A six-degree-of-freedom robot weak prior model grinding system and method

Abstract

This invention discloses a six-degree-of-freedom robot grinding system and method with a weak prior model. First, a comprehensive error model is established through robot kinematics and joint stiffness calibration. Based on workpiece point cloud data, a globally covering path is generated through region decomposition and path planning, and the desired tool pose is calculated by interpolation. During machining, laser displacement and six-dimensional force sensor data are simultaneously acquired, the actual local normal vector is fitted, and the actual distance is calculated. The true contact force is obtained through gravity compensation. The system uses a PI control law based on normal force deviation to correct the tool distance in real time, achieving adaptive tracking and constant force control on curved surfaces. Furthermore, based on the error model, a genetic algorithm is used to optimize the workpiece position and tool posture offline to reduce machining errors. Finally, commands are sent through the robot servo interface to integrate real-time perception, pose correction, and motion control. Experiments verify the system's control accuracy.

Description

Technical Field

[0001] This invention belongs to the field of robot control, specifically relating to a six-degree-of-freedom robot weak prior model grinding system and method. Background Technology

[0002] In the field of six-degree-of-freedom robotic grinding, existing technologies typically rely on accurate CAD models of the workpiece for offline programming and path planning. However, this method has limited applicability for complex free-form surfaces or workpieces with missing models. Furthermore, the absolute positioning accuracy of industrial robots is affected by kinematic errors, and under machining stress, their relatively low joint stiffness leads to further deformation. These factors collectively cause the actual machining trajectory to deviate from the theoretical path, severely impacting surface accuracy and consistency. While existing research has attempted to improve quality through offline error compensation, pose optimization, or online force control, a systematic solution is lacking that can simultaneously achieve globally covered path planning, systematic error suppression, and real-time surface tracking force control under weak prior model conditions. Therefore, developing a robotic intelligent grinding system and method that integrates offline planning and online perception control, while balancing geometric accuracy and contact force stability, has become a key requirement for improving the process capabilities in this field. Summary of the Invention

[0003] This invention relates to a six-degree-of-freedom robot grinding system and method with weak prior models, aiming to solve the problems of low machining accuracy and poor stability caused by uncertainties in robot kinematic errors, joint stiffness, and environmental interaction in existing technologies. This method achieves closed-loop management of the entire process from model construction, path planning, pose optimization to real-time force and position control through deep fusion of offline calibration and online sensing, significantly improving the robot's trajectory tracking accuracy and surface treatment quality.

[0004] To achieve the above objectives, the technical solution provided by this invention is as follows:

[0005] A six-degree-of-freedom robot weak prior model grinding system and method, characterized in that it includes:

[0006] Step S1: Conduct robot kinematics and stiffness calibration. First, design experiments to collect actual pose and force deformation data of the robot end effector under multiple configurations, and construct kinematic parameters and joint stiffness datasets. Use velocity-associated transformation to establish a kinematic error model, transform the pose error matrix into differential motion vectors through differential kinematics, and use the least squares method to fit and identify kinematic error parameters to eliminate no-load positioning errors. Then, use force-associated transformation to establish a joint stiffness model, measure the deformation of the end effector by applying a known external force, and identify the stiffness coefficients of each joint by combining the Jacobian matrix and the joint stiffness matrix. Finally, obtain a comprehensive robot body model containing kinematic errors and joint stiffness, and standardize the data.

[0007] Step S2: Obtain the workpiece surface boundary and perform global path planning. Obtain the boundary point set of the processing area through manual teaching or laser scanning and construct a point cloud dataset. Use Boustrophedon cell decomposition to divide the two-dimensional polygonal region into non-overlapping cells and generate parallel sweep paths. Use Dijkstra's algorithm to solve for the shortest connection path between cells to form a continuous two-dimensional path skeleton. Map the two-dimensional path points back to the three-dimensional surface. Use bilinear interpolation to obtain the three-dimensional position of the path points and the surface normal vector. Construct the desired tool posture matrix to ensure that the tool axis is antiparallel to the normal vector and the feed direction is aligned with the path tangent. Finally, standardize the global path points.

[0008] Step S3: Optimize the workpiece position and tool posture, calculate the comprehensive machining error of the global path points, which includes kinematic error and joint stiffness error, the latter taking into account the coupling effect of grinding force changing with actual grinding depth; perform error compensation by selecting the workpiece frame position, convert the absolute error into relative error, and construct an optimization function with the goal of minimizing the error; set operability indicators, joint limits and collision avoidance conditions as constraints, use a genetic algorithm to solve for the optimal workpiece position and optimal tool posture, and finally evaluate the optimization effect by the mean absolute error and root mean square error indicators;

[0009] Step S4: Real-time surface tracking and force control are achieved. During the robot's machining process, data from the laser displacement sensor and the six-dimensional force / torque sensor are synchronously acquired in servo cycles. The plane normal vector and actual distance are calculated by fitting the local plane based on the three-point laser measurement values, and the desired pose is generated by combining it with the desired posture from step S2. Gravity compensation is performed on the six-dimensional force data to obtain the actual contact force. After projecting it onto the normal, a parallel force control strategy is adopted. The distance correction is calculated through a PI control law and superimposed on the desired distance. The corrected desired pose is sent to the robot controller to achieve the unification of surface tracking and force control, and the tracking and force control errors are evaluated by indicators.

