ensures that the robot can still maintain basic positioning and motion control capabilities, avoiding task interruption due to a single sensor failure and improving the robustness of the navigation method.

[0083] Preferably, in step 5, a three-dimensional terrain feature vector is constructed based on multi-sensor data, and the terrain type is determined according to the three-dimensional terrain feature vector. The multi-sensor data refers to the comprehensive data from various sensors in a multi-source sensor system, covering multiple dimensions such as position, environment, depth, and attitude. Specifically, it can be absolute position data from an RTK measuring instrument, environmental point cloud data from a LiDAR, depth information from a binocular vision camera, attitude and motion change data from an inertial measurement unit, and auxiliary positioning data from a GPS positioning unit. The three-dimensional terrain feature vector is vector data of key terrain feature parameters used to quantify terrain attributes, specifically including slope, roughness, obstacle height, and texture entropy information. The terrain type is a scene category classified according to the terrain feature vector, specifically including fire scenes, earthquake scenes, and disaster-free scenes. The terrain type is the basis for motion pattern matching.

[0084] It should be noted that by fusing multi-dimensional sensor data to construct terrain feature vectors, accurate identification of complex disaster scenarios and normal scenarios can be achieved, providing an accurate basis for subsequent motion pattern matching and ensuring the robot's adaptability to different terrains.

[0085] More preferably, when the terrain type is a fire scene, the sensor priority is adjusted and the motion mode is set to the first type of motion mode; navigation is performed according to the first type of motion mode, and the local path is dynamically adjusted using the dynamic window method (DWA) until the global path is traversed.

[0086] Among these measures, adjusting sensor priorities involves resetting the operational priorities of each sensor based on the environmental characteristics of a fire scenario, ensuring that critical sensors (such as high-temperature resistant and interference-resistant sensors) operate first. The first type of movement mode is a dedicated movement mode designed for fire scenarios, including no-entry zone markings, detour strategies, gait adjustments, and jumping methods. The dynamic window method is a local path planning algorithm that dynamically adjusts the local path by analyzing environmental changes in real time to ensure driving safety.

[0087] Specifically, when the terrain recognition result indicates a fire scenario, the robot automatically adjusts the working priority of each sensor, prioritizing the operation of high-temperature resistant and interference-resistant sensors. It then activates the first type of movement mode: by detecting the ambient temperature through sensors, it marks molten areas exceeding the safe range as no-go zones; it identifies areas of accumulated water or steam and plans detour routes; it activates a high-temperature protective gait, increasing the robot's step frequency and reducing ground contact time; and when encountering burning debris, it switches to a single-leg support hop mode to overcome obstacles. During navigation, a dynamic window method is used to analyze changes in the surrounding environment in real time, dynamically adjusting the local path to avoid sudden danger zones until the global path traversal is completed.

[0088] It should be noted that by optimizing sensor priorities, accurately avoiding dangerous areas, and adjusting adaptive motion patterns, the robot can effectively cope with risks such as high temperatures, smoke, and sudden obstacles in fire scenarios, ensuring safe movement and navigation continuity.

[0089] More preferably, when the terrain type is an earthquake scenario, the sensor priority is adjusted, the second type of motion mode is enabled for navigation, and the local path is dynamically adjusted using a dynamic window method until the global path is traversed. The second type of motion mode is a dedicated motion mode designed for earthquake scenarios, which includes crack bypass, inclined ground marking, debris stability judgment, and dedicated gait.

[0090] Specifically, when the terrain identification result indicates an earthquake scenario, sensor priorities are adjusted to prioritize the operation of sensors used for terrain detection, such as LiDAR and inertial measurement units. A second type of movement mode is activated: sensors identify the actual width of concrete cracks and the ground slope, marking areas exceeding safety thresholds as areas requiring detours or dangerous zones; the variance of LiDAR point cloud data is calculated to determine the stability of the piled gravel, marking areas with excessive variance as unstable areas; a trial-and-support gait is used in gravel areas, and a bipedal cooperative mode is activated to traverse cracks. During navigation, a dynamic window method is used to adjust the local path in real time, avoiding unstable areas and dangerous terrain, until the global path traversal is completed.

