A neural field guided humanoid robot adaptive door opening control method

Abstract

The present application relates to the technical field of robots, in particular to a neuro field guided humanoid robot adaptive door opening control method, comprising: acquiring point cloud information of a door opening area, constructing a neuro field and detecting a door handle pose, wherein the neuro field is formed by boundary information and normal information of a door body structure; generating a collision-free approach path using the boundary information, the normal information and the door handle pose, and using quadratic programming to control the robot to approach and hold the door handle; using the normal information, applying an initial pulling force through a preset admittance control strategy, and combining modeling optimization to estimate the door shaft position; based on the door shaft position, constructing a structure consistency circular arc path in combination with the boundary information and the normal information; combining the boundary information, and using weighted quadratic programming to control the robot to perform actions according to the collision-free approach path and the structure consistency circular arc path. The present application embeds the neuro field in key control links, and realizes task level dynamic optimization under perception driving.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robots, in particular to a neural field guided humanoid robot adaptive door opening control method. BACKGROUND

[0002] With the rapid development of robot technology in various fields, humanoid robots have superior environmental adaptability and task flexibility due to their structural bionics, and have become a research focus. Among them, door opening operation as a basic and challenging task has attracted widespread attention. When a humanoid robot performs a door opening task, it mainly faces the following two main challenges: on the one hand, the different reaction forces of the movable and immovable directions of the door limit the degrees of freedom of the robot's end motion, and the humanoid robot needs to adjust flexibly during the process of pushing and pulling the door to adapt to the changes in the reaction force of the door; on the other hand, the reaction force of the door fluctuates dynamically with the change of the motion state during the door opening process, further increasing the uncertainty in the operation process. Therefore, the humanoid robot needs to identify the movable trajectory of the door online during the door opening process, while considering maintaining appropriate contact force between the robot and the door, so as to complete the set task.

[0003] Related technical personnel have made some research on the control method of robot door opening operation. Existing technology one determines the door panel data and door handle data around the robot through point cloud data, so as to calculate the door opening motion trajectory and door opening grasping mode. This method measures and calculates the motion size of the door according to visual information in advance. When there is a deviation between the measurement result and the true value, the robot will generate a larger deviation force at the work gripping point, affecting the work, which may lead to the termination or failure of the work task. Existing technology two identifies the target object through vision, calculates its pose information, reads the relative position of the door handle and the door opening mode stored therein, and then generates the travel route of the mechanical arm pressing the door handle, executes grasping and opens the door in the preset mode. This method completes the door opening operation by reading the stored information. The robot can only complete the known door opening task and cannot adapt to new environments.

[0004] In addition, the current humanoid robot adaptive door opening control method generally has the problem of independent operation trajectory planning and movement trajectory planning. Due to the lack of coordination and optimization of the upper body and the lower body, in the actual door opening process, it often leads to incoherent whole body movement, insufficient operation precision, and difficulty in fully exerting the overall dynamic control advantage of the humanoid robot.

[0005] Therefore, in view of the problems of inaccurate perception, insufficient environmental adaptability and poor whole body coordination of the humanoid robot in the door opening task in the prior art, it is urgent to propose a humanoid robot adaptive door opening control method that can realize real-time perception-flexible response-whole body coordination, so as to improve the ability of the humanoid robot to autonomously complete the door opening task in a complex environment. SUMMARY

[0006] The object of the present application is to provide a neural field guided humanoid robot adaptive door opening control method, which models the door body in a continuous and differentiable structure by constructing a neural signed distance function, and directly embeds the neural field into key control links such as path generation, control target function, posture guidance, and weight adjustment, to realize task-level dynamic optimization under perception driving.

[0007] To achieve the above object, the present application provides the following scheme:

[0008] A neural field guided humanoid robot adaptive door opening control method, comprising:

[0009] Obtaining point cloud information of a door opening area;

[0010] Constructing a neural field based on the point cloud information, wherein the boundary information and normal information of the door body structure output by the neural signed distance function form the neural field;

[0011] Detecting the door handle pose based on the point cloud information using a pre-trained PoseCNN algorithm;

[0012] Generating a collision-free approach path using the boundary information, normal information, and door handle pose, and using quadratic programming to control the robot to approach and grasp the door handle;

[0013] Using the normal information, applying an initial pulling force through a preset admittance control strategy, and combining modeling optimization to estimate the door shaft position;

[0014] Based on the door shaft position, constructing a structure-consistent circular arc path in combination with the boundary information and normal information;

[0015] In combination with the boundary information, using weighted quadratic programming to control the robot to perform actions according to the collision-free approach path and structure-consistent circular arc path, to complete the adaptive door opening control of the humanoid robot.