[0010] Step S5: Implement collaborative control and safety handling. Set a 20ms control cycle to ensure the real-time performance of sensor data acquisition, pose calculation, and command transmission. Monitor the effective measurement range of the sensors, the normal contact force threshold, and the joint angle limit in real time. If the range is exceeded or the signal is lost, immediately trigger deceleration and stop. Record abnormal events to the log file and control the robot to stay in place and sound an alarm when communication is interrupted, thus realizing the safety handling function of the entire process.

[0011] Preferably, step S1 involves: designing and conducting robot kinematics and stiffness calibration experiments; collecting actual pose data and force-deformation data of the robot end effector under multiple configurations; constructing a kinematic parameter and joint stiffness dataset; and establishing a kinematic error model for the dataset using velocity-adjoint transformation, defining the pose error matrix of the robot end effector in the base coordinate system {B} as follows:

[0012]

[0013] Where q is the nominal kinematic parameter, and Δq is the kinematic error to be identified; according to differential kinematics, the error matrix can be expressed as...

[0014]

[0015] in, E ΔT q The differential motion matrix corresponds to the differential motion vector.

[0016] E D q =[ E d x , E d y , E d z , E δ x , E δ y , E δ z ] T

[0017] And satisfy

[0018]

[0019] Δq is obtained by fitting using the least squares method, and kinematic parameter compensation is completed to eliminate the no-load positioning error;

[0020] A joint stiffness model is established using force-adjoint transformation on the dataset. A known external force is applied to the robot's end effector, and the deformation of the end effector is measured. The generalized force borne by the end effector is defined as follows: E F, the relationship between end-effector deformation and joint stiffness is:

[0021]

[0022] Where J is the Jacobian matrix, K θ =diag(k1, ..., k6) is the joint stiffness matrix. B F E The generalized force borne at the end is represented in the base coordinate system; in the end coordinate system {E}, the deformation vector satisfies the rotation transformation.

[0023] E D k =Sr( E R) B D Ek

[0024] Wherein, Sr is the spatial rotation transformation operator; through multiple sets of loading experiments, the stiffness coefficients of each joint are identified, and a comprehensive robot body model containing kinematic errors and joint stiffness is obtained;

[0025] Preferably, step S2 involves: acquiring a set of boundary points of the processing area through manual guidance or laser scanning, with the boundary point set represented by polygons, and constructing a point cloud dataset of the processing area; projecting the boundary points onto a path planning plane, and defining the projection transformation as...

[0026]

[0027] Among them, P 0p,i Let P be the three-dimensional position of the i-th boundary point in the base coordinate system. 0p R is the geometric center of the boundary point set. 0p The rotation matrix of the path planning coordinate system relative to the base coordinate system (its x... p y p The axes are parallel to the eigenvectors corresponding to the minimum and second smallest eigenvalues ​​of the point set's inertia tensor, respectively; after projection, the vertex v of the two-dimensional polygon is obtained. i ;

[0028] The two-dimensional polygon is decomposed using Boustrophedon cell decomposition along x p A sweep along the axis divides the region into non-overlapping cells, using critical points (polygon vertices) as cell boundaries, with cell widths set to 80% of the tool diameter; within each cell, a parallel axis is generated. p The sweep path of the axis, the spacing d between adjacent paths p Based on the tool radius r t The processing overlap rate η is determined to be d. p =2r t (1-η);

[0029] Dijkstra's algorithm is used to find the shortest connection path between cells, forming a continuous two-dimensional path skeleton. The two-dimensional path points are then mapped back to a three-dimensional surface, and the three-dimensional position p of each path point is obtained by bilinear interpolation using the boundary point cloud. 0s With the surface normal vector e s_plan Construct the desired tool pose matrix

[0030] R 0d =[e xd e yd ezd ], e zd =-e s_plan , e xd =e yd ×e zd

[0031] Among them, e t The path tangential unit vector ensures that the tool axis is antiparallel to the surface normal vector and the feed direction is aligned with the path tangential vector. The global path points are standardized and normalized to ensure that the position and pose data dimensions are consistent, providing input for subsequent pose optimization.