[0091] It should be noted that by setting targeted sensor priorities, accurately identifying dangerous terrain, and adjusting gait accordingly, the robot can effectively cope with risks such as cracks, tilted ground, and unstable debris in earthquake scenarios, ensuring safe navigation in ruins environments.

[0092] Preferably, in step 5, after completing the traversal of the global path, the following sub-steps are also included:

[0093] The robot uses a fixed device and pressure sensors to monitor the grabbed materials in real time. If the pressure sensor detects abnormal pressure fluctuations that last for more than a preset time, it is determined that the materials are at risk of falling, and the robot's movement posture is adjusted or its movement speed is reduced.

[0094] Upon reaching the target area, images are captured using a binocular vision camera, and acoustic signals are transmitted and received using a sonic positioning device to analyze the shape and posture characteristics of the trapped personnel. The robot's posture is then adjusted, and the robotic arm unloads the supplies into the reach of the trapped personnel.

[0095] Wait for confirmation signal, record task data based on confirmation signal result, return to preset standby point or retry delivery; task data includes path information, terrain recognition results, movement mode switching status and material transportation status.

[0096] The securing device is a mechanical structure on the robot used to secure materials, working in conjunction with pressure sensors to monitor the material's status. Abnormal pressure fluctuations occur when the pressure sensors detect that the material's weight pressure exceeds the normal fluctuation range, indicating a potential risk of the material falling. The acoustic positioning device is a device that uses the transmission and reception of acoustic signals to locate targets, assisting in identifying the location of trapped personnel.

[0097] Specifically, after completing the traversal of the global path, the following implementation process can also be adopted:

[0098] After the robot grabs the materials, it fixes the materials in place using the mounting device. Pressure sensors monitor the pressure changes of the materials in real time and continuously provide pressure data.

[0099] If the pressure sensor detects abnormal pressure fluctuations that last for more than a preset duration, it determines that there is a risk of the material falling and automatically adjusts the robot's movement posture or reduces its movement speed to ensure the stability of the material.

[0100] Upon reaching the target area, the binocular vision camera acquires images of the surrounding environment, and the acoustic positioning device emits and receives acoustic signals. By combining the image and acoustic data, the shape and posture characteristics of the trapped person are analyzed to determine their location.

[0101] Adjust the robot's posture and control the robotic arm to unload the supplies into a range that the trapped personnel can reach;

[0102] Waiting to receive confirmation signals from the trapped personnel: If a confirmation signal is received, record the task data (including path information, terrain recognition results, movement mode switching status, and material transportation status) and return to the preset standby point; if no confirmation signal is received, readjust the posture and attempt delivery.

[0103] It should be noted that real-time material monitoring can prevent materials from falling during transportation, and the combination of binocular vision and acoustic positioning can achieve accurate location of trapped personnel. The retry mechanism can improve the success rate of material delivery and ensure the reliability and closed-loop management of emergency rescue missions.

[0104] Example 2

[0105] This embodiment is based on embodiment 1:

[0106] This embodiment provides a navigation method for a robot in complex scenarios, including:

[0107] Step 1: Receive material transportation instructions;

[0108] Step 2: Activate the multi-source sensor system;

[0109] Step 3: Determine the positioning plan and positioning;

[0110] Step 4: Generate global paths;

[0111] Step 5: Move according to the global path and adapt to the scene;

[0112] Step 6: Monitoring the status of materials;

[0113] Step 7: Execute the target task.

[0114] The navigation method in this embodiment is described in detail below.

[0115] I. Scene setting.

[0116] In a simulated disaster environment at a factory, humanoid robots are deployed to perform material transportation and rescue assistance tasks. The factory contains simulated fire (created by burning barrels and smoke generators) and earthquake (artificially constructed rubble) areas. The robots need to transport materials from material points (containing materials of different types, sizes, and weights) to target points (simulating the location of trapped personnel). The server (cloud) acts as a command sending and partial data processing center, communicating and interacting with the robots.

[0117] II. Equipment Preparation.

[0118] The equipment involved includes robots and servers (in the cloud), with the robots equipped with the following sensors and modules:

[0119] Binocular vision: 2-megapixel binocular camera (baseline distance 12cm, frame rate 30fps) is used to acquire depth information and recognize texture features.