[0016] Optionally, the neural signed distance function is constructed based on a fully connected multilayer perceptron, and is:

[0017] ;

[0018] Wherein, the input is an arbitrary three-dimensional space coordinate point x; the output is the signed shortest distance d = Φ(x) from the point to the door body surface, and Φ(x) is the boundary function calculated by the neural signed distance function for the point x; is a neural signed distance function, θ is a network parameter, R represents a one-dimensional space, and R 3 represents a three-dimensional space.

[0019] Optionally, outputting the boundary information and normal information of the door structure comprises:

[0020] The neural signed distance function calculates any three-dimensional space coordinate point x to obtain the boundary function Φ(x), if Φ(x) = 0, it indicates that the point is located on the door surface; if Φ(x) < 0, it indicates that the point is located inside the door; if Φ(x) > 0, it indicates that the point is located outside the door; and the door surface normal direction at the point x, i.e. the structure gradient .

[0021] Optionally, the boundary information, normal information and door handle pose are used to generate a collision-free approach path by a path planning objective function under the constraint of structure consistency, wherein the constraint of structure consistency is:

[0022] Avoiding crossing the area of Φ(x) < 0; keeping the end pose and relatively constant or slowly changing in the feasible space;

[0023] The path planning objective function is:

[0024] ;

[0025] Wherein, is the path planning loss, is the boundary function calculated by the neural signed distance function for the point , is the Cartesian space position of the robot end at time step t; is the expected motion direction of the current position in path planning; δ is the minimum safety distance of the structure boundary; λ and α are proportional factors, is the structure gradient at the point .

[0026] Optionally, the robot is controlled to approach and hold the door handle by using the quadratic programming comprises:

[0027] Converting the robot control problem into a quadratic programming optimization problem, and dynamically adjusting the control strategy according to the real-time contact state between the robot end and the door handle in the process that the robot holds the door handle, wherein the optimization objective of the quadratic programming optimization problem is:

[0028] ;

[0029] Wherein, is the joint velocity vector, is the Jacobian matrix of the end effector, is the speed of the expected trajectory;

[0030] The conditional constraint of the quadratic programming optimization problem is:

[0031] ;

[0032] wherein q is a joint position variable, and represent the lower and upper limits of the joint position, respectively, and represent the lower and upper limits of the joint velocity, respectively.

[0033] Optionally, the admittance control strategy is implemented through a position control inner loop and an admittance control outer loop, comprising:

[0034] When receiving an input displacement instruction, an expected displacement is formed by comparing the input displacement instruction with an actual feedback displacement; in the position control inner loop, an expected position and pose of the robot end are converted into motion parameters of each joint, a control signal is generated based on the motion parameters to drive the robot motion, and the actual position of the robot end is calculated in real time according to the actual motion of each joint of the robot, compared with the expected displacement, and closed-loop control of the position is realized;

[0035] In the admittance control outer loop, when the robot is subjected to a load force during motion, a preset admittance control model adjusts the motion of the robot according to the size and direction of the force; in the admittance control model , represents a virtual mass; represents a viscous damping coefficient; is a virtual stiffness; is a complex variable in Laplace transform.

[0036] Optionally, the optimization objective function for estimating the door shaft position in combination with modeling optimization is:

[0037] ;

[0038] ;

[0039] ;

[0040] wherein, is a trajectory circular arc consistency error, is a structural normal consistency error, is a structural consistency weight adjustment factor, r is a fitting radius, and c is a door shaft center, is a Cartesian space position of the robot end at time step t, is a point structure gradient.

[0041] Optionally, constructing the structural consistency circular arc path comprises:

[0042] constructing a circular arc path based on the door shaft position;

[0043] introducing a boundary function Φ(x) to perform validity verification on the circular arc path, and adopting a structure gradient guiding the direction of the circular arc path, and optimizing the discretized circular arc path to obtain the structure-consistent circular arc path.