[0032] Preferably, step S3 involves: using global path points as the target point set {p1,...,p}. n} Calculate the uncompensated machining error at each target point, and the machining error caused by kinematic error is:

[0033]

[0034] Among them, E mz Let z be the transformation vector from the end coordinate system to the target coordinate system in the z-direction.

[0035] satisfy E pi0 For target point p i The position vector in the terminal coordinate system, [E] pi0 ] is its antisymmetric matrix; E dq and E δq These represent the translation error and rotation error in the end coordinate system, respectively, satisfying... This expression can be further written as

[0036]

[0037] Among them, Ad V Let D be the velocity-adjoint transformation matrix. z =[0, 0, 1, 0, 0, 0] T The machining error caused by joint stiffness is

[0038]

[0039] Where Ad is the force-associated transformation matrix, satisfying The nominal generalized force (force vector f) obtained by offline calibration based on design process parameters i and torque vector m i The overall error is a. qk =a q +a kConsidering the coupling effect of grinding force varying with actual grinding depth, the stiffness error is further corrected to...

[0040]

[0041] Among them, a des To design the grinding depth, X F The force-error mapping matrix is ​​used; error compensation is performed by selecting the workpiece frame position p0, so that the error at point p0 is zero after compensation, and the error at each point after compensation is...

[0042] a qkθ (p i ) = a qk (p1)-a qk (p0)

[0043] Convert the absolute error to a relative error, a qk (p0) represents the machining error at the workpiece frame position; an optimization objective function is constructed for the compensated machining error.

[0044]

[0045] Where r is the extreme radius of the tool in the horizontal plane, α is the tool azimuth angle, and the optimization constraints include the operability index κ(J)= / / J(q) / / × / / J + (q) / / ≤κ0、Joint limit θ j，low ≤θ ij ≤θ j，up The collision avoidance conditions and chatter stability region are determined. A genetic algorithm (population size 50, iteration 100 generations) is used to solve for the optimal solution, obtaining the optimal workpiece position p0 and the optimal tool posture (r, α), realizing the global pose optimization function. The optimization results are evaluated using the mean absolute error and root mean square error to ensure the convergence of the optimization effect.

[0046] Preferably, step S4 involves: during the robot's machining process, synchronously acquiring three sets of laser displacement sensor data and six-dimensional force / torque sensor data at servo cycles to construct a real-time perception dataset; and loading the optimal workpiece position p0 and optimal tool posture (r, α) obtained in step S3 into the control system as a global optimization benchmark.

[0047] Calculate the positions of three measurement points in the base coordinate system based on the laser displacement sensor data.

[0048]

[0049] Where, p 0e and R 0e p represents the current end-effector position and attitude, respectively. e,SF p e,SL pe,SR Let d represent the translational offset of the three sensors in the end coordinate system. F d L d R e represents the laser distance measurement value. z = [0, 0, 1] T ; Fit a local plane using three points, and calculate the plane normal vector e. s_meas and the actual distance d from the tool to the plane act ;

[0050] The desired position p is constructed by combining the local surface geometry information with the optimization benchmark. 0d =p 0s_opt +d des ·e s_meas , where d des p is the desired distance 0s_opt Based on the theoretically expected contact point determined by the optimal workpiece position p0 and the current path point, the desired pose is generated by combining the desired attitude matrix constructed in step S2 with the optimized tool azimuth angle α; gravity compensation is performed on the six-dimensional force / torque data to obtain the actual contact force f. c Projecting onto the normal direction yields the normal contact force.

[0051] f n =-(e s_meas ) T ·f c

[0052] A parallel force control strategy is adopted to achieve the desired normal force f. d Compared with the measured normal force f n The deviation is used as input, and a PI control law is used to calculate the distance correction.

[0053] d c =k P (f d -f n )+k I ∫(f d -f n )dτ

[0054] Where, k P For proportional gain, k I For the integral gain, the distance correction is superimposed on the desired distance, and an S3 force-error mapping model X is introduced. F Coupling compensation amount δp f The corrected expected position is obtained.

[0055] p 0c =p 0s_opt +(d des +d c )·e s_meas +δpf ;

[0056] The corrected desired pose is sent to the robot controller through the robot servo interface to achieve the unification of surface tracking and force control; the tracking error and force control error are evaluated using the mean absolute error and root mean square error to verify the control accuracy.

[0057] Preferably, step S5 involves: setting the control cycle to 20ms to ensure that sensor data acquisition, pose calculation, and command transmission are completed within the same cycle, thus guaranteeing system real-time performance; real-time monitoring of the effective measurement range of the laser sensor, the normal contact force threshold, and the joint angle limit; if any sensor exceeds the range or loses a signal, immediate deceleration and stop are triggered; abnormal events are recorded in a log file, and the robot remains in place and triggers an alarm when communication is interrupted, thus achieving full-process safety handling functions.