[0120] GPS: As an auxiliary positioning source in open areas, the module is placed at the auxiliary antenna position on the back of the robot.

[0121] IMU: 6-axis IMU (accelerometer range ±16g, gyroscope range ±2000° / s), installed at the center of gravity of the fuselage.

[0122] LiDAR: A top-mounted 360° surround-view LiDAR used to detect terrain, obstacles, and other information.

[0123] RTK antenna: The RTK antenna at the highest point of the back, together with 5 RTK base stations (using 5.8GHz wireless band communication) distributed in the four corners and center of the factory, to achieve high-precision positioning.

[0124] The server (cloud) stores factory CAD maps and disaster area heat maps (such as fire-prone high-temperature zones, earthquake-prone collapse zones, etc.) for generating material transportation instructions and performing some data processing.

[0125] III. Specific Implementation Steps.

[0126] like Figure 2 As shown, the navigation method in this embodiment can be implemented using the following steps.

[0127] Step 1: Receive material transportation instructions.

[0128] The server (cloud-based) generates material transportation instructions based on the factory's simulated disaster scenario and rescue needs, and sends these instructions to the robot. The instructions include, for example, the target point information as coordinates (X1, Y1, Z1), the material point information as the material location as (X2, Y2, Z2), the material type as a medical first-aid kit, the material size as 50cm long, 40cm wide, and 30cm high, and the material weight as 10kg. The robot receives these instructions via its communication module, parses them to determine key information such as material location, target point location, material type, material size, and material weight, preparing for subsequent task execution.

[0129] Step 2: Activate the multi-source sensor system.

[0130] Upon receiving the command, the robot automatically activates its multi-source sensor system, sequentially performing self-tests on the RTK, LiDAR, binocular vision, IMU, and GPS. Specifically, the RTK module attempts to establish a communication connection with the RTK base station within the factory, verifying signal reception strength and data transmission stability; the LiDAR initiates a 360° surround scan, checking motor rotation, laser emission, and reception components to ensure normal point cloud data acquisition; the binocular vision camera is powered on and starts, detecting image acquisition and transmission links, verifying frame rate, pixel count, and other parameters; the IMU module performs a self-test, checking the operating parameters of the accelerometer and gyroscope through internal diagnostics to confirm that range, zero bias, etc., are within normal ranges; and the GPS module searches for satellite signals, verifying the number of received signals and initial positioning accuracy, ensuring that each sensor is functioning properly to provide data support.

[0131] Step 3: Determine the positioning scheme and positioning.

[0132] Sensor status acquisition: The robot has a built-in sensor status monitoring module that acquires the current status values ​​of RTK, LiDAR, binocular vision, IMU and GPS in real time after each sensor completes its self-test, including information such as signal strength, data output frequency and self-test fault codes.

[0133] Path generation scheme and location selection:

[0134] Assuming all sensors are functioning normally: If RTK, LiDAR, binocular vision, IMU, and GPS are all detected to be functioning normally, the robot employs an EKF + factor graph optimization fusion localization scheme. In the short term (within 1 second), RTK and IMU data are fused using Extended Kalman Filter (EKF) to quickly output high-frequency (100Hz) pose information, ensuring real-time motion control. In the medium to long term (over 10 seconds), factor graph optimization incorporates LiDAR point cloud matching errors and binocular vision reprojection errors as constraint factors into the localization solution process, effectively correcting cumulative drift and ensuring a cumulative error of ≤0.5m within 1 hour, resulting in high-precision robot localization.

[0135] Preferably, the short-term positioning process is as follows:

[0136] 1. RTK and IMU positioning data can be acquired in a short period of time. RTK outputs high-precision coordinate data, while IMU calculates attitude angles, velocity, and short-term position changes through acceleration and angular velocity data streams.

[0137] 2. State Prediction: Based on IMU positioning data, using Newton's kinematics model, the robot's pose state (position, velocity, and attitude angle) is predicted at the next moment according to acceleration and angular velocity information. For example, the velocity change is obtained by integrating the acceleration, and then the position change is obtained. Combined with the angular velocity integral, the attitude angle change is obtained, and the pose at time t+Δt is predicted.