[0044] Optionally, the objective function of the weighted quadratic programming control is:

[0045] ;

[0046] wherein, and is a structure-aware term weight hyperparameter; is a boundary consistency loss; is a gradient guidance loss; , are Jacobian matrices of the upper body and the lower body of the robot respectively; , are expected speeds of the upper body and the lower body respectively; , are control weights of the upper body and the lower body respectively, is a joint speed vector.

[0047] Optionally, the upper body and the lower body adopt a dynamic adjustment mechanism when the action is performed, and the dynamic adjustment mechanism dynamically adjusts the upper body and the lower body according to the consistency of the force direction of the end of the robot and the normal direction of the door structure, so as to achieve the balance and stability of the robot.

[0048] The present application has the following advantages:

[0049] (1) The present application improves the perception and understanding ability of the door structure by using neural field modeling: the present application introduces N-SDF to construct a continuous geometric representation model of the door body, which can complete the modeling and identification of key structures such as door surface, door shaft and door handle without relying on CAD model or explicit template. Compared with the traditional point cloud + geometric fitting method, this modeling method has higher expression accuracy and differentiability, so that the robot can still achieve high robustness in perception and interaction when facing different sizes, opening and closing modes or local occlusion of the door structure.

[0050] (2) The application can realize structure-guided path planning and task execution: the neural field model established by the application is not only used for door body recognition, but also guides the trajectory generation of the robot end and the optimization of the interaction direction through the output structure boundary function and gradient information. In the path planning process, the robot can automatically avoid obstacles, calibrate the trajectory direction according to the door structure, and generate a circular arc path conforming to the opening and closing law of the door, so as to realize adaptive operation on complex doors (such as inward opening doors, outward opening doors, rotating doors, etc.).

[0051] (3) The application introduces a structure perception mechanism in control optimization to improve operation stability and coordination: based on the WQP control framework, the application introduces a neural field boundary consistency term and a gradient direction guide term in the objective function, so that the controller fully considers the door structure constraint when executing actions. In addition, by dynamically adjusting the upper and lower body control weights through the structure-force coupling index, the dynamic trade-off between operation accuracy and body stability is realized, which significantly enhances the autonomous control performance and fault tolerance of the humanoid robot in complex operation processes. BRIEF DESCRIPTION OF DRAWINGS

[0052] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed in the embodiments. Obviously, the drawings described below are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained from these drawings without creative labor.

[0053] Figure 1 A neural field guided humanoid robot adaptive door opening control method flow chart of an embodiment of the application;

[0054] Figure 2 A control block diagram of a neural field guided humanoid robot adaptive door opening control method of an embodiment of the application;

[0055] Figure 3 An N-SDF model construction flow chart of an embodiment of the application;

[0056] Figure 4 A preset admittance control strategy block diagram of the robot preliminary door opening of an embodiment of the application;

[0057] Figure 5 A planning path determination process schematic diagram of an embodiment of the application;

[0058] Figure 6 A neural field guiding mechanism action schematic diagram of an embodiment of the application. DETAILED DESCRIPTION

[0059] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative work fall within the protection scope of the present application.

[0060] In order to make the above objectives, characteristics and advantages of the present application more apparent, the present application will be described in further detail below with reference to the accompanying drawings and specific embodiments.

[0061] The embodiment provides a neural field guided humanoid robot adaptive door opening control method, as shown in the figure, which comprises the following steps: Figure 1

[0062] Obtain point cloud information of a door opening area;

[0063] Construct a neural field based on the point cloud information, wherein the boundary information and normal information of a door body structure output by a neural symbolic distance function form the neural field;

[0064] Detect a door handle pose by using a pre-trained PoseCNN algorithm based on the point cloud information;

[0065] Generate a collision-free approach path by using the boundary information, normal information and door handle pose, and control the robot to approach and hold the door handle by using quadratic programming;

[0066] Apply an initial pulling force by using a preset admittance control strategy by using the normal information, and estimate a door shaft position in combination with modeling optimization;

[0067] Construct a structure-consistent circular arc path based on the door shaft position in combination with the boundary information and normal information;

[0068] Control the robot to perform actions according to the collision-free approach path and structure-consistent circular arc path by using weighted quadratic programming in combination with the boundary information, so as to complete adaptive door opening control of the humanoid robot.