[0058] A six-degree-of-freedom robot grinding system with a weak prior model, characterized in that it includes:

[0059] The kinematics and stiffness calibration module is configured to: establish a robot body model containing kinematic errors and joint stiffness through calibration experiments;

[0060] The global path planning module is configured to perform cell decomposition and path planning based on the boundary point set to generate the desired tool pose covering the processing area.

[0061] The workpiece position and tool posture optimization module is configured to optimize the workpiece position and tool posture using a genetic algorithm based on the machining error model and error compensation.

[0062] The real-time surface tracking and force control module is configured to fuse laser and force sensor data to achieve surface tracking and parallel force control.

[0063] The collaborative control and safety processing module is configured to perform cycle synchronization control, sensor monitoring, and anomaly safety processing.

[0064] Compared with the prior art, the present invention has the following beneficial effects:

[0065] 1. This invention establishes a comprehensive error model by using a kinematics and stiffness calibration module and velocity-associated transformation and force-associated transformation. This model can simultaneously compensate for kinematic errors and machining errors caused by joint stiffness, thereby significantly improving the positioning accuracy and machining quality of the robot body.

[0066] 2. This invention employs a global path planning method that combines Boustrophedon cell decomposition with Dijkstra's algorithm. This method can automatically cover complex freeform surface processing areas and obtain the three-dimensional position and normal vector of path points by combining bilinear interpolation, thereby achieving fully automatic path generation under a weak prior model.

[0067] 3. This invention uses a workpiece position and tool posture optimization module to construct an objective function based on the compensated machining error and uses a genetic algorithm to solve for the optimal pose, which effectively reduces the average machining error, converts the absolute error into the relative error, and improves the overall machining consistency. Attached Figure Description

[0068] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings of the embodiments will be briefly described below.

[0069] Figure 1 This is a schematic diagram of a six-degree-of-freedom robot weak prior model grinding system and method provided in an embodiment of the present invention. Detailed Implementation

[0070] The following describes the implementation of the present invention through specific embodiments and in conjunction with the accompanying drawings.

[0071] To address the problems of insufficient machining accuracy and poor process stability in robotic grinding caused by kinematic errors, joint stiffness deformation, unknown surface geometry, and uncontrollable contact forces, this invention provides a high-precision collaborative control method for robotic grinding of complex surfaces. This method employs a collaborative control architecture combining "offline global pose optimization" and "online local real-time correction" to suppress system errors at their source and dynamically adapt to unknown geometry and disturbances during machining, thereby achieving high-precision and highly consistent automated grinding.

[0072] Figure 1 A six-degree-of-freedom robot weak prior model grinding system and method provided in the embodiments of the present invention will be described in detail below.

[0073] The present invention provides a six-degree-of-freedom robot weak prior model grinding system and method, comprising the following steps performed sequentially:

[0074] Step S1: Robot Comprehensive Error Modeling and Compensation

[0075] The purpose of this step is to establish and identify a comprehensive pose error model of the robot under grinding stress conditions, providing a quantitative basis for subsequent offline compensation. The core is to construct kinematic error models and joint stiffness models respectively.

[0076] First, a kinematic parameter error model is established based on the velocity-dependent transformation. The nominal kinematic parameter of the robot is defined as q, and its actual deviation is Δq. The resulting pose error matrix of the robot's end effector in the base coordinate system {B} is then given by...

[0077]

[0078] Differential kinematics can be used to map this pose error matrix into a differential motion vector in the end-effector coordinate system {E}. Furthermore, the kinematic parameter error Δq can be identified and compensated using the least squares method based on the calibration experimental data.

[0079] Secondly, a joint stiffness-deformation model is established based on force-dependent transformation. (This is related to the grinding contact force.) B F E Under the influence of force, the robot joints will deform due to their flexibility. Let the robot joint stiffness matrix be K. θ (Taking the ABB 6660 robot as an example, K) θ =diag(1,2,2,4,4,4)×10 6 Given N·m / rad), the robot's Jacobian matrix is ​​J. Then, in the base coordinate system, the pose deviation D caused by joint deformation at the end effector is... Ek for

[0080]

[0081] This deformation vector can be transformed to be described in the end coordinate system {E} using the spatial rotation transformation operator Sr, resulting in... E D k .

[0082] Finally, by integrating the two errors mentioned above, a comprehensive pose error model for the robot's end effector is obtained. This model can accurately predict the actual positional deviation of the robot's end effector in the grinding depth direction under specific poses and forces, and serves as the theoretical basis for subsequent offline pose optimization.

[0083] Step S2: Coverage path planning based on region boundaries

[0084] This step aims to generate continuous, complete grinding paths for machining areas of unknown surfaces without requiring a workpiece CAD model. The planning process is performed in a two-dimensional plane and then mapped back to three-dimensional space.

