A method and device for autonomous navigation path planning of a humanoid robot in a complex environment

Abstract

The present invention proposes a method, device and equipment for autonomous navigation path planning of a humanoid robot in a complex environment. The method determines the navigation task by obtaining the current position and path requirements of the robot; constructs a constraint pre-map based on body size constraints, joint motion constraints and dynamic balance constraints, and precalculates the robot's feasible motion space; obtains environmental perception data for change trend analysis, predicts the future motion trajectory of dynamic obstacles and generates a dynamic environment prediction sequence; establishes a two-way planning mechanism for simultaneous forward and reverse path search, and generates multiple candidate path hypotheses through meeting point detection and path fusion; analyzes human behavior patterns to identify potential traffic conflict areas, establishes a social negotiation strategy, and uses intermediate meeting points as social key nodes to adaptively adjust candidate paths; determines the optimal safe path through multi-dimensional safety assessment of fall risk, collision risk and energy consumption risk, and realizes intelligent autonomous navigation of the humanoid robot in a complex environment.

Description

Technical Field

[0001] The present invention relates to the field of robot navigation technology, and in particular to a method, device and equipment for autonomous navigation path planning of a humanoid robot in a complex environment. Background Art

[0002] Humanoid robots, intelligent robots with humanoid morphology and behavioral capabilities, demonstrate enormous potential for application in services, healthcare, education, and other fields. Compared to wheeled or tracked robots, humanoid robots are better able to adapt to the human living environment. However, their complex body structure, multi-joint motion, and dynamic balance requirements also present unprecedented technical challenges for autonomous navigation path planning. Traditional robotic path planning methods, primarily designed for mobile platforms with simple geometries, are unable to effectively address the multi-dimensional constraints and dynamic characteristics of humanoid robots.

[0003] Existing humanoid robot path planning technology suffers from significant technical deficiencies and application limitations: A lack of comprehensive modeling methods for specific constraints such as humanoid robot body size, range of motion, and dynamic balance means that planned paths often violate the robot's physical limitations. Inadequate dynamic environment perception and prediction capabilities make it difficult to handle complex dynamic scenarios such as the flow of people and obstacles. Path search efficiency is low, with traditional one-way search algorithms taking excessively long computational times in large-scale environments. A lack of socially adaptable design means robot behavior doesn't align with human social habits and spatial requirements, easily leading to human-robot conflicts. An incomplete safety risk assessment system prevents comprehensive evaluation of multi-dimensional safety factors such as posture stability, collision risk, and energy consumption risk. These issues severely restrict the practical deployment of humanoid robots in complex, human-occupied environments like hospitals, shopping malls, and office buildings. Summary of the Invention

[0004] The present invention provides a method, device and equipment for autonomous navigation path planning of humanoid robots in complex environments, aiming to solve the problem of intelligent navigation of humanoid robots in complex dynamic environments. It integrates core technologies such as constraint pre-modeling, dynamic environment prediction, two-way path search, social behavior adaptation, and multi-dimensional safety assessment to construct a complete path planning solution for the special needs of humanoid robots, and realizes full-process intelligent management of constraint condition preprocessing, environmental change prediction, search efficiency optimization, social conflict avoidance, and safety risk control, forming a new generation of humanoid robot autonomous navigation path planning technology system with environmental adaptability, social friendliness, safety and reliability.

[0005] A first aspect of the present invention provides a method for autonomous navigation path planning of a humanoid robot in a complex environment, comprising the following steps:

[0006] Obtaining the current position of the robot as a starting position, receiving a path requirement request for a humanoid robot navigation task, and determining a target position according to the path requirement request;

[0007] Based on the humanoid robot's body size constraints, joint motion constraints, and dynamic balance constraints, the robot's feasible motion space is constructed and a constraint pre-map is generated.

[0008] Acquire environmental perception data, perform change trend analysis on the environmental perception data, predict environmental change trajectories in future periods, and generate a dynamic environmental prediction sequence;

[0009] Establishing a bidirectional planning mechanism based on the constraint pre-map and the dynamic environment prediction sequence, performing a forward path search from the starting position to the target position, and simultaneously performing a reverse path search from the target position to the starting position;

[0010] performing path fusion according to intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses;

[0011] identifying potential conflicting areas based on analyzing human behavior patterns in the environment using the environmental perception data, establishing a social negotiation strategy for the environmental perception data and the potential conflicting areas, using an intermediate meeting point between the forward path search and the reverse path search as a social key node, and performing social adaptive adjustment on the multiple candidate path hypotheses based on the social negotiation strategy and the social key node to obtain an adjusted path;

[0012] A multi-dimensional safety risk assessment is performed on the adjusted path, and an optimal safety path is determined based on the multi-dimensional safety risk assessment to complete the autonomous navigation path planning of the humanoid robot.

[0013] A second aspect of the present invention provides a humanoid robot autonomous navigation path planning device in a complex environment, comprising:

[0014] A demand receiving module is used to obtain the current position of the robot as the starting position, receive the path demand request of the humanoid robot navigation task, and determine the target position according to the path demand request;

[0015] The constraint construction module is used to construct the robot's feasible motion space based on the humanoid robot's body size constraints, joint motion constraints, and dynamic balance constraints, and generate a constraint pre-map;

[0016] A prediction and analysis module is used to obtain environmental perception data, perform change trend analysis on the environmental perception data, predict environmental change trajectories in future periods, and generate a dynamic environmental prediction sequence;

[0017] a bidirectional planning module, configured to establish a bidirectional planning mechanism based on the constraint pre-map and the dynamic environment prediction sequence, and to perform a forward path search from the starting position to the target position and a reverse path search from the target position to the starting position;

[0018] a path fusion module, configured to perform path fusion according to intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses;

[0019] a social adjustment module configured to analyze human behavior patterns in an environment based on the environmental perception data to identify potential conflict areas, establish a social negotiation strategy based on the environmental perception data and the potential conflict areas, use an intermediate meeting point between the forward path search and the reverse path search as a social key node, and perform social adaptive adjustment on the multiple candidate path hypotheses based on the social negotiation strategy and the social key node to obtain an adjusted path;

[0020] The path optimization output module is used to perform a multi-dimensional safety risk assessment on the adjusted path, determine the optimal safety path based on the multi-dimensional safety risk assessment, and complete the autonomous navigation path planning of the humanoid robot.

[0021] The third aspect of the present invention proposes a computer device, comprising a memory, a processor, and a computer program stored in the memory and runnable on the processor, wherein when the processor executes the program, it implements the steps of a method for autonomous navigation path planning of a humanoid robot in a complex environment disclosed in the first aspect.

[0022] The beneficial effects of the present invention are reflected in the following points: First, by establishing a constraint pre-map construction technology based on body size, joint movement and dynamic balance, combined with a dynamic environment prediction sequence generation method, it innovatively solves the core technical problem that traditional path planning methods cannot effectively handle the complex constraints and dynamic environmental changes of humanoid robots, and realizes a technological leap from static constraint processing to dynamic constraint prediction, thereby improving the accuracy, adaptability and predictability of path planning, and enabling humanoid robots to maintain stable and reliable navigation performance in complex and changing environments.

[0023] Secondly, through the organic combination of bidirectional planning mechanism and path fusion technology, the problems of low efficiency of one-way search and single candidate path are solved, the search space is compressed and diversified path solutions are generated, and the computational complexity of traditional path planning is reduced from exponential level to square root level. At the same time, multiple differentiated candidate paths are provided to adapt to different task requirements and environmental conditions, which improves the computational efficiency and solution flexibility of path planning.

[0024] Finally, through the deep integration of social negotiation strategies and multi-dimensional safety risk assessments, the intermediate meeting points of two-way search are innovatively identified as social key nodes, and the human-computer interaction strategies in these spatial bottleneck areas are optimized. This systematically addresses the key issues of humanoid robots’ lack of social adaptability and incomplete safety assessments, and establishes a complete technical framework encompassing human behavior analysis, traffic conflict identification, social optimization of meeting points, adherence to social rules, and comprehensive safety evaluation. This achieves a fundamental shift from simple obstacle avoidance navigation to socially friendly navigation, especially in spatially constrained meeting point areas. Through early prediction and focused optimization, the acceptance, safety, and service quality of humanoid robots in human-computer coexistence environments are improved.

[0025] It should be understood that the foregoing general description and the following detailed description are exemplary and explanatory only and are not restrictive of the present application. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] The accompanying drawings herein illustrate specific examples of the technical solutions described in the present invention, and together with the specific implementation methods constitute a part of the specification, and are used to explain the technical solutions, principles and effects of the present invention.

[0027] Unless otherwise specified or defined, the same reference numerals in different drawings represent the same or similar technical features, and the same or similar technical features may also be represented by different reference numerals.

[0028] Figure 1 It is a flow chart of a method for autonomous navigation path planning of a humanoid robot in a complex environment according to the present invention.

[0029] Figure 2 This is a structural block diagram of a humanoid robot autonomous navigation path planning device in a complex environment according to the present invention.

[0030] Figure 3 It is a structural schematic diagram of a computer device of the present invention. DETAILED DESCRIPTION

[0031] In the following description, specific details such as specific system structures and techniques are provided for purposes of illustration rather than limitation to facilitate a thorough understanding of the embodiments of the present application. However, it will be apparent to those skilled in the art that the present application may be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to avoid obscuring the description of the present application with unnecessary detail.

[0032] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, integers, steps, operations, elements and / or components, but does not preclude the presence or addition of one or more other features, integers, steps, operations, elements, components and / or collections thereof.

[0033] It will also be understood that the term "and / or" used in this specification and the appended claims refers to and includes any and all possible combinations of one or more of the associated listed items.

[0034] As used in this specification and the appended claims, the term "if" can be interpreted as "when" or "upon" or "in response to determining" or "in response to detecting," depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "upon determination" or "in response to determining" or "upon detection of [described condition or event]" or "in response to detecting [described condition or event]," depending on the context.

[0035] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0036] References to "one embodiment" or "some embodiments" in this specification mean that a particular feature, structure, or characteristic described in conjunction with that embodiment is included in one or more embodiments of the present application. Thus, phrases such as "in one embodiment," "in some embodiments," "in other embodiments," and "in other embodiments" appearing in various places in this specification do not necessarily refer to the same embodiment, but rather mean "one or more but not all embodiments," unless otherwise specifically emphasized. The terms "including," "comprising," "having," and variations thereof all mean "including but not limited to," unless otherwise specifically emphasized.