[0138] 3. Measurement Update: The pose data output by RTK is used as the measurement value and compared with the predicted pose state. The Kalman gain is calculated (based on the covariance matrix of measurement noise and process noise). Then, the predicted pose state is corrected using the Kalman gain to obtain the fused pose estimate. Since the IMU data frequency is high (500Hz) and the RTK data frequency is relatively low (10Hz), the EKF performs a measurement update every time it receives RTK data. During the interval between two RTK data sessions, state prediction is performed using IMU data, thereby outputting high-frequency (100Hz) pose information to ensure the real-time performance of motion control.

[0139] Preferably, the long-term positioning process is as follows:

[0140] 1. Data Input and Factor Construction:

[0141] The pose sequence data obtained by short-term fusion (EKF) is used as the initial node data for factor graph optimization.

[0142] LiDAR point cloud matching error: By using a point cloud registration algorithm, the matching error between the LiDAR point cloud and the map (or keyframe point cloud) at different times is calculated to obtain the position and attitude constraint factors, which reflect the robot's motion deviation relative to the environment.

[0143] Binocular vision reprojection error: For feature points (such as texture feature points, corner points, etc.) in binocular vision images, calculate their reprojection error (the deviation between the actual projection position and the projection position predicted based on pose estimation) in the image at different times. This is used as a visual constraint factor to reflect the consistency between robot pose estimation and visual observation.

[0144] 2. Factor diagram optimization process:

[0145] Factor graph construction: The robot's pose is used as a node, and the LiDAR point cloud matching error and binocular vision reprojection error are used as edges (constraint factors) to construct a factor graph model. Each node represents the robot's pose state at a certain moment, and the edges represent the constraint relationships between poses at different moments or between pose and environmental observations.

[0146] Optimization Solution: A nonlinear optimization algorithm (such as the Gauss-Newton algorithm) is employed to minimize the sum of errors from all constraint factors. The state values ​​of the pose nodes are adjusted to minimize errors such as LiDAR point cloud matching and binocular vision reprojection. Through iterative optimization, pose deviations caused by IMU drift and short-term fusion error accumulation are corrected. This ensures that the robot's cumulative positioning error is ≤0.5m in the medium to long term (e.g., 1 hour), guaranteeing consistent positioning and allowing the robot to accurately determine its location during long-distance, long-duration rescue missions, thus ensuring accurate path planning and task execution.

[0147] 3. Based on this, after the factor graph optimization is completed, the optimized historical pose endpoint data (such as at time t-Δt) is used as the new initial value for EKF fusion, resetting the state estimation of EKF. For example, assuming that a factor graph optimization is completed at t=10s, resulting in a more accurate pose P_opt at t=10s, then after t=10s, EKF fusion uses P_opt as the initial value to re-fuse the real-time data streams of RTK and IMU, generating high-frequency pose data at time t>10s.

[0148] The optimized historical pose trajectory forms a precise reference trajectory. When real-time EKF generates pose data for t>10s, incremental optimization using factor maps locally constrains the newly generated pose against the historical optimized trajectory. For example, using real-time point clouds from LiDAR, the matching error between the point cloud and the historical optimized trajectory is calculated and used as an incremental constraint to correct the current pose predicted by EKF, ensuring consistency between the current pose and the historical optimized trajectory. This achieves an effect similar to predicting and correcting the current positioning based on the optimized trajectory.

[0149] Preferably, if sensor anomalies exist: Assume the RTK is detected to be malfunctioning due to the burning of some base stations in a fire scenario (signal interruption or extremely low strength), while the LiDAR, binocular vision, IMU, and GPS are functioning normally. The robot, based on a pre-configured list of methods, determines a backup positioning scheme using the normal sensor combination, i.e., IMU + LiDAR fusion positioning (GPS is not prioritized due to its near-ineffectiveness in fire scenarios). The IMU provides short-term inertial navigation data, and the LiDAR obtains relative position information through point cloud matching. The fusion calculation yields the robot's positioning result. Although the accuracy is slightly lower than when the RTK is normal, it maintains the robot's basic positioning and motion control requirements, awaiting RTK signal recovery or changes in the task environment.