[0085] First, the boundary point set {p} of the processing area is obtained through collaborative robot hand-guided teaching or 3D vision measurement. 0p,i Projecting these three-dimensional boundary points onto an optimal two-dimensional programming plane, the projection formula is as follows:

[0086]

[0087] Where, p 0p Let R be the geometric center of the boundary point set, and let R be the rotation matrix. 0p The boundary point cloud is constructed from the eigenvectors (i.e., principal component directions) of the covariance matrix of the boundary point set to ensure that the projection plane best fits the boundary point cloud.

[0088] Next, on the two-dimensional plane, the projected set of points is used to construct a polygonal region. Using the Boustrophedon cell decomposition method, the polygon is decomposed along a single sweep direction, generating a series of non-overlapping simple cells. Within each cell, a set of parallel, equidistant cells (with a spacing of d) is generated. p The process involves a zigzag sweep path. Then, Dijkstra's shortest path algorithm is used to plan the optimal global path connecting the sweep paths of each cell, forming a continuous path Γ covering the entire processing area from the starting point to the ending point. 2D .

[0089] Finally, through inverse projection transformation, the two-dimensional path Γ is transformed. 2D Mapping back to the 3D robot workspace yields the global 3D path point set γ(s).

[0090] For the tool posture at each point on the path, to ensure that the grinding tool is always perpendicular to the surface at the desired angle, its desired posture matrix R... 0d The construction is as follows

[0091] R 0d =[e xd e yd e zd ], e zd =-e s_plan , e xd =e yd ×e zd

[0092] Among them, e t The tangent direction of the current path point (given by path planning), e s_plan This is the theoretical surface normal vector obtained through bilinear interpolation based on the boundary point cloud. This construction guarantees the tool's Z-axis (e zd The direction of the curve is opposite to the normal of the surface, and the Y-axis (e) yd Perpendicular to the path direction, thus establishing a stable tool feed coordinate system.

[0093] Step S3: Offline global pose optimization based on comprehensive error

[0094] This step utilizes the error model established in step S1 to perform offline optimization of the workpiece installation position and the initial tool posture, aiming to minimize the systematic machining error caused by the robot's own errors at a global level.

[0095] The three-dimensional path point set {p1, ..., p2} obtained in step S2 is used to... n} is the target point. For any point p... iUsing the model from step S1, the robot's overall error in the grinding depth direction (assuming it's the Z-axis direction of the tool coordinate system, D) can be calculated without compensation. z The resulting prediction error a qk (p i This error is the kinematic error a. q (p i ) and joint stiffness deformation error a k (p i ) and

[0096] a qk (p i ) = a q (p i )+a k (p i )

[0097] in,

[0098]

[0099] in, The nominal grinding force and torque are obtained through offline calibration based on the design process parameters.

[0100] To simplify the compensation strategy, the origin p0 of the workpiece coordinate system is selected as the reference point, and the absolute error of each point on the entire path is converted into a relative error relative to that point.

[0101] a qkθ (p i ) = a qk (p i )-a qk (p0)

[0102] This operation is equivalent to fine-tuning the workpiece position p0 so that the overall error distribution fluctuates around zero, rather than accumulating offset.

[0103] The optimization problem is formulated with the objective of minimizing the sum of the absolute values ​​of the relative errors along the entire path.

[0104]

[0105] The optimization variables include the workpiece position p0 and the tool attitude parameters (polar radius r and azimuth angle α). The optimization process must meet constraints such as robot maneuverability, joint limits, collision avoidance, and process stability.

[0106] A genetic algorithm (population size 50, 100 generations) was used to solve the above problem, obtaining the optimal workpiece mounting position p0 and the initial tool posture (r, α). This result provides the robot with a globally optimal initial machining posture, which can significantly suppress system errors and will serve as the initial reference for online control in step S4.

[0107] Step S4: Online Real-Time Surface Tracking and Adaptive Force Control

[0108] This step is an online correction process, designed to utilize sensors to detect and compensate for workpiece geometric errors, clamping errors, and residual model errors in real time during machining, achieving precise surface tracking and constant force grinding. This step uses the optimal workpiece position p0 and optimal tool posture (r, α) obtained from offline optimization in the preceding step S3 as a global benchmark, and performs online local sensing and compensation based on this benchmark.

[0109] The robot's end effector integrates three laser displacement sensors and a six-dimensional force / torque sensor. All sensor data are synchronously acquired within each robot servo control cycle (e.g., 20ms). No workpiece CAD model is required; the entire process is based on real-time measurement.

[0110] Based on the current actual pose of the end effector (p) fed back by the robot controller 0e R 0e The three-dimensional coordinates p of the three laser measuring points in the base coordinate system can be calculated. OPF p OPL ,p OPR The local plane fitting method is used to estimate the surface geometry in real time.