[0037] The technical solutions of the embodiments of this application are introduced below.

[0038] like Figure 1 As shown, an embodiment of the present invention provides a method for autonomous navigation path planning of a humanoid robot in a complex environment, comprising the following steps S110 to S170:

[0039] Step S110 , obtaining the current position of the robot as a starting position, receiving a path requirement request for a humanoid robot navigation task, and determining a target position according to the path requirement request.

[0040] Specifically, the robot's current position is determined using multi-sensor fusion positioning technology. This technology combines a GPS global positioning module, an inertial measurement unit (IMU), a visual odometry system, and a lidar simultaneous localization and mapping (SLAM) algorithm to achieve high-precision positioning. The GPS positioning module utilizes RTK differential positioning technology, achieving centimeter-level positioning accuracy in open environments with an update frequency of 10Hz. Integrated Beidou and GLONASS multi-constellation reception enhances positioning reliability. The IMU sensor utilizes a three-axis gyroscope and a three-axis accelerometer, achieving angular velocity accuracy of 0.01 degrees / second and acceleration accuracy of 0.1m / s². High-frequency sampling (200Hz) provides continuous attitude and motion state information, supporting dead reckoning for inertial navigation. The visual odometry module integrates a binocular stereo camera and a monocular camera, employing the ORB-SLAM3 algorithm for real-time feature point extraction and matching. This provides stable relative position estimation indoors and in GPS-denied environments, achieving sub-pixel feature point extraction accuracy and supporting loop closure detection and map relocalization. The LiDAR uses a 360-degree scanning laser rangefinder with a range accuracy of 2cm, a scanning frequency of 40Hz, and an effective detection range of 30 meters. It achieves precise positioning by fusing laser point cloud data with map information through a particle filter algorithm. Multi-sensor data fusion utilizes the Extended Kalman Filter (EKF) algorithm, which weightedly fuses position estimates from different sensors to eliminate the effects of individual sensor errors and noise. Coordinate system conversion and time synchronization are also performed, resulting in a final positioning accuracy of 5cm outdoors and 10cm indoors. After completing precise positioning, the position acquisition module encapsulates and standardizes the robot's three-dimensional position coordinates (X, Y, Z) and attitude angles, automatically marking and storing them as the starting position for the current navigation mission.

[0041] Route request reception utilizes a multimodal human-computer interaction interface and a standardized task description protocol, supporting multiple input methods including voice commands, touchscreen operation, gesture recognition, and remote control commands. Voice command processing integrates a natural language understanding (NLU) module, supporting multiple languages ​​including Chinese and English. With a speech recognition accuracy exceeding 95%, it can interpret natural language navigation requests such as "Please take me to Conference Room A" and "Go to the café on the third floor." Speech recognition utilizes a deep neural network model, supporting joint decoding of acoustic and language models. It can handle voice commands in noisy environments, achieving a noise suppression of 20dB. The touchscreen interactive interface utilizes a graphical task selection method, allowing users to enter their target location through map selection, menu navigation, and shortcut buttons. The interface has a response time of less than 100ms and supports multi-touch and gesture swiping. The gesture recognition system captures the user's hand movements using an RGB-D camera and employs a convolutional neural network (CNN) for gesture classification. It supports the recognition and semantic interpretation of basic gestures such as pointing, waving, and making a fist. The remote control interface supports WiFi, Bluetooth, and 4G / 5G network communications, utilizes the standard ROS (Robot Operating System) message format, and supports task descriptions in XML and JSON data formats, with communication latency of less than 50ms. The task description protocol defines a standardized navigation task format, including fields such as task type (guidance, delivery, inspection, etc.), target description (coordinates, landmarks, relative position, etc.), time constraints, priority, and special requirements. In a hospital nursing assistant application, when a nurse issues a delivery task using the voice command "Please deliver the medicine to Ward 305," the natural language understanding module interprets the task type as "delivery" and the target location as "Ward 305." The task priority analysis system automatically assigns it a "high priority" task. The entire command parsing and task generation process completes within 200ms.

[0042] The target location is determined based on a comprehensive processing flow of semantic map matching, coordinate system conversion and reachability verification, which converts the abstract location description into precise coordinate information that can be understood by the robot navigation system. Semantic map matching adopts an ontology-based map representation method to construct a multi-level map model containing geographic coordinates, semantic labels, functional attributes and association relationships. The map database stores detailed information about the building, including room numbers, functional partitions, floor plans and three-dimensional structure information. Each semantic node contains attributes such as a unique identifier, coordinate information (X, Y, Z and direction angle), reachable path and access rights. The location matching algorithm uses fuzzy matching and similarity calculation to achieve accurate matching by comprehensively evaluating text similarity, spatial relationships and contextual information. The similarity calculation adopts a weighted fusion method: , Where q represents the query semantics entered by the user, c represents the candidate location in the map database, Stext is the text matching score based on edit distance and word semantic similarity, Sspatial is the location relevance score based on spatial proximity, and Scontext is the contextual matching score based on functional attributes and usage scenarios. The weighting coefficients α=0.5, β=0.3, and γ=0.2 are dynamically adjusted based on the actual application scenario, supporting precise matching of fuzzy location descriptions such as "conference room," "café," and "elevator entrance." The coordinate system conversion module converts the semantic matching results into a precise location in the robot's navigation coordinate system. It uses affine and projection transformation algorithms to handle conversions between different coordinate systems, including unified conversions between the GPS coordinate system, the building coordinate system, and the robot's body coordinate system, achieving centimeter-level accuracy. Reachability verification evaluates the reachability and safety of the target location by analyzing the terrain conditions, channel width, obstacle distribution, and the humanoid robot's motion constraints. The verification algorithm considers the robot's body dimensions (height, width, center of gravity height), joint range of motion, and stability requirements. Through geometric constraint checks and physical feasibility analysis, it ensures that the target location meets the humanoid robot's safe navigation requirements. In a hotel service robot application, when a guest requests "towels delivered to room 1208," the target determination system first matches the "1208" room identifier in the hotel map database, obtaining the room coordinates (X:89.2m, Y:156.7m, Floor: 12F). This coordinate conversion is then used to map these coordinates to the target point (X:2847.3, Y:4563.1, Z:36.0) in the robot's navigation coordinate system. The system also verifies that the width of the passage leading to the room is 1.2 meters, meeting the robot's minimum passage width of 0.8 meters, and the floor height is 2.8 meters, meeting the robot's height requirement of 1.7 meters. Finally, the system confirms the reachability and safety of the target location, completing the precise determination of the target location.

[0043] Step S120 , constructing a feasible motion space of the robot based on the body size constraints, joint motion constraints, and dynamic balance constraints of the humanoid robot, and generating a constraint pre-map.

[0044] Specifically, the humanoid robot's pre-constrained map comprehensively analyzes the robot's physical limitations and motion capabilities, precalculating and storing the robot's feasible and prohibited motion areas within the environment. By pre-building this pre-constrained map, infeasible motion areas can be quickly eliminated during the path planning phase, significantly improving path search efficiency and safety.

[0045] In some embodiments, the method of constructing a feasible motion space of a robot based on the body size constraints, joint motion constraints and dynamic balance constraints of a humanoid robot and generating a constraint front map includes: analyzing the body width, height and center of gravity position parameters of the humanoid robot to determine the body size boundary; determining the joint motion boundary based on the joint angle range and motion speed limit of the humanoid robot; analyzing the posture constraint conditions for stable walking according to the dynamic balance algorithm of the humanoid robot; and generating a constraint front map based on the body size boundary, the joint motion boundary and the posture constraint conditions.

[0046] First, the humanoid robot's body width, height, and center of gravity position parameters are analyzed to determine the body size boundaries. Body size analysis uses a three-dimensional geometric modeling method to establish an accurate geometric model based on the robot's CAD design data and actual measurement results. Body width measurements include the maximum shoulder width, waist width, and hip width. A typical humanoid robot has a shoulder width of 45-55cm, a waist width of 35-40cm, and a hip width of 40-45cm. The maximum width value is selected in the size boundary calculation and a safety margin of 5-10cm is added. Body height analysis includes the robot's total height, center of gravity height, and the height distribution of each joint. The total height is typically 160-180cm, and the center of gravity height is approximately 55-60% of the total height. The leg joint distribution includes accurate measurements of the hip, knee, and ankle heights. Center of gravity position parameters are determined through mass distribution analysis and balance testing. These include the static center of gravity position (X, Y, and Z coordinates) and the dynamic center of gravity range. The static center of gravity is typically located at the center of the pelvis, while the dynamic center of gravity swings within ±8-12 cm laterally and ±15-20 cm forward and backward during walking. Body dimension boundaries are generated using an envelope calculation method. The robot's spatial occupancy range in various postures is statistically analyzed to generate a three-dimensional envelope containing the safety boundary. This envelope calculation takes into account the range of joint motion, dynamic swing amplitude, and collision detection requirements, ultimately forming a geometric constraint boundary for spatial occupancy checks.

[0047] Then, based on the humanoid robot's joint angle range and motion speed limits, the joint motion boundaries are determined. This joint motion boundary analysis covers the motion capabilities of all major joints of the humanoid robot, including the neck, shoulder, elbow, wrist, waist, hip, knee, and ankle joints. The angle ranges of each joint are dually constrained by mechanical design limits and safety control algorithms. The neck joint has a left-right rotation range of ±90 degrees and a vertical pitch range of -30 to +45 degrees. The shoulder joint has a fore-and-aft swing range of -45 to +180 degrees and a left-right extension range of ±180 degrees. The hip joint has a fore-and-aft swing range of -30 to +120 degrees and a left-right extension range of ±45 degrees. The knee joint has a flexion range of 0 to +150 degrees. Joint motion speed limits are based on motor performance parameters and safety control requirements. The maximum angular velocity of large joints (such as the hip and knee) is 120-180 degrees / second, and the maximum angular velocity of small joints (such as the ankle and wrist) is 200-300 degrees / second. Joint acceleration is limited to 2-3 times the angular velocity limit. Motion boundary calculation also considers coupling constraints and coordination restrictions between joints. An inverse kinematics algorithm is used to analyze the accessible space of each joint's combined motion and identify inaccessible areas due to joint interference or kinematic singularities. Joint motion boundaries are stored in two forms: a joint space configuration and a Cartesian space reachable domain. The joint space configuration records the angle and speed limits of each joint, while the Cartesian space reachable domain describes the reachable range and motion capabilities of the end effector (such as the hand or foot) in three-dimensional space.