[0069] Specifically, as shown in the figure, Figure 1 , Figure 2 The neural field guided humanoid robot adaptive door opening control method provided by the embodiment comprises the following steps:

[0070] ​Step 101, perception and structure modeling: using a robot-mounted RGB-D camera to obtain point cloud information of the door opening area, constructing a neural signed distance function (N-SDF) based on deep learning, i.e., N-SDF model, to realize continuous implicit modeling of key structures such as door handle and door plane. The model outputs the door body boundary function Φ(x) and its structure gradient , as the basis for subsequent path planning and control.

[0071] As shown in Figure 3 , the construction process of the N-SDF model mainly includes the following steps:

[0072] (1) Point cloud collection and preprocessing: using a robot head or upper body mounted RGB-D camera to collect dense point cloud data of the door opening area , where is the i-th original point coordinate, N is the number of original point clouds, including door surface, door handle, door frame and other structures, and filtering and noise reduction operations are performed on the original point cloud.

[0073] (2) Spatial sampling to construct neural field input points: constructing a 3D sampling boundary box around the point cloud coverage area, and adaptively sampling a large number of spatial point cloud data as input query points for the N-SDF network, where is the j-th spatial point coordinate, and M is the number of spatial point clouds.

[0074] (3) Constructing N-SDF model: using a fully connected multilayer perceptron to define the neural signed distance function:

[0075] ;

[0076] where the input is an arbitrary three-dimensional coordinate point x; the output is the signed shortest distance d=Φ(x) from the point to the door body surface, Φ(x) is the boundary function calculated by the neural signed distance function for the coordinate point x; is the neural signed distance function, θ is the network parameter, R represents one-dimensional space, and R 3 represents three-dimensional space.

[0077] (4) N-SDF model training: for each spatial point , the nearest neighbor from the point cloud can be constructed to represent the supervised target point relative to the door body "inside / outside", and "how far away":

[0078] ;

[0079] where, For the point Estimate the normal.

[0080] The training loss function is:

[0081] ;

[0082] (5) Model output: Obtain the structure boundary and normal information: After the model training or deployment is completed, The SDF value of any point in space: If Φ(x)=0, it means that the point is located on the door surface; if Φ(x)<0, it means that the point is inside the door body (illegal area); if Φ(x)>0, it means that the point is outside the door body (passable area). Structure gradient Indicates the door surface normal direction at point x.

[0083] (6) Output interface integration to downstream links: Function Φ(x) and structure gradient As an interface output: The path planner uses Φ(x) to determine whether the trajectory enters the illegal area; the WQP controller uses Φ(x) as a boundary constraint term to guide the structure consistency; the admittance controller uses Adjust the end-effector force response direction to maintain natural interaction; the weight adjuster uses Calculate the force-structure consistency index.

[0084] (7) Door handle pose detection: Use the PoseCNN algorithm to detect the 6D space pose of the door handle. PoseCNN trains a deep learning model to extract key features from point cloud data and predicts the pose of the object based on these features.

[0085] Step 102, approach and grasp: Based on the door handle pose and door surface normal information output by N-SDF, generate a collision-free approach path. Use the quadratic programming (QP) method to convert robot control into an optimization problem, realize the mapping and tracking of Cartesian space path to robot joint space, control the robot end-effector to approach and grasp the door handle. Specifically includes:

[0086] (1) Path planning: Based on the pose of the door handle in step 101 and the N-SDF output door body structure function Φ(x) and door surface normal, i.e. structure gradient , generate a collision-free trajectory under the constraint of structure consistency;

[0087] The structure consistency constraint condition is: the path should approach along the outer normal direction of the door surface; avoid crossing the area where Φ(x)<0 (i.e. inside the door body); keep the end-effector pose and relatively unchanged or slowly changing in the feasible space.

[0088] Path planning objective function (the first term keeps the end outside the door body, the second term guides the consistency of the direction, and the minimum objective function obtains the planning path):

[0089] ;

[0090] wherein, is the path planning loss, is the boundary function calculated by the neural symbolic distance function for point , is the Cartesian space position of the robot end at time step t; is the expected motion direction of the current position in path planning; δ is the minimum safety distance of the structure boundary; λ and α are proportional factors, is the structure gradient at point .

[0091] (2) QP control: assuming that the expected trajectory of the robot end effector is , and the actual trajectory is , the trajectory error is defined as:

[0092] ;

[0093] In order to make the end effector track the expected trajectory, the two-norm of the trajectory error needs to be minimized, and the robot control problem is converted into a QP optimization problem, and the optimization objective is:

[0094] ;

[0095] wherein, is the joint velocity vector, is the Jacobian matrix of the end effector, is the velocity of the expected trajectory (obtained according to the planning trajectory).