[0150] Step 4: Generate global paths.

[0151] Based on the parsed target point information (target point coordinates, etc.), material point information (material location, etc.), and its own positioning information, the robot calls a path planning algorithm to generate a global path.

[0152] Specifically, taking the A* algorithm as an example, using a factory CAD map as the base map data and combining it with a disaster area heat map provided by the server (cloud) (e.g., high-temperature fire zones are marked as high-risk areas, and earthquake-prone collapse zones are marked as areas of corresponding danger levels), a path cost function is constructed: Path cost function = distance × 0.3 + terrain hazard coefficient × 0.7. When searching for a path from the material point to the target point, the algorithm comprehensively considers distance factors and terrain hazard levels, prioritizing paths with shorter distances and passing through areas with low hazard coefficients. Ultimately, it plans the shortest safe path from the material point to the location of the trapped personnel, such as avoiding high-temperature fire zones and earthquake-prone collapse zones, generating a continuous global path trajectory for subsequent robot movement.

[0153] Step 5: Move according to the global path and adapt to the scene.

[0154] Terrain recognition and motion pattern matching: As the robot travels along the global path, the terrain recognition module is activated in real time. The data input layer receives LiDAR point clouds (to obtain terrain height distribution), binocular vision images (to extract texture features), and IMU pose data (to obtain tilt angles), and constructs a three-dimensional terrain feature vector containing four dimensions: slope, roughness, obstacle height, and texture entropy.

[0155] Then, the recognition algorithm layer uses an improved PointNet++ deep learning model to identify and classify the current terrain. This model is pre-trained offline with 10 types of factory disaster terrain samples (flat ground, gravel piles, concrete cracks, etc.). During online recognition, the constructed 3D terrain feature vector is input into the model to quickly determine the terrain type, with an accuracy rate of ≥95% and a recognition latency of ≤200ms.

[0156] If the ground is identified as flat, the motion mode output layer matches the preset motion mode: 1. Use wheel-type movement, limit the linear velocity and angular velocity, such as the forward linear velocity of 0.5m / s, the backward velocity of 0.3m / s, the clockwise rotation angular velocity of 1rad / s, and the turning angular velocity of 30° / s.

[0157] If the terrain is identified as a pile of gravel, the robot moves using a foot-based motion with a step frequency of 1.5Hz, a step length of 30cm, and a continuous movement speed of 1.2m / s. If the terrain is identified as a larger pile of gravel, the robot switches to a foot-based motion mode with a step frequency of 1.0Hz, a step length of 15cm, and enhanced foot cushioning to ensure that the obstacle clearance height is ≤8cm per step. The motion mode parameters are accurately matched according to different terrain types to ensure the robot moves stably.

[0158] The robot can determine the type of disaster scenario (fire or earthquake) in real time.

[0159] Preferably, in a fire scenario, when entering a fire simulation scenario area, the robot automatically adjusts the sensor priority to: LiDAR (anti-smoke interference) > IMU > RTK (some base stations may be burned out, but can still assist if there is a weak signal) > binocular vision (discontinued when the texture recognition accuracy drops to 70% due to smoke) > GPS (basically ineffective).

[0160] Accordingly, in terrain recognition, the system focuses on identifying areas with high-temperature molten materials (by using lidar point cloud temperature inversion, areas ≥80℃ are marked as no-go zones, and the robot automatically avoids them during path planning) and areas with accumulated water / water vapor (using binocular vision to identify reflective features; if detected, the path is adjusted to detour). The movement mode is adjusted to enable a high-temperature protective gait, increasing the step frequency to 1.8Hz to reduce the contact time between the feet and the ground, lowering the risk of injury from high-temperature ground. When encountering burning debris that needs to be crossed, a single-leg support hop mode is used, with a maximum jump distance of 50cm for safe crossing. For obstacle avoidance, the Dynamic Window Method (DWA) is used, adjusting the local path within 0.5s based on sudden obstacles detected by lidar (such as fallen pipes), with an adjustment range ≤2m, ensuring a safe distance of ≥50cm from obstacles and guaranteeing safe movement.