[0048] Next, based on the humanoid robot's dynamic balance algorithm, the posture constraints for stable walking were analyzed. This dynamic balance analysis is based on the zero-moment point (ZMP) theory and the center of mass dynamic stability criterion, establishing a balance constraint model for the humanoid robot. The ZMP calculation utilizes the mass distribution, center of gravity position, and acceleration information of each robot component. The Newton-Euler dynamic equations are used to calculate the location of the zero-moment point within the support polygon. The ZMP must lie within the support polygon to ensure the robot's static and dynamic stability. The support polygon is dynamically determined based on the robot's foot contact. For single-foot support, it is the sole area; for bipedal support, it is the convex polygon formed by the line connecting the two feet. The safety margin of the support polygon is typically indented by 5-10 cm within the actual margin to provide a stability margin. The center of mass dynamic stability analysis predicts the trajectory and acceleration of the robot's center of mass to assess stability risks under given motion patterns. The horizontal acceleration of the center of mass is limited to 0.3-0.5 times the acceleration of gravity, and the vertical acceleration is limited to 0.8-1.2 times the acceleration of gravity. Posture constraints include trunk tilt angle limits, gait parameter constraints, and motion continuity requirements. Trunk tilt angles are limited to ±15 degrees fore-aft and ±10 degrees for side-to-side. Step lengths range from 20 to 60 cm, and cadences range from 0.5 to 2.0 steps per second. The balancing algorithm also integrates predictive and feedback control mechanisms. Through forward-looking gait planning, it predicts stability over the next few steps and dynamically adjusts based on real-time sensor feedback to ensure the robot's walking stability in complex terrain and under external disturbances.

[0049] Finally, a constraint-based pre-map is generated based on the body size boundaries, joint motion boundaries, and posture constraints. This pre-map generation utilizes a multi-layer overlay and constraint fusion approach to integrate the various constraint information into a unified spatial constraint description. Map generation begins with a spatial gridding of the environment, with a grid resolution of 5-10 cm to ensure sufficient spatial accuracy while limiting computational complexity. Each grid cell records the feasibility status and constraint type at that location. Body size constraints are checked using a geometric collision detection algorithm, which intersects the robot's envelope with environmental obstacles and marks grid cells with collision risk as infeasible regions. Joint motion constraint verification is achieved through inverse kinematics and reachability analysis. For each grid location, the required joint configuration required to reach that location is calculated, verifying whether the joint angle and velocity constraints are met. Locations that do not meet these constraints are marked as motion-infeasible regions. Posture constraint verification is achieved using a stability analysis algorithm, which simulates the robot's standing and walking postures at each grid location, calculates the ZMP position and stability margin, and marks locations with insufficient stability as balance risk regions. The constraint fusion process uses Boolean operations and weighted superposition methods to logically combine the results of multiple constraint types to generate a comprehensive feasibility rating. The rating is divided into three levels: feasible, risky, and prohibited. Each level corresponds to a different path planning strategy and safety factor.

[0050] Step S130: Acquire environmental perception data, perform change trend analysis on the environmental perception data, predict environmental change trajectories in future time periods, and generate a dynamic environmental prediction sequence.

[0051] Specifically, environmental perception data is acquired using a multi-sensor fusion architecture, integrating multiple sensors, including lidar, RGB-D camera, ultrasonic sensors, and millimeter-wave radar, to achieve omnidirectional awareness of both static obstacles and dynamic objects in the environment. The lidar utilizes a multi-beam rotational scanning method, with a vertical field of view of ±15 degrees, a horizontal field of view of 360 degrees, a ranging accuracy of ±2 cm, a maximum detection range of 100 meters, and a scanning frequency of 20 Hz, capable of generating high-precision 3D point cloud data. The RGB-D camera integrates color image and depth information acquisition, with an image resolution of 1920×1080 pixels, a depth measurement range of 0.5-8 meters, a depth accuracy of ±1 cm, and a frame rate of 30 fps, supporting object recognition and semantic segmentation. An ultrasonic sensor array, consisting of 16 units, is positioned around the robot, offering a detection range of 0.2-4 meters, a ranging accuracy of ±3 mm, and an update frequency of 40 Hz. It is primarily used for close-range obstacle detection and blind spot correction. Millimeter-wave radar operates at a frequency of 77 GHz, has a detection range of 200 meters, a velocity measurement accuracy of ±0.1 m / s, and an angular resolution of 1 degree. It can penetrate smoke, rain, and snow, and is specifically designed for measuring the velocity and trajectory of dynamic targets. Sensor data fusion utilizes time synchronization and spatial registration technologies. Hardware clock synchronization ensures temporal consistency across sensor data with a synchronization accuracy of 1 ms. Sensor calibration and coordinate transformation achieve spatial unification of data from different sensors. The fused environmental perception data includes multi-dimensional information such as the target's three-dimensional position, geometric dimensions, color and texture, motion speed, and acceleration.

[0052] In some embodiments, the changing trend analysis of the environmental perception data, prediction of the environmental change trajectory in future time periods, and generation of a dynamic environment prediction sequence include: extracting the position information and motion information of dynamic obstacles from the environmental perception data; performing time series analysis on the position information and the motion information to extract motion pattern features; predicting the future motion trajectory of the dynamic obstacle based on the motion pattern features; and arranging the future motion trajectories in a time series to generate a dynamic environment prediction sequence.

[0053] First, the location and motion information of dynamic obstacles are extracted from environmental perception data. Dynamic obstacle detection utilizes background modeling and motion detection algorithms, identifying moving targets by comparing the current frame with the background model. Background modeling utilizes the Gaussian Mixture Model (GMM) algorithm, which is adaptable to lighting variations and environmental noise. Object tracking utilizes the Multi-Object Tracking (MOT) algorithm, which includes three steps: detection, data association, and trajectory management. The detection step uses the YOLO deep learning model to identify dynamic objects such as people, vehicles, and pets. Data association utilizes a combination of the Hungarian algorithm and the Kalman filter, establishing inter-frame object correspondence through position prediction and feature matching. Extracted position information includes the target's 3D coordinates (X, Y, Z), bounding box dimensions (length, width, height), and orientation angle. Position measurement accuracy is ±5 cm, and orientation angle accuracy is ±3 degrees. Motion information extracted includes instantaneous velocity vectors (Vx, Vy, Vz), acceleration vectors (Ax, Ay, Az), and angular velocity. Velocity is measured with an accuracy of ±0.1 m / s and acceleration with an accuracy of ±0.2 m / s². These are calculated from the position sequence using numerical differentiation and filtering smoothing algorithms. Motion information also includes the target's motion type classification, such as linear motion, curved motion, accelerated motion, decelerated motion, and stationary state.

[0054] The position and motion information is then analyzed in a time series manner to extract motion pattern features. This time series analysis utilizes a sliding window approach, with a time window length of 2-5 seconds, corresponding to 40-100 data frames. The data within the window is used to analyze the target's short-term motion pattern. Motion pattern feature extraction encompasses three aspects: trajectory geometry, dynamics, and statistical features. Trajectory geometry includes trajectory length, curvature variation, turning radius, and rate of change of direction. Discrete trajectory points are fitted using cubic spline interpolation to calculate the trajectory's smooth curve and geometric parameters. Curvature calculation accuracy is ±0.01 m⁻¹. Dynamic features include average speed, maximum speed, acceleration variance, and energy consumption. Average speed reflects the target's overall motion intensity, acceleration variance reflects the smoothness of motion, and energy consumption is calculated by integrating motion parameters. Statistical features include motion periodicity, randomness, and predictability. Periodicity is calculated using spectral analysis and autocorrelation functions, randomness is assessed using entropy and information content, and predictability is quantified using linear prediction error and model fit metrics. Motion pattern classification uses machine learning to build a feature library encompassing typical motion patterns such as walking, running, cycling, and driving. Pattern recognition is performed using a support vector machine (SVM) algorithm. Feature extraction also considers environmental context, such as the impact of factors like surface type, traffic regulations, and pedestrian density on motion patterns. This context-aware algorithm improves the accuracy and robustness of feature extraction.

[0055] The future trajectory of the dynamic obstacle is then predicted based on its motion pattern characteristics. Trajectory prediction utilizes a multi-model fusion approach, integrating physical, statistical, and machine learning models. The physical model, based on Newton's laws of motion and motion constraints, assumes the target maintains its current motion state for a short period of time. Linear prediction is performed using constant velocity, constant acceleration, and constant angular velocity models, with a prediction window of 1-3 seconds. The statistical model uses a Kalman filter and an extended Kalman filter to establish a state-space model of the target's motion. Future states are calculated using state estimation and prediction equations. The process noise covariance and observation noise covariance of the filter are set based on actual measurement accuracy, resulting in a state prediction accuracy of ±10 cm. The machine learning model uses a long short-term memory network (LSTM) and a recurrent neural network (RNN) to predict future trajectories by learning from historical trajectory data. The network inputs include position, velocity, acceleration, and motion pattern characteristics, and the output is a sequence of future trajectory points. Prediction accuracy is ±15 cm within 2 seconds and ±25 cm within 3 seconds. Multi-model fusion employs weighted averaging and probabilistic reasoning, dynamically adjusting fusion weights based on the historical prediction accuracy of each model and its matching degree to the current motion pattern. Physical models are weighted more heavily during uniform motion, statistical models are weighted more heavily during variable speed motion, and machine learning models are weighted more heavily during complex motion patterns. Trajectory prediction also considers environmental constraints and interactions, such as road boundary constraints, traffic light effects, and pedestrian avoidance behavior. Constraint optimization and social force modeling are used to refine the predicted trajectory, enhancing its rationality and credibility.