[0096] The conditional constraint is:

[0097] ;

[0098] wherein, q is the joint position variable, and represent the lower limit and upper limit of the joint position respectively, and represent the lower limit and upper limit of the joint velocity respectively.

[0099] In the process of grasping the door handle, the control strategy is dynamically adjusted according to the real-time contact state of the robot end effector and the door handle, so as to prevent unstable grasping or damage to the door handle.

[0100] Step 103, preliminary door opening and door shaft estimation: based on the preset admittance control strategy, an initial pulling force is applied in the normal direction of the door plane, and the motion trajectory is adjusted through the feedback of the end six-dimensional force sensor to ensure smooth and effective force control process. At the same time, combined with the robot end trajectory and the door body geometric information output by the neural field, the door shaft position estimation is carried out by adopting the structural consistency optimization method. Specifically, it includes:

[0101] (1) Preliminary door opening admittance control strategy: to ensure the dynamic stability and smooth response in the process of interacting with the door body, the end admittance control model is adopted in this embodiment to realize soft force application.

[0102] As shown in Figure 4 , the flow of the preset admittance control strategy for the robot to preliminarily pull the door is carried out around the position control inner loop and the admittance control outer loop. First, the robot receives the input displacement instruction (the initial pulling force direction in this embodiment is set to be consistent with the normal direction of the door plane output by the neural field), which forms the expected displacement after being compared with the actual feedback displacement. In the position control inner loop, the expected displacement is transmitted to the inverse kinematics module, which converts the expected position and attitude of the robot end into motion parameters of each joint. These parameters are sent to the position controller to generate specific control signals to drive the robot motion. At the same time, the forward kinematics calculates the actual position of the robot end effector in real time according to the actual motion of each joint of the robot, and feeds back to the system, which is compared with the expected displacement to realize the closed-loop control of the position, so as to ensure that the robot reaches the expected position as accurately as possible. In the admittance control outer loop, when the robot is subjected to load force in the motion process (the force information sensed by the robot end six-dimensional force sensor in this embodiment), the preset admittance control model will adjust the motion of the robot according to the size and direction of the force. This makes the robot not only accurate positioning, but also flexible to change the motion state and adjust the robot end motion trajectory in real time when interacting with the external environment.

[0103] In the admittance control model , represents the virtual mass (or equivalent inertia), which simulates the inertia characteristics of the robot in the motion process; represents the viscous damping coefficient, which embodies the damping effect of the system in the motion process due to friction and other factors; is the virtual stiffness (or equivalent stiffness), which describes the ability of the system to resist deformation; is a complex variable in Laplace transform.

[0104] (2) Neural field consistency based door axis estimation method: Traditional door axis estimation methods mostly use pure geometric fitting (such as RANSAC for hand trajectory circle fitting), but this method is easily disturbed by point noise and path deviation, and has poor robustness. To solve this problem, the embodiment introduces a structure consistency optimization strategy guided by neural field perception for door axis estimation.

[0105] Define the sequence of robot end grip point trajectories in the preliminary door opening process as , and the door axis center as an unknown variable , which is expected to satisfy the following two types of consistency constraints:

[0106] 1) Trajectory arc consistency: assuming that the end trajectory is a circular arc rotating around the door axis, minimize the consistency error of the end trajectory and concentric circles:

[0107] ;

[0108] Where r is the fitting radius.

[0109] 2) Structural normal consistency: assuming that the normal of the rotation plane should be perpendicular to the end motion direction, add the normal consistency constraint consistent with the neural field:

[0110] ;

[0111] Ensure that the direction of the connecting line between the door axis center and the trajectory point is orthogonal to the door surface direction.

[0112] In summary, the comprehensive optimization objective function is:

[0113] ;

[0114] Where is the structural consistency weight adjustment factor, and the comprehensive optimization objective function is solved by least squares method.