[0111] n p =(p OPL -p OPF )×(p OPR -P OPF ),

[0112] Where e s That is, the estimated unit normal vector e of the surface at the current contact point. s_meas ;

[0113] Furthermore, the directed distance d from the tool's center point to the local plane can be calculated. e,zp .

[0114] Local surface reference point p 0s The determination will be based on the theoretical path points derived from the optimized workpiece position p0 and the actual sensor measurement data.

[0115] Based on perceptual surface geometry, the desired tool position p is generated. 0d =p 0s_opt +ddes ·e s_meas , where d des The set desired tracking distance.

[0116] Simultaneously, inertial force and gravity compensation are performed on the six-dimensional force / torque sensor signal to obtain the actual grinding contact force f. c Projecting this force onto the normal direction of the surface yields the normal contact force.

[0117] A parallel force control strategy is adopted. In the position loop, with the aforementioned p... 0d The desired position. In the force ring, with the desired normal force f. d Compared with the measured f n The deviation is taken as input, and a PI controller calculates the position correction d required to maintain constant force. c

[0118] d c =k P (f d -f n )+k I ∫(f d -f n )dτ

[0119] Typical experimental parameters can be taken as follows: proportional gain k P =0.2mm / N, integral gain k I = 2mm / (N·s).

[0120] In addition, a force-error mapping model X based on S3 is introduced. F Coupled feedforward compensation δp f =X F ·(f n -f d The correction amount of the force loop is superimposed on the model feedforward amount to the desired position of the position control loop, and the final corrected desired position is sent to the robot servo.

[0121] p 0c =p 0s_opt +(d des +d c )·e s_meas +δp f

[0122] Combine the desired attitude R constructed by the method in step S2 and incorporating the azimuth angle α of the optimization tool. 0d To form the complete corrected desired pose (p) 0c R 0d This allows for the maintenance of a constant normal grinding force while tracking a globally optimized theoretical path.

[0123] To ensure real-time control and safety, the system adopts a hierarchical synchronous architecture. High-speed sensor signals (such as force sensors) are acquired and preprocessed at the programmable logic controller (PLC) with a shorter cycle (such as 2ms). The filtered and coordinate transformed data, along with the robot's current state information, is sent to the robot controller. Within the robot controller's servo cycle, the surface geometry calculation, force control algorithm calculation, and pose command generation and issuance in step S4 are executed sequentially. At the same time, the system monitors the validity of all sensor signals, whether the contact force exceeds the limit, and whether the robot joints are approaching the limit, etc., in real time. Once an abnormality is detected, a safety shutdown is immediately triggered to ensure the safe and reliable processing.

[0124] To verify the effectiveness of this invention, an experiment was conducted on an industrial robot system equipped with a six-dimensional force / torque sensor, a laser displacement sensor, and a pneumatic grinding tool. The robot controller's servo cycle was 20ms, and sensor data was synchronously acquired by a PLC at a 2ms cycle.

[0125] In benchmark tests without the application of the method of this invention, the maximum machining error in the grinding depth direction reached 0.522 mm due to robot kinematic errors and joint stiffness deformation. After applying the method of this invention, the maximum deviation caused by such systematic errors can be reduced to 0.061 mm simply through offline global pose optimization in step S3. Based on this, online real-time surface tracking and force control in step S4 are initiated, and the robot successfully achieves stable tracking of the unknown geometric surface. The fluctuation of the normal contact force is controlled within the target range of ±2N. Experimental results show that the offline / online collaborative control method proposed in this invention can significantly improve the absolute accuracy and process stability of robot grinding.

[0126] The present invention has been described above with reference to preferred embodiments, but these embodiments are merely exemplary. Various substitutions and modifications can be made to the present invention based on these embodiments, all of which fall within the scope of protection of the present invention.