[0161] Preferably, for earthquake scenarios, if entering an earthquake simulation area, the sensor priority is adjusted to: RTK (remaining base stations, if any are operational) > LiDAR (constructs a 3D map of the ruins, providing detailed terrain data for path planning and obstacle avoidance) > IMU (for inertial navigation during frequent signal interruptions, maintaining short-term positioning) > binocular vision (identifies crack edges, assisting in judging dangerous terrain) > GPS (assistance in open outdoor areas, but its effect is weak in indoor earthquake scenarios). Terrain recognition focuses on concrete cracks (width > 30cm marked as needing to be detoured, the robot automatically plans a path to avoid them), sloping ground (slope > 35° marked as dangerous, adjusting the path or movement mode to cope), and the stability of stacked rubble (judged by point cloud variance; variance > 0.1m² indicates an unstable area, the planned path avoids it or a specific movement mode is used to pass through).

[0162] Accordingly, the movement pattern was adjusted to a "test-and-hold" gait in the scree area, first touching the ground with 50% power to test the ground and confirming the stability of the scree before bearing weight and moving forward; when crossing cracks, a bipedal coordination mode was activated, with the front leg probing the ground and the rear leg pushing off, extending the stride length to 60cm to safely cross the cracks. Similarly, the Dynamic Window Method (DWA) was used for local dynamic obstacle avoidance, quickly adjusting the local path to ensure a safe distance in the event of sudden obstacles (such as collapsed building components) that may appear in the earthquake scenario.

[0163] Preferably, in disaster-free scenarios, if the vehicle is in a normal factory area free from disasters during the journey, the default sensor priority (RTK priority, combined with LiDAR, binocular vision, IMU, etc. working in coordination) is maintained, and a conventional cruise mode is adopted to travel according to preset stable motion parameters to ensure transportation efficiency.

[0164] Step 6: Monitoring the status of materials.

[0165] During the robot's movement, the grabbed supplies (such as the aforementioned medical first-aid kit) are monitored in real time. Using a mounting system on the robot, along with pressure sensors and other equipment, parameters such as changes in pressure exerted by the supplies on the mounting system are monitored to determine if there is a risk of the supplies falling.

[0166] For example, when the robot switches its movement mode or crosses obstacles, if the pressure sensor detects abnormal pressure fluctuations and the duration exceeds a certain threshold (such as 0.5s), it is determined that the materials are likely to fall. The robot can automatically adjust its movement posture or reduce its movement speed to ensure the stability of the materials, prevent the materials from falling, and ensure the reliability of the transportation task.

[0167] Step 7: Execute the target task.

[0168] Identifying Trapped Personnel: Once the robot reaches the target area, it activates a sensor array to identify trapped personnel. Visual recognition (binocular vision camera) is used to acquire and analyze images of the target area, identifying the shape, posture, and other features of suspected trapped personnel. Simultaneously, combined with acoustic positioning equipment, the robot determines the precise location of the trapped personnel by transmitting and receiving acoustic signals, achieving accurate identification and positioning of the trapped personnel.

[0169] Material Delivery and Return: Based on the identified location of the trapped personnel, the robot adjusts its posture and uses its robotic arm to deliver supplies, smoothly unloading medical first-aid kits and placing them in a suitable location (within reach of the trapped personnel). It then waits for confirmation from the trapped personnel or rescuers. If a successful signal is received, the robot records the task data (including path information, terrain recognition results, motion mode switching, and material transport status) and returns to a pre-set standby point within the factory, awaiting the next task instruction. If a failure signal is received (e.g., no confirmation signal or an abnormal signal is detected), the robot initiates a backtracking procedure, readjusts its position, and attempts to deliver the supplies again, up to a maximum of three retries (reference value). If all three attempts fail, the robot again records the task data and returns to the standby point for subsequent analysis of the cause of the failure and optimization of the task execution process.

[0170] IV. Testing and Verification

[0171] Specifically, this embodiment involves hardware deployment and testing verification, detailed as follows.

[0172] Hardware deployment: Five RTK base stations are installed at the four corners and the center of the factory. The base stations communicate with the robots using the 5.8GHz wireless band, which has strong anti-interference capabilities.