[0115] Step 104, path planning: based on the estimated door axis, combined with the door surface normal and boundary constraints in the neural field, a structure consistency circular arc path is constructed. The path planning process is controlled by the Φ(x) control point sampling and direction guidance, ensuring that the motion trajectory conforms to the door structure logic and avoids border crossing. Specifically, it includes:

[0116] (1) Path generation framework: as shown in Figure 5 , assuming that the door axis estimation result is , the path should be a circular arc trajectory with c as the center and r as the radius, the trajectory starting point is the current end grip point , the end point is determined by the desired door opening angle , and the plane of the circular arc trajectory is parallel to .

[0117] (2) Path screening under the constraint of neural field: In order to ensure that the planned path does not collide with the door body, the N-SDF boundary function Φ(x) is introduced to check the path validity:

[0118] ;

[0119] wherein, is the sequence of sampling points of the circular arc path; is a small positive value representing the minimum safety distance; if a certain path point , it will be considered that the point is not feasible, triggering the path correction or obstacle avoidance strategy.

[0120] (3) Structural gradient guided path direction: The motion direction between path points should be consistent with the normal direction of the door surface structure to avoid phenomena such as transverse cutting and attitude deviation. The neural field structural gradient is used to guide the path direction, and the following direction consistency loss is constructed for the path optimization objective:

[0121] ;

[0122] (4) Discrete representation and optimization of circular arc trajectory: The path trajectory is discretized into a series of trajectory points:

[0123] ;

[0124] wherein, is the rotation matrix of rotating an angle around the door axis, is the angle sequence uniformly divided from 0 to the desired opening angle , and T is the total time. The optimization objective is as follows:

[0125] ;

[0126] wherein, the first term limits the boundary legality, the second term maintains the path direction consistency, and the final output is the planned path .

[0127] Step 105, whole body coordinated control enhanced by structure perception: After the path generation is completed, the robot needs to coordinate the upper body (mainly the arm and torso) and the lower body (mainly the legs and support surface) to complete the task. Based on the Weighted Quadratic Programming (WQP) control framework, this embodiment introduces a neural field structure perception term to construct a coordinated controller, ensuring the dynamic optimal coordination of trajectory accuracy, attitude consistency, and support stability. Specifically, it includes:

[0128] (1) WQP control framework basis: the goal is to solve joint velocity command , so that the robot end (hand) follows the desired path as much as possible while maintaining body coordination and balance, in the form of:

[0129] ;

[0130] where, , are the Jacobian matrices of the upper and lower body of the robot, respectively; , are the desired velocities of the upper and lower body, respectively; , are the control weights of the upper and lower body, respectively, satisfying: , .

[0131] (2) Structure-aware WQP control: add a structure-aware term to the WQP control to build a coordinated controller that can make the robot respond to the door geometry in real time during path execution, ensuring the physical legality, direction consistency of end motion, and dynamic coordination of whole body control. The structure-aware term mainly includes:

[0132] 1) Boundary consistency loss term:

[0133] ;

[0134] where, ensures a safe buffer zone to avoid collisions between the path and the door. In essence, this term will form a soft constraint on the end position in the optimization, controlling the end path point to always be outside the door structure boundary defined by the neural field.

[0135] 2) Gradient guidance loss term:

[0136] ;

[0137] This term is used to control the consistency of the end motion direction and the normal direction of the door surface , and the parameter 𝛼 controls the motion speed scale.

[0138] 3) Weight dynamic adjustment mechanism: to achieve the trade-off between task flexibility and support stability, a mechanism is introduced to adjust the control weights of the upper and lower body according to the neural field and force feedback:

[0139] ;

[0140] where, represents the current ZMP (zero moment point) center of gravity offset; is the six-axis force sensor data at the end; Ensure the force direction is consistent with the door surface normal; and are the corresponding weights; represents the Sigmoid function for compressing the weight range.

[0141] Through the above weight dynamic adjustment mechanism, the consistency relationship between the door body structure direction and the end interaction force can be adjusted in real time, combined with the ZMP center of gravity offset, to realize the automatic and real-time adjustment of the upper and lower body control weight of the robot, so as to adapt to the different needs of stability and dexterity in different task stages. When the door is heavy or the path is long, , the lower limb support dominates, and the posture is stable; when the door is light or needs to be operated finely , the arm control dominates, and the dexterity is improved.

[0142] The final structure-aware WQP control objective function is:

[0143]

[0144] wherein, and are structure-aware item weight hyperparameters.