[0056] Finally, the future motion trajectories are arranged in a time series to generate a dynamic environment prediction sequence. The prediction sequence generation uses a discrete time sampling method, with a time step of 0.1 seconds and a prediction time of 5 seconds. Each prediction sequence contains 50 time steps of environmental state information. The environmental state at each time step includes the predicted position, velocity, shape, and confidence level of all dynamic obstacles. Position information is represented as 3D coordinates, and shape information is represented as bounding boxes or polygons. The confidence level reflects the reliability of the prediction result. The prediction sequence also contains uncertainty information, using a covariance matrix and confidence intervals to describe the distribution of prediction errors. The confidence level decreases over time, with a built-in confidence level of 90% for 1 second, 70% for 3 seconds, and 50% for 5 seconds. The dynamic environment prediction sequence is stored in a hierarchical data structure, consisting of a global prediction layer, a local prediction layer, and a detailed prediction layer. The global prediction layer stores a coarse prediction of the large-scale environment, the local prediction layer stores a fine-grained prediction within a 5-meter radius around the robot, and the detailed prediction layer stores a high-resolution prediction of key areas. The prediction sequence supports real-time updates and rolling predictions. When new sensory data arrives, the prediction model parameters and prediction results are automatically updated, ensuring the timeliness and accuracy of the prediction information. In a shopping guide scenario, when a robot detects a customer walking slowly with a shopping cart, the prediction sequence analyzes the customer's walking pattern and the cart's trajectory, predicting that the customer will continue to move approximately 2 meters in the current direction within the next 3 seconds. The robot then adjusts its path accordingly to avoid potential intersections.

[0057] Step S140 , a bidirectional planning mechanism is established based on the constrained pre-map and the dynamic environment prediction sequence, and a forward path search is performed from the starting position to the target position, and a reverse path search is performed from the target position to the starting position.

[0058] Specifically, while traditional one-way search requires traversing the entire search space from the starting point to the target, bidirectional search, through the mutually expanding search trees, reduces the search space to the square root of the original size, significantly improving the efficiency and success rate of path planning. The bidirectional planning mechanism combines the spatial constraints of the pre-map with the temporal constraints of the dynamic environment prediction sequence to achieve spatiotemporal coupled path search optimization.

[0059] In some embodiments, the bidirectional planning mechanism is established based on the constrained pre-map and the dynamic environment prediction sequence, including: setting a starting search node and a target search node in the constrained pre-map; performing a forward path extension search based on the dynamic environment prediction sequence starting from the starting search node; performing a reverse path extension search based on the dynamic environment prediction sequence starting from the target search node; monitoring the node overlap of the forward path extension search and the reverse path extension search, and establishing a bidirectional planning mechanism.

[0060] First, the starting and target search nodes are set in the constrained pre-map. Search node setting utilizes a spatial discretization method, converting continuous spatial coordinates into discrete grid nodes with a grid resolution of 10-20 cm, balancing search accuracy and computational efficiency. The starting search node is set based on the robot's current position coordinates. Coordinate quantization and grid mapping are used to convert the precise position coordinates (X, Y, θ) into corresponding grid indices (i, j, k), where θ represents the robot's heading angle, discretized to an accuracy of 15 degrees, corresponding to 24 directional components. During node setting, feasibility verification is performed to check whether the starting node is within the feasible region of the constrained pre-map. If the starting position is in a prohibited or risky area, the nearest feasible node is searched around it, serving as the actual starting search node. The search radius is set to 50 cm. The target search node is set based on the determined target position, using the same spatial discretization method and feasibility verification process to ensure the reachability and safety of the target node. Node attributes include spatial coordinates, grid index, constraint type, feasibility level, and adjacency. Adjacency relationships are established using either 8- or 26-neighborhood connections, supporting diagonal movement and multi-directional expansion. Nodes also contain time attributes, recording access timestamps and estimated arrival times, supporting spatiotemporal path search. The data structure for search nodes utilizes a combination of a priority queue and a hash table. The priority queue maintains the search expansion order, while the hash table allows for fast node status queries and avoids duplicate accesses. The access time complexity of this data structure is O(log n).

[0061] Then, starting from the starting search node, a forward path expansion search is performed based on the dynamic environment prediction sequence. This forward search utilizes an improved A* algorithm, combining heuristic functions and dynamic constraints for intelligent search. The A* algorithm selects the optimal expansion node using the evaluation function f(n) = g(n) + h(n), where g(n) is the actual cost from the starting node to the current node, and h(n) is a heuristic estimate. The heuristic function h(n) uses a weighted combination of Euclidean and Manhattan distances to calculate the estimated distance from the current node to the target node. The weight coefficients are dynamically adjusted based on the complexity of the environment, with Euclidean distance receiving a higher weight in open environments and Manhattan distance receiving a higher weight in narrow passages. The cost function g(n) consists of three components: movement cost, time cost, and risk cost. The movement cost is calculated based on the path length, with a cost of 1.0 for straight-line movement and 1.414 for diagonal movement. The time cost is calculated based on the robot's velocity and acceleration constraints, and the risk cost is calculated based on the risk level of the constraint-based map and the collision probability predicted by the dynamic environment. During the search expansion process, dynamic constraint checks are performed, using dynamic environment prediction sequences to verify the safety of path nodes at corresponding time points. If the prediction indicates the presence of a dynamic obstacle at a certain time point, the node is marked as temporarily infeasible and the search strategy is adjusted. Forward search supports multi-resolution expansion, using a coarse-grained search for rapid approach when the target is far away, and switching to a fine-grained search to ensure accuracy when approaching the target. During the search process, an open list and a closed list are maintained. The open list stores candidate nodes to be expanded, while the closed list stores already visited nodes. This list management avoids cyclic searches and repeated calculations.

[0062] Next, a reverse path extension search is performed, starting from the target search node and based on the dynamic environment prediction sequence. This reverse search employs a time-reversal search strategy, performing path planning in reverse order from the target time point to the starting time point. The reverse search heuristic function h(n) calculates the estimated distance from the current node to the starting node, using the same distance metric as the forward search, but in the opposite direction. The temporal processing of the reverse search is more complex, requiring consideration of the time reversal of the dynamic environment prediction sequence. The forward time series is converted to a reverse time series through a time axis transformation, ensuring that temporal constraint checking during the reverse search is consistent with that during the forward search. The cost function design of the reverse search takes into account the a posteriori optimization properties of the path, assigning higher weights to path segments close to the starting location, encouraging the search algorithm to prioritize the initial portion of the path. The reverse extension process employs the same dynamic constraint checking mechanism, but the temporal reference point is the corresponding instant on the reverse time axis. A time mapping function is used to establish the correspondence between forward and reverse time. The reverse search supports a target-guided optimization strategy, leveraging the geometric characteristics of the target location and environmental constraints to guide the search direction, prioritizing path branches that are more likely to connect to the starting location. The reverse search process maintains separate open and closed lists, independent of the forward search data structure, to prevent interference between search states. The reverse search also considers the reversibility constraints of the robot's motion. Certain motion patterns may be difficult or infeasible when executed in reverse. Reversibility analysis ensures the physical feasibility of the reverse path.

[0063] Finally, the node overlap of the forward path expansion search and the reverse path expansion search is monitored, and a bidirectional planning mechanism is established. Node overlap detection uses a spatial index and time window matching method to identify the intersection of the two search trees by comparing the node coordinates and time information generated by the forward search and the reverse search. The spatial tolerance of the overlap detection is set to the size of a grid unit, and the time tolerance is set to 0.2-0.5 seconds, allowing node matching within a certain error range. The overlap detection algorithm uses a bidirectional hash table query. When the forward search expands a new node, it queries the node set of the reverse search for a matching node. The reverse search uses the same query strategy, and the query time complexity is O(1). The quality assessment of overlapping nodes uses a multi-criteria evaluation method, including indicators such as total path length, total time cost, total risk cost, and path smoothness. The comprehensive score of the overlapping nodes is calculated by weighted summation. The bidirectional planning mechanism also supports multiple overlapping point detection. When there are multiple overlapping nodes, the node with the best comprehensive score is selected as the best connection point, while the suboptimal node is retained as an alternative. During the overlap detection process, path connectivity verification is performed to ensure that the forward and reverse path segments can be smoothly connected at the overlap points. Connectivity verification includes checks for position continuity, velocity continuity, and acceleration continuity. A dynamic search strategy adjustment mechanism is established in the bidirectional planning mechanism. When no overlapping nodes are detected for a long time, the search parameters and expansion strategies are automatically adjusted to increase the search coverage and exploration depth. In the navigation task of a large shopping mall, when the robot needs to reach a designated store on the third floor from the elevator entrance on the first floor, the forward search starts exploring feasible paths from the elevator entrance, and the reverse search starts reverse planning from the target store. The bidirectional search detects overlapping nodes in the atrium area on the second floor and quickly determines two feasible paths through the stairs and escalators.

[0064] The forward path search in the bidirectional planning mechanism serves as the starting search component, expanding the path from the starting location toward the target location while maintaining information synchronization and strategy coordination with the reverse search. The forward search utilizes a modified A* algorithm combined with a bidirectional optimization strategy. The heuristic function not only considers the distance to the target location but also dynamically adjusts the search weight and expansion direction by referencing the progress of the reverse search in real time. During the forward search process, a shared node pool is established, and node information generated by the search is passed to the reverse search in real time, supporting overlap detection and optimizing search efficiency. The search strategy utilizes an adaptive expansion mechanism. When the reverse search detects rapid progress in a certain area, the forward search increases the expansion intensity in that direction, promoting a rapid convergence of the two search trees. The forward search cost function integrates bidirectional optimization objectives, evaluating not only the distance cost to the target but also the possibility of connecting to the reverse search tree, prioritizing exploration of node branches that are more conducive to forming a complete path. In a shopping mall navigation scenario, when the forward search starts exploring from the entrance hall, it receives path information from the target store in real time from the reverse search. It finds that the reverse search expands faster in the second-floor area. The forward search immediately adjusts its strategy to prioritize exploring the path to the second floor, significantly improving the convergence speed of the two-way search.

[0065] The reverse path search in the bidirectional planning mechanism serves as the target-side search component, performing reverse path expansion from the target location toward the starting location, forming a collaboratively optimized search pair with the forward search. The reverse search utilizes a time reversal strategy and a bidirectional coordination algorithm. It obtains real-time status information from the forward search through a shared node pool, dynamically adjusting its own search strategy and expansion priorities. The reverse search heuristic function comprehensively considers the distance from the starting location and the progress of the forward search. When the forward search has a high expansion density in a certain area, the reverse search will intensify its exploration of that area, promoting rapid convergence of the search tree. The reverse search handles complex time-constrained inversion problems, transforming the dynamic environment prediction sequence on a time axis to ensure temporal consistency and compatibility between the reverse path planning and the forward search. The bidirectional coordination mechanism enables the reverse search to adjust its expansion strategy based on real-time feedback from the forward search. When the forward search encounters a complex obstacle area, the reverse search will proactively approach it to find a more optimal connection path. The reverse search also performs path quality pre-assessment, utilizing the environmental characteristics and constraints of the target location to provide path optimization recommendations for the forward search. In the hospital corridor navigation task, the reverse search starts from the ward location and is planned towards the nurse station. The forward search status at the corridor intersection is monitored in real time. When the forward search is found to be expanding slowly due to obstruction by medical equipment, the reverse search actively expands to the intersection area. Through two-way information synchronization and strategy coordination, it provides optimized search results for subsequent path fusion processing.