[0173] The robot's sensor layout: The lidar is mounted on the top to achieve 360° surround view; the binocular camera is placed on the chest with a horizontal field of view of 120°; the RTK antenna is mounted at the highest point on the back; the IMU is placed at the center of gravity of the body; and the GPS module is located at the auxiliary antenna position on the back.

[0174] Simulation test: Construct a digital twin scenario of a factory with an area of ​​1000m² in the simulation environment, simulate various combinations of disaster terrain, and require a navigation success rate of ≥98% and an average path deviation of ≤10cm.

[0175] Field testing: Fire (using burning barrels + smoke generators) and earthquake (artificially piled ruins) simulation scenarios were set up in the factory training field. 50 material transportation tests were conducted for each scenario. The terrain recognition accuracy rate must be ≥92%, the motion mode switching response time must be ≤300ms, the obstacle crossing success rate (15-25cm obstacles) must be ≥95%, and no materials must be dropped throughout the process (verified by using fixed devices in conjunction with motion modes).

[0176] The above detailed embodiments fully demonstrate the entire process of the robot in a factory simulated disaster scenario, from receiving instructions, sensor activation and positioning, path planning, scene adaptation and movement, material monitoring to target point task execution. This verifies the feasibility and effectiveness of the technical solution, which can be promoted and applied in actual factory disaster relief and other scenarios to improve the robot's operational capabilities and reliability.

[0177] The above description is merely a preferred embodiment of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the concept described herein through the above teachings or related technologies or knowledge. Modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.

[0178] It should be noted that, for the sake of simplicity, the foregoing method embodiments are described as a series of actions. However, those skilled in the art should understand that this application is not limited to the described order of actions, as some steps may be performed in other orders or simultaneously according to this application. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions and modules involved are not necessarily essential to this application.

Claims

1. A navigation method for robots in complex scenarios, characterized in that, include: Step 1: Receive and parse the material transportation instruction to determine the location of the material and the target point; Step 2: Activate the multi-source sensor system and perform self-tests on each sensor within the system to obtain the status values ​​of each sensor; the sensors in the multi-source sensor system include RTK measuring instruments, lidar, binocular vision cameras, inertial measurement units, and GPS positioning units; Step 3: Determine and execute the positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result; Step 4: Based on the location of the materials, the location of the target point, and the robot's localization results, call the path planning algorithm to generate a global path; Step 5: While moving along the global path, the terrain recognition module identifies the current terrain to obtain the terrain type; the movement mode corresponding to the terrain type is matched, and navigation is performed according to the movement mode until the global path is traversed. In step 5, the terrain recognition module identifies the current terrain to obtain the terrain type, including: A three-dimensional terrain feature vector is constructed based on multi-sensor data. The three-dimensional terrain feature vector includes slope, roughness, obstacle height, and texture entropy information. The multi-sensor data includes absolute position data from RTK measuring instruments, environmental point cloud data from lidar, depth information from binocular vision cameras, attitude and motion change data from inertial measurement units, and auxiliary positioning data from GPS positioning units. The terrain type is determined based on the three-dimensional terrain feature vector, and the terrain type includes fire scene, earthquake scene and no disaster scene; In step 5, the motion pattern corresponding to the terrain type is matched, and navigation is performed according to the motion pattern until the global path is traversed, including: When the terrain type is a fire scenario, the priority of the sensors is adjusted and the movement mode is set to the first type of movement mode. The first type of movement mode includes: marking areas of molten material with a temperature higher than a preset temperature as no-entry zones, identifying areas of water accumulation or water vapor accumulation and detouring around them; activating high-temperature protective gait and increasing the step frequency; and using a single-leg support jump when crossing burning debris. Navigate according to the first type of motion mode, and dynamically adjust the local path using the dynamic window method until the global path is traversed. In step 5, the motion pattern corresponding to the terrain type is matched, and navigation is performed according to the motion pattern until the global path is traversed, including: When the terrain type is an earthquake scenario, the sensor priority is adjusted and the motion mode is set to the second type of motion mode. The second type of motion mode includes: identifying concrete cracks with an actual width greater than a preset width and marking them as needing to be detoured; identifying sloping ground with an actual slope greater than a preset slope and marking it as dangerous; judging the stability of stacked gravel by the variance of the point cloud data of the lidar, and marking areas with a point cloud data variance greater than a preset variance as unstable areas; adopting a trial-and-support gait in the gravel area, and enabling bipedal coordination when crossing cracks; Navigation is performed according to the second type of motion mode, and the local path is dynamically adjusted using the dynamic window method until the global path is traversed.