[0145] As shown in Figure 6 , the neural field guiding mechanism proposed in this embodiment is embodied in the following four core levels in the whole door opening control process:

[0146] (1) Guide path planning (structure-consistent trajectory generation);

[0147] The boundary function Φ(x) and normal gradient output by the N-SDF model are used to construct a circular arc path that fits the door surface structure, and to perform path screening and correction:

[0148] If the path sampling point satisfies , it is located inside the door body and is considered as an illegal path point. This path will be excluded;

[0149] If the plane of the path is not consistent with the door surface normal , the trajectory is adjusted by introducing a direction consistency loss term.

[0150] (2) Guide the interaction direction prediction in admittance control;

[0151] During the initial pulling of the door, the admittance controller adjusts the motion response according to the feedback of the six-axis force sensor at the end. This embodiment introduces the derived in the neural field as the door surface normal direction into the interaction modeling:

[0152] ​If the direction of the force applied at the end significantly deviates from , a direction compensation mechanism is triggered to effectively improve the alignment between the direction of the force applied at the end and the rotation trend of the door body, making the interaction more smooth and natural.

[0153] (3) Guidance control objective function construction (structure perception control constraint);

[0154] In the objective function of the WQP controller, this embodiment introduces two types of structure perception constraints based on neural fields:

[0155] Structure boundary soft constraint term (limit the end to cross the door body):

[0156] ;

[0157] Structure gradient guidance term (guide the trajectory direction to fit the door surface):

[0158] ;

[0159] The controller will automatically respond to the door structure field when performing path tracking and posture optimization, improving the physical consistency and environmental adaptability of trajectory execution.

[0160] (4) Dynamic adjustment mechanism for guiding upper and lower body weights;

[0161] This embodiment constructs a control weight adjustment function based on a structure-force coupling index, which is used to dynamically balance dexterity and stability:

[0162] ;

[0163] wherein, represents the current ZMP center of gravity offset; is the six-dimensional force sensor data at the end; ensures the consistency of the force direction with the normal direction of the door surface; and are the corresponding weights; represents a Sigmoid function used to compress the weight range.

[0164] used to measure the consistency of the force direction applied at the end with the normal direction of the door structure: if the consistency is low (the direction is not aligned), the lower body weight is automatically increased , enhancing the support force and the stability of the center of gravity; if the consistency is high, the lower body constraint is appropriately released, improving the flexibility and operation precision of the upper body. This mechanism allows the controller to automatically adjust the coordination strategy of the upper and lower bodies according to different door weights, path lengths, and support states, achieving optimal dynamic control of the whole body under the structure drive.

[0165] The above described embodiments are only to illustrate the preferred modes of the present application, and are not intended to limit the scope of the present application. Any modification and improvement made by those skilled in the art to the technical solutions of the present application without departing from the design spirit of the present application shall fall within the protection scope of the present application as defined by the claims.

Claims

1. A neural field-guided adaptive door opening control method for a humanoid robot, characterized in that, include: Obtain point cloud information of the door opening area; A neural field is constructed based on the point cloud information, wherein the neural field is formed by the boundary information and normal information of the gate structure output by the neural symbol distance function; Based on the point cloud information, the door handle pose is detected using the pre-trained PoseCNN algorithm; Using the boundary information, normal information, and door handle pose, a collision-free approach path is generated, and quadratic planning is used to control the robot to approach and grasp the door handle. Using the normal information, an initial tension is applied through a preset admittance control strategy, and the door hinge position is estimated by combining modeling optimization. Based on the gate hinge position, and combined with the boundary information and normal information, a structurally consistent circular arc path is constructed. Based on the boundary information, a weighted quadratic programming approach is used to control the robot to perform actions along the collision-free approach path and the structurally consistent circular arc path, thereby completing the adaptive door opening control of the humanoid robot.

2. The neural field-guided adaptive door opening control method for humanoid robots according to claim 1, characterized in that, The neural symbol distance function is constructed based on a fully connected multilayer perceptron and is as follows: ； The input is any three-dimensional spatial coordinate point x; the output is the signed shortest distance d=Φ(x) from the point to the surface of the gate, where Φ(x) is the boundary function calculated for point x using the neural symbolic distance function. Let θ be the neural symbol distance function, θ be the network parameters, and R represent a one-dimensional space. 3 Represents three-dimensional space.