[0066] Step S150 : performing path fusion based on the intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses.

[0067] Specifically, through intelligent meeting point detection and path connection technology, the path fragments generated by the forward search tree and the reverse search tree are organically combined to form a complete navigation path. The multiple candidate paths generated provide a rich selection space for subsequent path optimization and risk assessment.

[0068] In some embodiments, the generating multiple candidate path hypotheses by performing path fusion based on the intermediate meeting points of the forward path search and the reverse path search includes: detecting the optimal meeting point of the forward path search and the reverse path search; connecting the forward path search and the reverse path search based on the optimal meeting point to form a primary candidate path; analyzing the suboptimal meeting point of the forward path search and the reverse path search; generating an alternative candidate path based on the suboptimal meeting point, and combining the alternative candidate path with the primary candidate path to form multiple candidate path hypotheses.

[0069] First, the optimal meeting points for both the forward and reverse path searches are detected. This optimal meeting point detection utilizes a multi-criteria optimization algorithm, comprehensively considering multiple evaluation metrics, including path length, time cost, connection quality, and safety. A set of candidate meeting points is generated through a spatial proximity search. Within a search radius of 50-100 cm, node pairs from the forward and reverse search trees are searched for within this radius to form a set of potential meeting points. Meeting point evaluation utilizes a weighted scoring method, with a weight of 0.3 for path length, 0.25 for time cost, 0.25 for connection smoothness, and 0.2 for safety margin. These weights can be adjusted based on specific task requirements. Path length evaluation calculates the total length of the forward and reverse path segments, prioritizing meeting points with shorter total distances. Time cost evaluation considers the robot's movement time and acceleration / deceleration times across different path segments, including the time costs of turning, obstacle avoidance, and waiting. Connection smoothness evaluation evaluates the smoothness of the path connection by calculating the azimuth, velocity, and acceleration differences at the meeting point. Meeting points with azimuth differences less than 30 degrees and velocity differences less than 0.5 m / s receive higher scores. Safety margin assessment analyzes the distribution of obstacles and dynamic environmental risks around the meeting point, selecting a location with sufficient safety distance and low collision risk. The optimal meeting point is determined by ranking based on a comprehensive score, with the highest-scoring meeting point selected as the optimal connection location. In the airport terminal navigation task, optimal meeting point detection identified three candidate meeting points in the central area of ​​the terminal. Through comprehensive evaluation, a location close to the boarding gate, with moderate surrounding traffic density and a relatively flat connection angle was selected as the optimal meeting point.

[0070] The forward and reverse path searches are then connected based on the optimal meeting point to form a primary candidate path. Path connection utilizes a cubic spline interpolation algorithm to achieve a smooth transition between the forward and reverse path segments at the optimal meeting point. The connection algorithm first extracts path points within 1-2 meters before and after the meeting point as interpolation control points. Boundary conditions are set to ensure position, velocity, and acceleration continuity at the connection. The spline curve parameters are determined by minimizing the curvature and acceleration changes, generating a smooth trajectory that conforms to the robot's kinematic characteristics. The length of the connection segment is controlled between 0.5 and 1.5 meters to ensure smooth connection while avoiding excessive deviation from the original path. Feasibility verification is performed during the path connection process, using inverse kinematics and dynamic simulation to verify that the connected path meets the robot's joint angle, velocity, and acceleration constraints. Connection quality is assessed by calculating the curvature distribution, velocity change rate, and acceleration continuity of the connected segment to ensure that the generated primary candidate path has good kinematic characteristics. The primary candidate path also needs to pass collision detection and safety verification. The safety of the path is checked using a constrained pre-map and dynamic environment prediction sequence, and local optimization adjustments are performed on path segments with safety risks. During the nursing robot's navigation in a hospital corridor, a primary candidate path with a total length of 42 meters and an estimated travel time of 85 seconds was generated based on the path connection based on the optimal meeting points. The maximum curvature at the path connection was 0.8 m⁻¹, and the maximum lateral acceleration was 1.2 m / s², meeting the requirements for smooth operation of the nursing robot.

[0071] Next, we analyze the suboptimal meeting points from both forward and reverse path searches. This analysis employs a ranking and differentiation evaluation method to select alternative connection locations with varying characteristics and advantages from a candidate set of meeting points. The selection criteria for suboptimal meeting points include dissimilarity from the optimal meeting point, independent optimization objectives, and specific environmental adaptability. The dissimilarity evaluation calculates the degree of difference between the candidate meeting points and the optimal meeting point in terms of spatial location, path characteristics, and performance metrics. Meeting points with the highest degree of dissimilarity are selected as suboptimal candidates, ensuring a diverse range of path options. Independent optimization objectives include shortest path, lowest energy path, highest safety path, and best comfort path, with each suboptimal meeting point corresponding to a specific optimization objective. Specific environmental adaptability considers the flexibility and robustness of the meeting point in responding to emergencies, environmental changes, and mission adjustments, selecting meeting points with strong adaptability as alternatives. The number of suboptimal meeting points is controlled between 2 and 5, and is dynamically adjusted based on environmental complexity and mission requirements. Suboptimal meeting point assessment utilizes the Analytic Hierarchy Process (AHP) to establish a multi-tiered evaluation system encompassing efficiency, safety, comfort, and flexibility. Expert knowledge and statistical analysis determine the weighting of each tier. In a nursing home service robot application, the suboptimal meeting point analysis identified three alternative connection locations: the most comfortable path through the sunroom, the safest path along the wall, and the most flexible path to avoid mealtime crowds. This provides adaptable options for different service scenarios.

[0072] Finally, alternative candidate paths are generated based on the suboptimal meeting points, and combined with the main candidate paths to form multiple candidate path hypotheses. The generation of alternative candidate paths uses the same connection algorithm and optimization process as the main candidate paths, but adjusts parameters for different meeting points and optimization goals. An alternative candidate path is generated for each suboptimal meeting point, and the specific advantages and applicable scenarios of the meeting point are emphasized during the path generation process. The quality control of alternative paths ensures that all alternative paths meet basic safety and executable requirements by setting minimum performance thresholds and feasibility conditions. Path deduplication and similarity checking calculate the similarity index between paths, eliminate highly similar redundant paths, and retain candidate paths with obvious differences and unique values. The diversity of the candidate path set is evaluated by statistical analysis of path length distribution, time distribution, risk distribution, and energy consumption distribution, and normalized variance is used as the diversity measurement indicator. Where: F1, F2, F3, and F4 are the length, time, risk, and energy consumption characteristics, respectively; Var(Fi) is the variance of the i-th characteristic; Fi is the mean of the i-th characteristic; and wi is the characteristic weight. This metric quantifies the diversity of the path set by calculating the relative dispersion of each characteristic. A larger Ddiversity value indicates greater differences in the performance characteristics of the candidate paths, ensuring that the path set covers different performance requirements and environmental conditions. Path labeling and attribute settings assign descriptive labels to each candidate path, including path type (optimal, safe, fast, comfortable, etc.), applicable scenarios, performance characteristics, and recommendation priority, to facilitate subsequent path selection and decision-making. The candidate path data structure uses a hierarchical storage structure, with the primary candidate path given the highest priority and alternative candidate paths sorted by performance score, supporting fast query and dynamic adjustment. For example, in the room service of a hotel service robot, the alternative candidate paths generated based on four suboptimal meeting points include the best experience path through the landscape corridor, the most unobstructed path avoiding the cleaning operation area, the most convenient path passing through the service desk for easy coordination, and the safest path with reserved emergency avoidance space. Combined with the main candidate path, a candidate path hypothesis set containing five different characteristic paths is formed, providing a rich path selection for personalized room service.

[0073] Step S160: Analyze human behavior patterns in the environment based on the environmental perception data to identify potential conflict areas, establish a social negotiation strategy based on the environmental perception data and the potential conflict areas, use the intermediate meeting points between the forward path search and the reverse path search as social key nodes, and perform social adaptive adjustments on multiple candidate path hypotheses based on the social negotiation strategy and the social key nodes to obtain an adjusted path.

[0074] Specifically, environmental perception data is used to analyze human behavior patterns within the environment to identify potential conflict zones. Multi-layered behavior recognition and spatial analysis technologies are employed to construct a human behavior prediction model and conflict risk assessment mechanism. Human behavior pattern recognition integrates visual recognition, motion tracking, and behavior analysis algorithms to extract human posture features, motion trajectories, and behavioral intentions. Posture recognition uses skeletal point detection technology to identify the spatial position and motion state of the head, shoulders, arms, torso, and legs. Key point detection accuracy reaches pixel-level, supporting simultaneous tracking and recognition of multiple individuals. Motion pattern analysis processes time series data to identify basic behavioral patterns such as walking, running, turning, stopping, and carrying, and then establishes a behavioral state transition model to predict behavioral trends. Intent recognition combines gaze direction, body orientation, movement speed, and environmental context to infer a person's travel goals and path preferences. The intention prediction window is set to 2-5 seconds. Conflict area identification uses spatiotemporal occupancy analysis to project the predicted trajectories of humans and robots onto a spatiotemporal grid, identifying areas of spatiotemporal overlap as potential conflict zones. The spatial scope of the conflict zone is set to an elliptical area with a radius of 1.5-2.0 meters around the human body, and the time range covers a prediction window of 3-8 seconds into the future. The conflict probability calculation takes into account the uncertainty of trajectory prediction and the randomness of human behavior, quantifying the likelihood of conflict through a probability distribution model. The conflict probability threshold is set at 0.3-0.6, and areas exceeding the threshold are marked as high-risk conflict areas. In the hospital outpatient lobby scenario, behavioral analysis identifies different behavioral patterns, such as patients walking slowly, family members searching quickly, and medical staff passing through urgently. Multiple potential conflict areas are identified at the intersection of elevator entrances, registration windows, and main corridors, providing an important environmental cognitive foundation for the robot's social navigation.