Claims

1. The six-degree-of-freedom robot weak prior model grinding system and method provided by the present invention are characterized in that: Includes the following steps: S1: Robot kinematics and stiffness calibration, wherein S1 includes: S11: Design and conduct robot kinematics and stiffness calibration experiments, collect actual pose data and force deformation data of robot end effector under multiple configurations, and construct kinematic parameters and joint stiffness datasets; S12: Establish a kinematic error model for the dataset using velocity adjoint transformation, and define the pose error matrix of the robot end effector in the base coordinate system {B} as follows: Where q is the nominal kinematic parameter, and Δq is the kinematic error to be identified; according to differential kinematics, the error matrix can be expressed as... wherein E ΔT q is the differential motion matrix corresponding to the differential motion vector E D q = [ E d x , E d y , E d z , E delta x , E delta y , E delta z ] T And satisfy Δq is obtained by fitting using the least squares method, and kinematic parameter compensation is completed to eliminate the no-load positioning error; S13: Establish a joint stiffness model for the dataset using force-adjoint transformation, apply a known external force to the robot end effector, measure the end effector deformation, and define the generalized force borne by the end effector as... E F, the relationship between end-effector deformation and joint stiffness is: Where J is the Jacobian matrix, K θ =diag(k1, ..., k6) is the joint stiffness matrix. B F E The generalized force borne at the end is represented in the base coordinate system; in the end coordinate system {E}, the deformation vector satisfies the rotation transformation. E D k =Sr( E R) B D Ek Wherein, Sr is the spatial rotation transformation operator; through multiple sets of loading experiments, the stiffness coefficients of each joint are identified, and a comprehensive robot body model containing kinematic errors and joint stiffness is obtained; S14: Standardize and normalize the kinematic parameters and joint stiffness data to ensure data consistency and provide an accurate robot body model basis for subsequent pose optimization. S2: Workpiece surface boundary acquisition and global path planning, wherein S2 includes: S21: Design and conduct an experiment to obtain the boundary of the workpiece surface. Obtain the boundary point set of the processing area through manual teaching or laser scanning. The boundary point set is represented by polygons, and a point cloud dataset of the processing area is constructed. S22: The boundary point set is divided into non-overlapping cells by using Boustrophedon cell decomposition, with the critical point as the cell boundary, the sweep direction is set to be parallel to the long axis of the polygon, and the cell width is set to 80% of the tool diameter. S23: Generate parallel sweep paths within the cells. The spacing between adjacent paths is determined based on the tool size and processing overlap rate. Use Dijkstra's algorithm to solve for the shortest connection path between cells to form a continuous two-dimensional path skeleton. S24: Map the two-dimensional path points back to the three-dimensional surface, and use bilinear interpolation to obtain the three-dimensional position of each path point and the surface theoretical normal vector e. s_plan Construct the desired tool pose matrix R 0d =[e xd e yd e zd ],in And zd =-e s_plan And xd =and yd ×e zd Among them, e t For the path tangent, ensure that the tool axis is antiparallel to the normal vector and the feed direction is aligned with the path tangent; S25: Standardize and normalize the global path points to ensure that the position and pose data dimensions are consistent, providing input for subsequent pose optimization; S3: Workpiece position and tool posture optimization, wherein S3 includes: S31: Use the global path points as the target point set {p1, ..., p} n } Calculate the uncompensated machining error at each target point, and the machining error caused by kinematic error is: This expression can be further written as Among them, Ad V Let D be the velocity-adjoint transformation matrix. z =[0, 0, 1, 0, 0, 0] T ; S32: Machining error caused by joint stiffness is The overall error is a qk =a q +a k Considering the coupling effect of grinding force varying with actual grinding depth, the stiffness error is further corrected to... Among them, a des To design the grinding depth, X F The force-error mapping matrix; S33: For the target point set, error compensation is performed by selecting the workpiece frame position p0, so that the error at point p0 is zero after compensation, and the error of each point after compensation is a. qkθ (p i ) = a qk (p i )-a qk (p0) S34: Construct an optimization objective function for the compensated processing error. Where r is the extreme radius of the tool in the horizontal plane, α is the tool azimuth angle, and the optimization constraints include the operability index κ(J)= / / J(q) / / × / / J + (q) / / ≤κ0、Joint limit θ j，low ≤θ ij ≤θ j，up Collision avoidance conditions and flutter stability zone; S35: The optimal solution is obtained by using a genetic algorithm for the objective function. The population size is set to 50, and the iteration is performed for 100 generations to obtain the optimal workpiece position p0 and the optimal tool posture (r, α), thereby realizing the global pose optimization function. S36: The optimization results are evaluated using the mean absolute error and root mean square error to ensure that the optimization effect converges. S4: Real-time surface tracking and force control, wherein S4 includes: S41: System initialization, load the optimal workpiece position p0 and optimal tool posture (r, α) obtained by optimization in S3. This optimal tool posture is used to construct the theoretical expected tool direction reference, and p0 is set as the reference origin of the workpiece coordinate system {W}. S42: Design and conduct real-time surface tracking and force control experiments. During the robot's machining process, collect three sets of laser displacement sensor data and six-dimensional force / torque sensor data synchronously with the robot's servo cycle to construct a real-time perception dataset. S43: Calculate the positions of the three measurement points in the base coordinate system based on the laser displacement sensor data. p 0，PF &=P 0e +R 0e ·P e，SF +d F ·R 0e ·e z p 0，PL &=P 0e +R 0e ·P e，SL +d L ·R 0e ·e z p 0，PR &=P 0e +R 0e ·p e，SR +d R ·R 0e ·e z Where, p 0e and R 0e p represents the current end-effector position and attitude, respectively. e，SF p e，SL P e，SR Let d represent the translational offset of the three sensors in the end coordinate system. F d L d R e represents the laser distance measurement value. z [0, 0, 1] T ; Calculate the local actual normal vector e by fitting a local plane using three points. s_meas and the actual distance d from the tool to the plane act ; S44: Calculate the desired pose based on the optimized baseline. First, based on the current path parameters and the optimization results (p0, r, α), calculate the theoretical expected contact position P of the current point. 0s_opt Simultaneously, the estimated local actual contact point p is obtained from the sensor data in step S43. 0s_meas Construct the desired position P 0d =p 0s_opt +d des ·e s_meas S45: Perform gravity compensation on the six-dimensional force / torque sensor data to obtain the actual contact force f. c Projecting onto the normal direction yields the normal contact force. f n =-(e s_meas ) T ·f c S46: A parallel force control strategy is adopted for the normal contact force to achieve the desired normal force f. d Compared with the measured normal force f n The deviation is used as input, and a PI control law is used to calculate the distance correction. d c =k P (f d -f n )+k I ∫(f d -f n )dτ Introducing force-error coupling compensation, based on the force-error mapping model X in S3. F Calculate the additional position compensation δp based on the current force deviation. f =X F ·(f n -f d The distance correction and coupling compensation are added together to the desired distance to obtain the corrected desired position. p 0c =p 0s_opt +(d des +d c )·e s_meas +δp f S47: The corrected desired pose is sent to the robot controller through the robot servo interface to achieve the unification of surface tracking and force control; S48: The tracking error and force control error are evaluated using the average absolute error and root mean square error to verify the control accuracy; S5: Collaborative control and security processing, wherein S5 includes: S51: Set the control cycle to 20ms. Sensor data acquisition, pose calculation and command transmission are completed within the same cycle to ensure system real-time performance. S52: Real-time monitoring of the effective measurement range of the laser sensor, the normal contact force threshold, and the joint angle limit. If any sensor exceeds the range or the signal is lost, deceleration and stop will be triggered immediately. S53: Record the abnormal events to the log file. When communication is interrupted, the robot remains in place and alarms, realizing the full-process safety handling function.