3. The neural field-guided adaptive door opening control method for humanoid robots according to claim 2, characterized in that, The boundary and normal information of the gate structure are output, including: The neural symbol distance function is used to calculate the boundary function Φ(x) for any three-dimensional spatial coordinate point x. If Φ(x) = 0, it means the point is located on the surface of the door; if Φ(x) < 0, it means the point is located inside the door; if Φ(x) > 0, it means the point is located outside the door. The direction of the door surface normal at point x is also obtained, i.e., the structural gradient. .

4. The neural field-guided adaptive door opening control method for humanoid robots according to claim 3, characterized in that, Using the boundary information, normal information, and door handle pose, a collision-free approach path is generated through a path planning objective function under structural consistency constraints. The structural consistency constraints are: Avoid crossing regions where Φ(x) < 0; maintain end-effector attitude within the feasible space. Relatively unchanged or slowly changing; The objective function for the path planning is: ； in, For path planning loss, To use the neural symbolic distance function to pair points The boundary function is calculated. Let be the Cartesian space position of the robot's end effector at time step t; δ represents the desired direction of movement at the current position in path planning; δ is the minimum safe distance from the structural boundary; λ and α are scaling factors. For point Structural gradient.

5. The neural field-guided adaptive door opening control method for humanoid robots according to claim 1, characterized in that, The control of the robot to approach and grasp the door handle using the aforementioned quadratic programming includes: The robot control problem is transformed into a quadratic programming optimization problem. During the robot's grasping of the door handle, the control strategy is dynamically adjusted based on the real-time contact state between the robot's end effector and the door handle. The optimization objective of the quadratic programming optimization problem is: ； in, For joint velocity vectors, The Jacobian matrix of the end effector. The velocity of the desired trajectory; The constraints of the quadratic programming optimization problem are: ； Where q is the joint position variable. and These represent the lower and upper limits of the joint position, respectively. and These represent the lower and upper limits of joint velocity, respectively.

6. The neural field-guided adaptive door opening control method for humanoid robots according to claim 1, characterized in that, The admittance control strategy is implemented through an inner loop of position control and an outer loop of admittance control, including: When an input displacement command is received, the input displacement command is compared with the actual feedback displacement to form the desired displacement. In the position control inner loop, the desired position and posture of the robot end effector are converted into motion parameters of each joint. Based on the motion parameters, control signals are generated to drive the robot to move. At the same time, based on the actual movement of each joint of the robot, the actual position of the robot end effector is calculated in real time and compared with the desired displacement to achieve closed-loop position control. In the admittance control outer loop, when the robot is subjected to load force during motion, the preset admittance control model... The robot's motion is adjusted according to the magnitude and direction of the force; admittance control model. middle, Indicates virtual quality; Represents the viscous damping coefficient; It is virtual stiffness; It is a complex variable in the Laplace transform.

7. The neural field-guided adaptive door opening control method for humanoid robots according to claim 1, characterized in that, The objective function for optimizing the estimation of the gate hinge position by combining modeling is: ； ； ； in, For trajectory arc consistency error, For structural normal consistency error, Here, r is the structural consistency weight adjustment factor, r is the fitting radius, and c is the gate axis center. Let t be the Cartesian space position of the robot's end effector at time step t. For point Structural gradient.

8. The neural field-guided adaptive door opening control method for humanoid robots according to claim 1, characterized in that, Constructing the structurally consistent circular arc path includes: Construct an arc path based on the door hinge position; A boundary function Φ(x) is introduced to verify the validity of the circular arc path, and a structural gradient is used. Guide the direction of the arc path, and optimize the discretized arc path to obtain the structurally consistent arc path.

9. The neural field-guided adaptive door opening control method for humanoid robots according to claim 1, characterized in that, The objective function of the weighted quadratic programming control is: ； in, and The weight hyperparameters for the structure-aware term; This is the boundary consistency loss; Gradient-guided loss; , These are the Jacobian matrices for the upper and lower halves of the robot, respectively; , The expected speeds are for the upper body and lower body, respectively. , The control weights are for the upper body and lower body, respectively. This is the joint velocity vector.

10. The neural field-guided adaptive door opening control method for a humanoid robot according to claim 8, characterized in that, When performing the action, the upper and lower body adopt a dynamic adjustment mechanism. The dynamic adjustment mechanism dynamically adjusts the upper and lower body according to the consistency between the direction of the force applied by the robot end and the normal of the facade structure, so as to achieve robot balance and stability.

Images