[0075] In some embodiments, establishing a social negotiation strategy for the environmental perception data and the potential conflict area includes: analyzing human traffic priority and space requirements based on the environmental perception data and the potential conflict area; determining the avoidance timing and avoidance method of the humanoid robot according to the traffic priority; generating an avoidance strategy, a waiting strategy and a collaborative traffic strategy based on the avoidance timing, the avoidance method and the space requirement; and combining the avoidance strategy, the waiting strategy and the collaborative traffic strategy to form a social negotiation strategy.

[0076] Exemplarily, the analyzing the traffic priority and space requirement of humans based on the environmental perception data and the potential traffic conflict area includes: identifying the travel direction and travel speed of humans in the traffic conflict area based on the environmental perception data; judging the urgency of the traffic of humans according to the travel direction and the travel speed; analyzing the items carried by humans and the physical condition of humans based on the environmental perception data to determine the space occupation requirement of humans; and determining the traffic priority and space requirement of humans based on the travel urgency and the space occupation requirement.

[0077] First, based on environmental perception data and potential conflict zones, the system analyzes human traffic priorities and space requirements. Based on environmental perception data, the system identifies the direction and speed of traffic within conflict zones. Human traffic is detected and tracked using RGB-D camera and LiDAR fusion data. Optical flow and Kalman filtering are used to calculate the motion vector of the human center of mass. The direction of traffic is calculated from the angle of position change between consecutive frames, and the speed is calculated as the ratio of displacement distance to time interval. The system then determines the urgency of traffic based on direction and speed, establishing an urgency assessment system. Rapid linear motion is classified as high urgency, slow motion as low urgency, and moderate speed with frequent directional changes as medium urgency. The system then analyzes the human's belongings and physical condition based on environmental perception data to determine space requirements. Object detection algorithms are used to identify assistive devices such as wheelchairs, strollers, luggage, and crutches, as well as items carried. Human posture estimation is then used to analyze physical condition. Overall space requirements are determined based on the basic dimensions of the human body, the size of its belongings, and the space occupied by assistive devices. Human traffic priority and space requirements are determined based on urgency and space requirements. A comprehensive assessment method weights urgency, special physical conditions, and social roles to generate a priority score. The final space requirement range is calculated by combining basic space, object space, motion compensation space, and safety buffer space. In a hospital emergency department, when a medical worker is detected pushing an emergency cart through a corridor at high speed, their high-speed linear motion pattern is identified and judged as the highest urgency. Combined with the cart's large footprint, the robot is ultimately determined to have the highest traffic priority and a large space requirement. Based on this, the robot proactively avoids the worker and reserves ample space for passage.

[0078] The humanoid robot then determines the avoidance timing and method based on the priority of the traffic flow. Avoidance timing decisions utilize a combination of predictive and responsive avoidance strategies. Predictive avoidance pre-plans avoidance maneuvers based on predicted human behavior, while responsive avoidance makes immediate adjustments based on real-time perception information. Trigger conditions for avoidance timing include conflict probability thresholds, temporal distance thresholds, and spatial distance thresholds. The avoidance mechanism is activated when the conflict probability exceeds 0.4, the expected encounter time is less than 4 seconds, or the spatial distance is less than 3 meters. The avoidance method selection considers the geometry of the conflict, environmental constraints, and social etiquette requirements, primarily including lateral avoidance, slowdown and waiting, stopping and yielding, and reversing paths. Lateral avoidance is suitable for scenarios with ample lateral space, with a set avoidance distance of 0.8-1.5 meters to ensure ample space for human traffic. Slowdown and waiting is suitable for scenarios with a short expected conflict time. The robot reduces its movement speed by 50-80%, delaying its arrival at the conflict zone. Stopping and yielding is suitable for scenarios involving high-priority traffic. The robot completely stops and maintains a polite waiting posture. Path detour is used when conflicts persist for extended periods or when space is severely limited. The robot proactively chooses an alternative path to avoid the conflict zone. The avoidance strategy is selected using a decision tree algorithm, which makes a multi-level judgment based on factors such as priority difference, space availability, time urgency, and social appropriateness. In a shopping mall environment, when the robot encounters a customer pushing a stroller, it uses a high priority score of 8 and chooses a sideways avoidance strategy, maintaining a safe distance of 1.2 meters while reducing its speed to avoid noise interference.

[0079] Next, avoidance strategies, waiting strategies, and cooperative traffic strategies are generated based on the avoidance timing, avoidance method, and space requirements. Avoidance strategy generation utilizes a combination of path replanning and motion parameter adjustment to calculate specific avoidance trajectories and motion control parameters based on the selected avoidance method. The lateral avoidance strategy generates a smooth S-shaped avoidance trajectory through lateral displacement planning. The maximum lateral acceleration and maximum lateral speed of the avoidance trajectory are limited to 1.0 m / s² and 0.5 m / s², ensuring smooth and safe avoidance maneuvers. The deceleration avoidance strategy achieves time coordination through longitudinal speed control, with a deceleration rate set at 0.5-1.5 m / s² and a minimum speed limit of 0.2 m / s to avoid traffic congestion caused by a complete stop. The waiting strategy is applicable to scenarios requiring stopping to yield, and includes three elements: waiting location selection, waiting posture control, and waiting time management. Waiting location selection is based on the principles of non-blocking traffic, easy observation, and easy resumption of movement. Walls, pillars, or designated waiting areas are preferred. Waiting posture control involves adjusting body orientation, arm position, and head direction, using a friendly waiting posture to convey the robot's polite intentions. Waiting time management sets a maximum waiting time of 30-60 seconds, after which an alternative strategy or route replanning is initiated. The collaborative passage strategy is suitable for scenarios where space is limited but coordinated passage is possible, achieving coordinated passage through interactive communication with humans. Collaborative passage includes three mechanisms: visual signal interaction, speed synchronization, and spatial coordination. Visual signals convey the robot's movement intentions through LED lights, displays, or gestures. Speed ​​synchronization ensures that the robot and human maintain a coordinated passage rhythm. Spatial coordination ensures sufficient passage space for both parties through real-time adjustments.

[0080] Finally, the avoidance strategy, waiting strategy, and collaborative passage strategy are combined to form a social negotiation strategy. This strategy combination utilizes a hierarchical decision-making and dynamic switching architecture, intelligently selecting and smoothly switching between different strategies based on real-time conditions and environmental changes. Strategy weights are dynamically adjusted based on environmental complexity, crowd density, and task urgency. Avoidance strategies are given higher weight in sparsely populated environments, while waiting strategies and collaborative passage strategies are given higher weight in crowded environments. The strategy switching mechanism sets state transition conditions and switching thresholds, automatically switching to a more appropriate strategy when environmental conditions change significantly. The social negotiation strategy also incorporates learning and adaptation mechanisms, continuously optimizing strategy parameters and decision-making logic by recording and analyzing human responses to the robot's behavior. Strategy evaluation utilizes comprehensive metrics such as human satisfaction, passage efficiency, and safety, establishing a feedback mechanism for strategy optimization. In nursing robot applications in nursing homes, the social negotiation strategy prioritizes waiting strategies and collaborative passage strategies based on the behavioral characteristics and psychological needs of the elderly, avoiding rapid avoidance maneuvers that could startle them. At the same time, the social negotiation strategy effectively communicates with the elderly through gentle voice prompts and friendly body language, creating a harmonious human-robot interaction environment.

[0081] Based on the social negotiation strategy and social key nodes, social adaptive adjustments are made to multiple candidate path hypotheses to obtain adjusted paths. Path re-evaluation and local optimization techniques are used to integrate the requirements of the social negotiation strategy into the optimization objectives of path planning. In particular, the intermediate meeting points between the forward path search and the reverse path search are identified as social coordination locations that require special attention, because meeting points are usually located in areas with narrow space or limited vision, which are high-incidence areas for human-machine interaction conflicts. Social adaptive adjustment first conducts a social risk assessment on the candidate paths, analyzes the advantages and disadvantages of each path in social negotiation, and assigns a higher assessment weight to the meeting point area. This is because: (1) Meeting points are often turning points or spatial bottlenecks of the path. Identifying and handling social risks in these areas in advance can avoid conflicts between robots and humans at the narrowest points; (2) Appropriate social adjustments at the meeting points have the greatest impact on the overall path quality. A successful meeting point coordination can avoid subsequent chain adjustments; (3) Humans have higher psychological expectations at these key locations, and good social performance can significantly improve user experience. Social risk is quantified by integrating the risk density of each point on the path: , where ρ(s) is the crowd density at path point s, Pconflict(s) is the potential conflict probability, Cetiquette(s) is the social etiquette compliance (between 0 and 1), Icrowd(s) is the crowd influence indicator, w1, w2, and w3 are risk weight coefficients, and Kmeet(s) is a weighted function for the neighborhood of the meeting point, taking values ​​of 2-3 near the meeting point and 1 elsewhere. By increasing the risk weight of the meeting point, the system prioritizes optimizing the social performance of these key areas. This integral calculates the accumulated social risk value along the entire path, with higher risk values ​​indicating poorer social adaptability. Social risk assessment encompasses multiple dimensions, including human-robot conflict probability, social etiquette compliance, crowd flow impact, and psychological comfort. Quantified scores are used to identify path segments requiring adjustment and areas for improvement. Path adjustment utilizes a combination of local replanning and global optimization, prioritizing social risks near the meeting point. This prioritization approach offers the following advantages: reduced computational resource consumption (focused optimization on key areas), improved adjustment efficiency (a single adjustment addresses major issues), and improved user perception (optimal performance is achieved at locations most likely to generate negative impressions). For encounter areas, the robot begins adjusting its motion strategy 5-8 meters in advance, including reducing speed, adjusting posture, and reserving more social space. Small path corrections are applied to localized social risk points, and paths with poor overall social adaptability are replanned. Path continuity and executability are maintained during the adjustment process, and smooth interpolation and constraint checking ensure that the adjusted path meets the robot's kinematic requirements. Social adaptability optimization considers different social scenarios and cultural backgrounds, establishing a configurable library of social rules and behavior patterns. Specific social strategies are formulated specifically for the spatial characteristics of encounter points (such as corridor intersections, elevator entrances, and doorways). This targeted design reduces the probability of human-robot conflict and supports personalized adjustments for different environments and tasks. Path adjustment also integrates social coordination in the temporal dimension, adjusting the robot's movement speed and behavior timing to achieve coordination with the rhythm of human activities. In the application of the welcoming robot in the hotel lobby, social adaptability adjustment changes the original shortest path that goes straight through the center of the lobby to a path that detours along the edge, avoiding the guests' rest and conversation areas. In particular, the three-stage social strategy of "early deceleration - lateral avoidance - polite waiting" is adopted in the meeting point area in the center of the lobby. At the same time, an arc approach trajectory is adopted when approaching the target guest, completing the welcoming task in a more friendly and polite manner. Although the adjusted path increases, it significantly improves guest satisfaction and service quality.