2. The six-degree-of-freedom robot weak prior model grinding system and method according to claim 1, characterized in that: In S11, the experimental device mainly consists of a six-degree-of-freedom industrial robot, a laser tracker, a six-dimensional force / torque sensor, a data acquisition system, and an end effector. The laser tracker enables the measurement of the actual pose of the robot end in multiple configurations, with a measurement accuracy of up to the micrometer level. A six-dimensional force / torque sensor was used to collect force and torque data of the end under different loading conditions in order to establish a force-deformation mapping relationship. A data acquisition system was used to synchronously record robot joint angles, end-effector pose, end-effector force, and deformation data to construct a comprehensive dataset. The kinematic parameters included Denavit-Hartenberg parameters such as link length, link offset, joint rotation angle, and joint torsion angle for each joint, as well as additional geometric parameters such as reduction ratio and coupling coefficient. The joint stiffness data included the stiffness coefficients of each joint under loading conditions. To cover a wide range of poses and force states within the robot's workspace, at least 30 representative joint configurations were selected in the experimental design. Generalized forces in multiple directions were applied to each configuration. The magnitude of the forces was set according to the robot's rated load and processing conditions, typically ranging from 50N to 500N, and the torque range was from 10N·m to 100N·m. Data was collected at least three times for each configuration to reduce random measurement errors. Finally, a comprehensive dataset containing robot joint angles, end-effector pose, end-effector force, and deformation was constructed, providing a foundation for subsequent kinematic error identification and joint stiffness identification.

3. The six-degree-of-freedom robot weak prior model grinding system and method according to claim 1, characterized in that: In step S24, the two-dimensional path points generated in S23 are mapped back to the three-dimensional surface using bilinear interpolation. Using the boundary point cloud data constructed in S21, bilinear interpolation is performed on each two-dimensional path point within its corresponding cell to obtain the three-dimensional position coordinates and the surface normal vector e corresponding to that path point. s_plan The bilinear interpolation is based on the position and normal of the four neighboring boundary points of the grid where the path point is located, to ensure that the mapped three-dimensional path points are continuously and smoothly distributed on the actual surface of the workpiece.

4. The six-degree-of-freedom robot weak prior model grinding system and method according to claim 1, characterized in that: In step S42, based on the three sets of laser displacement sensor data synchronously acquired in step S41 and the current end-effector pose of the robot, the three-dimensional position coordinates of the three measurement points in the base coordinate system are calculated. The three laser displacement sensors are fixedly installed on the end-effector and are located at the front end, left end, and right end, respectively. Each sensor has a known translational offset p in the end-effector coordinate system {E}. e，SF p e，SL p e，SR and rotation matrix R e，SF R e，SL R e，SR The measurement direction of each sensor is along the z-axis of its own coordinate system, and the measured value d F d L d R These represent the distances from the sensor's emission point to the measured point on the workpiece surface.

Images