2. The navigation method for a robot in complex scenarios according to claim 1, characterized in that, Step 3, which involves determining and executing a positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result, includes: When the state values ​​of all sensors in the multi-source sensor system indicate a normal state, the detection data of each sensor are fused by combining the extended Kalman filter method and the factor graph optimization method to obtain the robot localization result.

3. The navigation method for a robot in complex scenarios according to claim 2, characterized in that, The method of fusing the detection data from each sensor by combining the extended Kalman filter method and the factor graph optimization method to obtain the robot localization result includes: Given that the current time is in the first stage, the robot's pose information is obtained by fusing the RTK measurement instrument detection data and the inertial measurement unit detection data through the Kalman filter method, and this information is used as the robot's localization result. Given that the current time is in the second stage, based on the robot's pose information, the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are used as constraint factors. The robot's pose information is updated through the factor graph optimization method and used as the robot localization result. The first stage and the second stage are two consecutive time periods.

4. The navigation method for a robot in complex scenarios according to claim 3, characterized in that, The process of fusing RTK measurement instrument data and inertial measurement unit data using the Kalman filter method to obtain the robot's pose information includes: Based on the acceleration and angular velocity data output by the inertial measurement unit, the robot's pose state at the next moment is predicted by the Newtonian kinematics model. The coordinate data output by the RTK measuring instrument is used as the measured value. It is compared with the predicted pose state, the Kalman gain is calculated and the predicted pose state is corrected to obtain the fused pose estimate, which is used as the robot's pose information.

5. A navigation method for a robot in complex scenarios according to claim 3, characterized in that, The robot's pose information, based on its own body, uses the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera as constraint factors. The robot's pose information is updated using a factor graph optimization method, including: The robot's pose information obtained in the first stage is used as the initial node data, and the point cloud matching error of the LiDAR and the reprojection error of the binocular vision camera are used as constraint factors. A factor graph is constructed using robot pose as nodes and constraint factors as edges. The Gauss-Newton algorithm is used to minimize the sum of errors of each constraint factor, and the pose nodes are iteratively optimized to obtain the optimized factor graph. Based on the optimized factor graph, the historical pose end data of the first stage is obtained, and the initial value is reset using the historical pose end data. Based on the optimized factor graph, the optimized pose trajectory is determined as the reference trajectory. Based on the initial value, the reference trajectory, and the incremental optimization strategy, incremental constraints are determined. The robot's pose information is predicted and updated through the incremental constraints. The incremental constraints are used to ensure the consistency between the initial value and the subsequent updated value and the reference trajectory.

6. A navigation method for a robot in complex scenarios according to claim 1, characterized in that, In step 3, determining and executing the positioning strategy based on the state values ​​of each sensor to obtain the robot positioning result includes: when at least one sensor in the multi-source sensor system is unavailable, determining a backup positioning strategy after combining the normal state sensors according to the pre-configured sensor list.

7. A navigation method for a robot in complex scenarios according to claim 1, characterized in that, After completing the traversal of the global path, it also includes: The robot uses a fixed device and pressure sensors to monitor the grabbed materials in real time. If the pressure sensor detects abnormal pressure fluctuations that last for more than a preset time, it is determined that the materials are at risk of falling, and the robot's movement posture is adjusted or its movement speed is reduced. Upon reaching the target area, images are captured using a binocular vision camera, and acoustic signals are transmitted and received using a sonic positioning device to analyze the shape and posture characteristics of the trapped personnel. The robot's posture is then adjusted, and the robotic arm unloads the supplies into the reach of the trapped personnel. Wait for confirmation signal, record task data based on confirmation signal result, return to preset standby point or retry delivery; the task data includes path information, terrain recognition results, motion mode switching status and material transportation status.

Images