[0082] Step S170 , performing a multi-dimensional safety risk assessment on the adjusted path, determining the optimal safety path based on the multi-dimensional safety risk assessment, and completing the autonomous navigation path planning of the humanoid robot.

[0083] Specifically, a multi-dimensional safety risk assessment analyzes the fall risk, collision risk, and energy consumption risk of the adjusted path, establishing a comprehensive safety evaluation system to quantitatively assess the safety of each path. Based on the evaluation results, the path with the highest overall safety score is selected as the optimal navigation solution, ensuring that the robot can safely and reliably complete its autonomous navigation mission.

[0084] In some embodiments, the multi-dimensional safety risk assessment of the adjustment path includes: analyzing the posture stability of the humanoid robot in the adjustment path, and evaluating the fall risk level; detecting the safe distance between the adjustment path and obstacles, and evaluating the collision risk level; estimating the energy consumption and time cost of the joint movement of the adjustment path, and evaluating the energy consumption risk level; and performing a comprehensive safety score based on the fall risk level, the collision risk level, and the energy consumption risk level to complete the multi-dimensional safety risk assessment.

[0085] First, the posture stability of the humanoid robot during the adjustment path is analyzed to assess its fall risk. The posture stability analysis establishes a stability assessment model based on the robot's center of gravity, support polygon, and dynamic balance conditions. The center of gravity position analysis calculates the mass distribution and spatial position of each robot component to determine the trajectory of the robot's center of gravity under different motion states. The center of gravity trajectory must always lie within the support polygon to ensure static stability. The support polygon is dynamically determined based on the contact between the robot's feet and the ground. It is the contour of the sole of the foot in single-foot support and the convex polygon formed by the line connecting the two feet in bipedal support. The dynamic balance analysis uses the zero moment point (ZMP) theory to calculate the ZMP position and its trend during motion. The distance the ZMP deviates from the support polygon boundary reflects the stability margin. The posture stability assessment also considers the effects of external disturbances, including uneven ground, wind interference, and collision impact, on the robot's balance. The fall risk level is comprehensively assessed based on the stability margin, center of gravity height, motion speed, and environmental conditions, and is categorized as low, medium, high, and extremely high. Paths with sufficient stability margins, a low center of gravity, and smooth motion are rated low risk, while paths with insufficient stability margins or rapid turns, acceleration, and deceleration are rated high risk. In a nursing home service robot application, when the robot needs to navigate the intersection of carpet and tile, posture stability analysis found that changes in floor material could cause foot slippage, affecting stability. Therefore, the fall risk level for this path segment was rated medium, and it was recommended to reduce the passage speed and increase stability monitoring.

[0086] The robot then detects and adjusts the safe distance between the robot and obstacles, assessing the collision risk level. Safe distance detection utilizes three-dimensional spatial analysis and dynamic obstacle avoidance algorithms to calculate the minimum distance and safety margin between the robot and various obstacles during path execution. Static obstacle detection analyzes the obstacle distribution in the constraint pre-map and calculates the shortest distance between the robot's envelope and the obstacle boundary, ensuring that the robot avoids geometric interference during motion. Dynamic obstacle detection incorporates dynamic environment prediction sequences to analyze the relative relationship between the robot and moving obstacles in time and space, predicting potential collision risk points and occurrence times. Safe distance thresholds are determined based on robot size, movement speed, and reaction time, including both minimum safe distance and recommended safe distance criteria. Collision risk assessment considers obstacle type, material, and relative motion. Hard obstacles have a higher collision risk than soft obstacles, and fast-moving obstacles have a higher collision risk than slow-moving obstacles. Collision risk levels are classified into four levels: safe, caution, warning, and dangerous, based on minimum safe distance, collision probability, and potential damage. Paths with sufficient safety distance and no collision risk are rated safe, while paths with safety distances close to the threshold or with a slight potential for collision are rated caution. In a factory inspection robot application, when the robot needs to pass through narrow passages between production line equipment, a collision risk assessment finds that the safety distance to the conveyor equipment is only within the minimum threshold, and the conveyor movement may increase the risk of collision. Therefore, the collision risk level for this path is rated warning, and a detour or reduced speed is recommended.

[0087] Next, the joint motion energy and time costs of the adjusted path are estimated, and the energy risk level is assessed. Energy estimation calculates the total energy consumption during path execution by analyzing the motion parameters and load conditions of each robot joint. Joint motion analysis includes the angle change, angular velocity, angular acceleration, and load torque of each joint. Kinematic and dynamic calculations are used to determine the power requirements and energy consumption distribution of the joint motors. Time cost estimation considers the robot's motion speed limits, acceleration and deceleration constraints, and dwell time to calculate the total time required to complete the path. Energy risk assessment establishes energy consumption thresholds and time limits. Paths that exceed battery capacity limits or task time requirements are rated as high energy risk. Energy optimization analysis compares the energy efficiency of different paths, prioritizing efficient paths with low energy consumption and short time. Energy risk levels are comprehensively assessed based on total energy consumption, energy consumption per unit distance, critical joint loads, and task completion time, and are classified into four levels: economic, standard, demanding, and overloaded. Paths with moderate energy consumption and reasonable time are rated as economic, while paths with high energy consumption or long time are rated as demanding. During the hotel service robot's room cleaning task, energy consumption assessments found that frequent climbing and descending stairs significantly increases leg joint load and battery consumption, potentially affecting a full day's work endurance. Therefore, the energy consumption risk level of this path is assessed as strenuous, and it is recommended to prioritize the use of elevators or reduce the number of trips between floors.

[0088] Finally, a comprehensive safety score is calculated based on the fall risk level, collision risk level, and energy consumption risk level, completing a multi-dimensional safety risk assessment. The comprehensive score utilizes a hierarchical weighted calculation method, converting the risk levels of the three dimensions into standardized numerical scores for unified processing. The fall risk level is quantified using a stability coefficient: low risk corresponds to a stability coefficient of 1.0, medium risk corresponds to 0.7, high risk corresponds to 0.4, and extremely high risk corresponds to 0.1. This reflects the robot's balance safety during motion. The collision risk level is quantified using a safety distance coefficient: the safety level corresponds to a distance coefficient of 1.0, the caution level corresponds to 0.8, the warning level corresponds to 0.5, and the danger level corresponds to 0.2. This reflects the safety margin for the robot's interaction with the environment. The energy consumption risk level is quantified using an efficiency coefficient: the economic level corresponds to an efficiency coefficient of 1.0, the standard level corresponds to 0.8, the effort level corresponds to 0.6, and the overload level corresponds to 0.3. This measures the rationality of resource consumption during path execution. The weight distribution is dynamically adjusted according to the application scenario and safety requirements. The weight of fall risk is set to 0.5 in medical environments, the weight of collision risk is set to 0.5 in crowded areas, and the weight of energy consumption risk is appropriately increased in long-distance tasks. The comprehensive safety score is calculated by weighted summation to obtain the final score of each adjusted path. The score range is 0-1.0. The higher the score, the safer and more reliable the path. Multi-dimensional risk association analysis identifies the mutual influence between different risk types. When the risk of falling is high, the impact weight of collision risk is increased. When the energy consumption risk is overloaded, the acceptability of the overall path is reduced. A risk threshold control mechanism is established during the assessment process. Paths with a comprehensive score below 0.6 are marked as high-risk paths, and paths below 0.3 are judged as unacceptable paths. This ensures the safety bottom line of path selection and completes a multi-dimensional safety risk assessment.

[0089] Based on a multi-dimensional safety risk assessment, the optimal safe path is determined. Autonomous navigation path planning for the humanoid robot is accomplished using an intelligent decision-making mechanism driven by assessment results. The comprehensive safety score and risk analysis results are used to screen the optimal path and generate a solution. Path selection begins with a safety threshold filter. Based on the comprehensive score from the multi-dimensional safety risk assessment, candidate paths with scores below the safety threshold are eliminated to ensure that all paths entering the decision-making phase meet basic safety requirements. Qualified paths are ranked and analyzed based on their comprehensive scores. The path with the highest score is selected as the preferred solution, and the path with the second-highest score as the backup solution, forming a hierarchical path selection system. The optimal path determination also considers task-specific requirements and environmental constraints, balancing performance indicators such as path length, execution time, and energy efficiency while ensuring safety. Path optimization processes the selected optimal path through detailed adjustments and parameter calibration, including determining key parameters such as speed distribution, turning radius, stop point settings, and safety monitoring point placement. The final path solution includes a complete trajectory coordinate sequence, time-velocity distribution curve, joint angle change sequence, and a set of safety control instructions, forming a standardized control scheme that can be directly used for robot navigation execution. In the application of nursing robots in nursing homes, the optimal option is selected from four candidate paths based on multi-dimensional safety risk assessment. Among them, the path passing through the medical equipment area is excluded due to the high collision risk, the path passing near the stairs is downgraded to an alternative due to the high risk of falling, and the path passing through the rehabilitation training area is affected by its high energy consumption. Finally, the path passing through the main corridor and avoiding high-risk areas is selected as the optimal safe path. The comprehensive score of this path reaches 0.89, which can ensure that the nursing robot can complete the elderly care service tasks safely and efficiently and complete the autonomous navigation path planning of the humanoid robot.

[0090] In order to implement the humanoid robot autonomous navigation path planning method in a complex environment corresponding to the above method embodiment, to achieve the corresponding functions and technical effects. Figure 2 , Figure 2 The following is a block diagram of a humanoid robot autonomous navigation path planning device 200 in a complex environment provided by an embodiment of the present application. For ease of explanation, only the parts related to this embodiment are shown. The humanoid robot autonomous navigation path planning device 200 in a complex environment provided by an embodiment of the present application includes:

[0091] The demand receiving module 201 is used to obtain the current position of the robot as the starting position, receive the path demand request of the humanoid robot navigation task, and determine the target position according to the path demand request;

[0092] A constraint construction module 202 is used to construct a feasible motion space of the robot based on the body size constraints, joint motion constraints and dynamic balance constraints of the humanoid robot, and generate a constraint pre-map;

[0093] The prediction and analysis module 203 is used to obtain environmental perception data, perform change trend analysis on the environmental perception data, predict environmental change trajectories in future periods, and generate a dynamic environmental prediction sequence;

[0094] a bidirectional planning module 204 for establishing a bidirectional planning mechanism based on the constrained pre-map and the dynamic environment prediction sequence, performing a forward path search from the starting position to the target position, and simultaneously performing a reverse path search from the target position to the starting position;

[0095] A path fusion module 205 is configured to perform path fusion based on intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses;

[0096] a social adjustment module 206 configured to analyze human behavior patterns in the environment based on the environmental perception data to identify potential conflict areas, establish a social negotiation strategy based on the environmental perception data and the potential conflict areas, use an intermediate meeting point between the forward path search and the reverse path search as a social key node, and perform social adaptive adjustment on the multiple candidate path hypotheses based on the social negotiation strategy and the social key node to obtain an adjusted path;

[0097] The path optimization output module 207 is used to perform a multi-dimensional safety risk assessment on the adjusted path, determine the optimal safety path based on the multi-dimensional safety risk assessment, and complete the autonomous navigation path planning of the humanoid robot.

[0098] The above-mentioned humanoid robot autonomous navigation path planning device 200 in a complex environment can implement the humanoid robot autonomous navigation path planning method in a complex environment of the above-mentioned method embodiment. The optional options in the above-mentioned method embodiment also apply to this embodiment and are not described in detail here. The remaining contents of the embodiments of this application can refer to the contents of the above-mentioned method embodiment and are not repeated in this embodiment.

[0099] like Figure 3 As shown, the third embodiment of the present invention also provides a computer device, including a memory 301, a processor 302, and a computer program stored in the memory 301 and executable on the processor 302, characterized in that when the processor 302 executes the program, the steps of the method for autonomous navigation path planning of a humanoid robot in a complex environment described in the first embodiment of the present invention are implemented.

[0100] The purpose of the above embodiments is to exemplify and deduce the technical solution of the present invention, and to fully describe the technical solution, purpose and effect of the present invention. Its purpose is to enable the public to have a more thorough and comprehensive understanding of the disclosed content of the present invention, and it does not limit the scope of protection of the present invention.

[0101] The above embodiments are not exhaustive and may include many other embodiments not listed above. Any replacements and improvements made without violating the concept of the present invention are within the scope of protection of the present invention.

Claims

1. A method for autonomous navigation path planning of a humanoid robot in a complex environment, characterized by: include: Obtaining the current position of the robot as a starting position, receiving a path requirement request for a humanoid robot navigation task, and determining a target position according to the path requirement request; The method comprises: constructing a feasible motion space of the humanoid robot based on the body size constraints, joint motion constraints, and dynamic balance constraints of the humanoid robot, and generating a constraint pre-map, including: analyzing the body width, height, and center of gravity position parameters of the humanoid robot to determine the body size boundary; determining the joint motion boundary based on the joint angle range and motion speed limit of the humanoid robot; analyzing the posture constraint conditions for stable walking according to the dynamic balance algorithm of the humanoid robot; and generating a constraint pre-map based on the body size boundary, the joint motion boundary, and the posture constraint conditions. Acquire environmental perception data, perform change trend analysis on the environmental perception data, predict environmental change trajectories in future periods, and generate a dynamic environmental prediction sequence; A bidirectional planning mechanism is established based on the constrained pre-map and the dynamic environment prediction sequence, including: setting a starting search node and a target search node in the constrained pre-map; performing a forward path extension search based on the dynamic environment prediction sequence starting from the starting search node; performing a reverse path extension search based on the dynamic environment prediction sequence starting from the target search node; monitoring node overlap between the forward path extension search and the reverse path extension search, establishing a bidirectional planning mechanism, performing a forward path search from the starting position to the target position, and simultaneously performing a reverse path search from the target position to the starting position; performing path fusion according to intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses; Based on the environmental perception data, human behavior patterns in the environment are analyzed to identify potential traffic conflict areas, and a social negotiation strategy is established for the environmental perception data and the potential traffic conflict areas, including: analyzing the traffic priority and space requirements of humans based on the environmental perception data and the potential traffic conflict areas; determining the avoidance timing and avoidance method of the humanoid robot according to the traffic priority; generating an avoidance strategy, a waiting strategy and a collaborative traffic strategy based on the avoidance timing, the avoidance method and the space requirement; combining the avoidance strategy, the waiting strategy and the collaborative traffic strategy to form a social negotiation strategy, and using the intermediate meeting point of the forward path search and the reverse path search as the social negotiation strategy. A social key node, based on the social negotiation strategy and the social key node, performs social adaptive adjustment on the multiple candidate path hypotheses to obtain an adjusted path, wherein the analyzing the human's passage priority and space requirement based on the environmental perception data and the potential passage conflict area includes: identifying the human's travel direction and travel speed in the passage conflict area based on the environmental perception data; judging the human's passage urgency based on the travel direction and the travel speed; analyzing the human's carrying items and physical condition based on the environmental perception data to determine the human's space occupation requirement; and determining the human's passage priority and space requirement based on the passage urgency and the space occupation requirement; A multi-dimensional safety risk assessment is performed on the adjusted path, and an optimal safety path is determined based on the multi-dimensional safety risk assessment to complete the autonomous navigation path planning of the humanoid robot.

2. The method for autonomous navigation path planning of a humanoid robot in a complex environment according to claim 1, characterized in that: The step of analyzing the change trend of the environmental perception data, predicting the environmental change trajectory in the future period, and generating a dynamic environmental prediction sequence includes: extracting position information and motion information of dynamic obstacles from the environmental perception data; Performing time series analysis on the position information and the motion information to extract motion pattern features; predicting the future motion trajectory of the dynamic obstacle based on the motion pattern characteristics; The future motion trajectories are arranged in a time series to generate a dynamic environment prediction sequence.

3. The method for autonomous navigation path planning of a humanoid robot in a complex environment according to claim 1, characterized in that: The performing path fusion according to the intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses includes: detecting an optimal meeting point of the forward path search and the reverse path search; Connecting the forward path search and the reverse path search based on the optimal meeting point to form a primary candidate path; analyzing a suboptimal meeting point of the forward path search and the reverse path search; An alternative candidate path is generated based on the suboptimal meeting point, and is combined with the primary candidate path to form multiple candidate path hypotheses.

4. The method for autonomous navigation path planning of a humanoid robot in a complex environment according to claim 1, characterized in that: The multi-dimensional security risk assessment of the adjustment path includes: analyzing the posture stability of the humanoid robot in the adjustment path and assessing the fall risk level; detecting a safe distance between the adjusted path and an obstacle, and assessing a collision risk level; estimating the joint movement energy consumption and time cost of the adjustment path, and evaluating the energy consumption risk level; A comprehensive safety score is performed based on the fall risk level, the collision risk level, and the energy consumption risk level to complete a multi-dimensional safety risk assessment.

5. A humanoid robot autonomous navigation path planning device in a complex environment, characterized by: include: A demand receiving module is used to obtain the current position of the robot as the starting position, receive the path demand request of the humanoid robot navigation task, and determine the target position according to the path demand request; A constraint construction module is used to construct a feasible motion space of the robot based on the body size constraints, joint motion constraints, and dynamic balance constraints of the humanoid robot, and generate a constraint pre-map, including: analyzing the body width, height, and center of gravity position parameters of the humanoid robot to determine the body size boundary; determining the joint motion boundary based on the joint angle range and motion speed limit of the humanoid robot; analyzing the posture constraint conditions for stable walking according to the humanoid robot dynamic balance algorithm; and generating a constraint pre-map based on the body size boundary, the joint motion boundary, and the posture constraint conditions. A prediction and analysis module is used to obtain environmental perception data, perform change trend analysis on the environmental perception data, predict environmental change trajectories in future periods, and generate a dynamic environmental prediction sequence; A bidirectional planning module is configured to establish a bidirectional planning mechanism based on the constrained pre-map and the dynamic environment prediction sequence, comprising: setting a starting search node and a target search node in the constrained pre-map; performing a forward path extension search based on the dynamic environment prediction sequence starting from the starting search node; performing a reverse path extension search based on the dynamic environment prediction sequence starting from the target search node; monitoring node overlap between the forward path extension search and the reverse path extension search, establishing a bidirectional planning mechanism, performing a forward path search from the starting position to the target position, and simultaneously performing a reverse path search from the target position to the starting position; a path fusion module, configured to perform path fusion according to intermediate meeting points of the forward path search and the reverse path search to generate multiple candidate path hypotheses; A social adjustment module is used to analyze the human behavior pattern in the environment based on the environmental perception data to identify potential passage conflict areas, and establish a social negotiation strategy for the environmental perception data and the potential passage conflict areas, including: analyzing the passage priority and space requirements of humans based on the environmental perception data and the potential passage conflict areas; determining the avoidance time and avoidance method of the humanoid robot according to the passage priority; generating an avoidance strategy, a waiting strategy and a collaborative passage strategy based on the avoidance time, the avoidance method and the space requirements; combining the avoidance strategy, the waiting strategy and the collaborative passage strategy to form a social negotiation strategy, and combining the intermediate between the forward path search and the reverse path search The meeting point is used as a social key node, and social adaptive adjustment is performed on the multiple candidate path hypotheses based on the social negotiation strategy and the social key node to obtain an adjusted path, wherein the analyzing the human's passage priority and space requirement based on the environmental perception data and the potential passage conflict area includes: identifying the human's travel direction and travel speed in the passage conflict area based on the environmental perception data; judging the human's passage urgency based on the travel direction and the travel speed; analyzing the human's carrying items and physical condition based on the environmental perception data to determine the human's space occupation requirement; and determining the human's passage priority and space requirement based on the passage urgency and the space occupation requirement; The path optimization output module is used to perform a multi-dimensional safety risk assessment on the adjusted path, determine the optimal safety path based on the multi-dimensional safety risk assessment, and complete the autonomous navigation path planning of the humanoid robot.

6. A computer device, characterized in that: The method comprises a memory and a processor; the memory is used to store a computer program; the processor is used to execute the computer program and implement the method according to any one of claims 1 to 4 when executing the computer program.

Images