A method for planning a motion path of an industrial robot
By constructing a dynamic model of the hidden space of a flexible body and improving the path planning algorithm, the problems of real-time performance and safety in the flexible load handling of industrial robots were solved, and efficient and safe path planning was achieved.
Patent Information
- Application Number
- CN202611538716.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-30
- Publication Date
- 2026-11-17
AI Technical Summary
Existing industrial robot path planning methods cannot balance dynamic deformation perception accuracy and online real-time calculation when dealing with flexible loads, leading to risks of collision, breakage, and entanglement. Furthermore, their high computational complexity fails to meet the robot's underlying real-time interpolation cycle requirements.
A dynamic model of the latent space of a flexible body is constructed. The deformation and strain state of the flexible body are predicted by an encoder and decoder network. Combined with an improved fast extended random tree algorithm and nonlinear model predictive control, a motion path that meets the requirements of real-time performance and safety is generated.
It achieves sub-millisecond deformation prediction, reduces collision accidents, lowers the risk of breakage and entanglement, meets the real-time interpolation cycle requirements of robots, and the generated instructions can be directly executed by industrial controllers.
Smart Images

Figure CN123103679A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of industrial robot control technology, specifically to a method for planning the motion path of an industrial robot. Background Technology
[0002] As industries such as 3C electronics manufacturing, new energy vehicle assembly, and medical device assembly transform towards highly flexible and automated production, industrial robots are increasingly undertaking handling tasks involving highly flexible and large deformation loads. Typical loads include flexible bodies such as automotive wiring harnesses, flexible circuit boards, aerospace composite soft fabrics, and medical catheters. The basic task of industrial robot motion path planning is to search for a collision-free path that satisfies kinematic and dynamic constraints in the joint state space or Cartesian space of the robotic arm, given the initial and target configurations. Furthermore, the path is smoothed and interpolated to generate a sequence of position, velocity, and torque commands that can be executed by the controller.
[0003] Currently, mainstream path planning methods in the industry are all based on the rigid body assumption, which treats the end effector and its gripped load as a rigid envelope with a fixed shape. They use sampling-based path search algorithms (such as fast expanding random trees and their star-shaped improved algorithms) or optimization-based path planning algorithms (such as covariant Hamiltonian optimization motion planning algorithms and trajectory optimization algorithms) to perform collision detection and path optimization in joint space. Then, a smooth trajectory is generated through polynomial interpolation or spline fitting and output to the robot controller for execution.
[0004] The closest existing technical solution to this invention is a robotic arm path planning method based on a fast expanding random tree star topology algorithm. Its processing is roughly as follows: several candidate configuration points are randomly sampled in the joint state space of the robotic arm. Based on the rigid body envelope (i.e., the fixed geometric model of the end effector and load), a geometric collision detection algorithm is used to determine whether the candidate configuration collides with static obstacles in the workspace. A cost function is constructed based on the path length or motion energy consumption. The path that satisfies the cost condition is gradually approximated through the expansion and reconnection operations of the random tree. The discrete path point sequence obtained by the search is used to generate a continuous trajectory using polynomial or spline interpolation. The trajectory is output to the robot controller for interpolation. For cases involving flexible loads, some existing solutions introduce finite element numerical simulation methods to perform numerical calculations on the deformation of the flexible body in the offline stage, or use a method of slowing down the operation combined with repeated stops and corrections by an external vision camera in the online stage to avoid collision risks.
[0005] The aforementioned existing technical solutions have the following drawbacks: First, because existing path planning methods treat the end-effector load as a fixed geometric envelope, their collision detection models do not include the drooping and swinging behaviors of flexible bodies that dynamically change with their motion states (velocity, acceleration). This results in the robot, when transporting flexible loads at high speeds, having no collisions at the rigid body level in the planned path, but the actual deformation of the flexible body exceeds the preset envelope range, leading to collisions or scrapes with the environment. Second, because existing methods lack modeling of the internal tension and topological constraints of materials when multi-arm collaboratively transporting flexible components such as cables, the cost function lacks an evaluation mechanism for the strain state of the flexible body. This leads to the risk of fracture due to overstretching of the flexible body during coordinated motion, or self-entanglement and knotting due to slackness caused by improper motion coordination. Third, since accurate prediction of flexible body deformation usually relies on finite element numerical simulation methods, the computational complexity of such methods increases nonlinearly with the size of the discretized mesh, and the time taken for a single solution is usually on the order of tens to hundreds of milliseconds. As a result, such methods cannot be embedded in the interpolation cycle of the robot's underlying layer, which is on the order of milliseconds, for real-time operation. Existing industrial solutions can only be forced to adopt compromise solutions such as slow-down quasi-static operation or repeated pauses and corrections relying on external vision, sacrificing production cycle time and automation efficiency.
[0006] In summary, existing industrial robot path planning technologies generally suffer from a technical contradiction when handling flexible body load handling tasks: they cannot simultaneously achieve the accuracy of dynamic deformation perception of flexible bodies and the real-time performance of online calculations. There is an urgent need for a motion path planning method that can simultaneously realize dynamic collision avoidance of flexible bodies and safety assessment of material internal forces while meeting the real-time interpolation cycle requirements of the robot's underlying layer. Summary of the Invention
[0007] The technical problem this invention aims to solve is: how to accurately predict the dynamic deformation envelope and internal strain state of a flexible body during the motion path planning process of an industrial robot handling a highly flexible, large-deformation load, while meeting the real-time requirement of the robot's underlying millisecond-level interpolation cycle, and integrate the prediction results into the obstacle avoidance cost assessment and trajectory smoothing control stages of path search, thereby avoiding collisions between the flexible body and the environment, avoiding the risk of material breakage due to excessive stretching or entanglement due to relaxation, and ensuring that the generated motion commands can be smoothly executed by an industrial standard controller without exciting high-frequency vibrations in the flexible body.
[0008] To address the aforementioned technical problems, this invention provides a motion path planning method for industrial robots. The overall technical solution is divided into an offline stage and an online stage. The offline stage completes the construction and training of the dynamic model of the latent space of the flexible body. This model includes an encoder network and a decoder network. The encoder network is used to map the displacement field, velocity field, and external force field of the discrete sampling points of the flexible body at the current moment into low-dimensional latent variables. The decoder network is used to reconstruct the low-dimensional latent variables into the deformation field estimate and strain field estimate of the flexible body.
[0009] During training, the physical residual loss is constructed based on the deformation field estimate output by the decoder network and combined with the elastic dynamics control equation. The physical residual loss is then weighted and summed with the data fitting loss that measures the deviation between the reconstructed deformation field and the real simulation data to construct the total training loss. Based on this, the parameters of the encoder network and decoder network are iteratively updated so that the network output satisfies the basic physical laws while fitting the simulation data.
[0010] During the online phase, the current joint status of the robotic arm, the current deformation status of the flexible body, and information on obstacles in the workspace are first obtained.
[0011] Subsequently, based on the improved fast extended random tree star algorithm, random sampling is performed in the joint state space of the robotic arm to obtain candidate configuration points. For the end motion state corresponding to the candidate configuration points, the trained flexible body latent space dynamic model is called to perform deformation prediction, and the corresponding deformation field prediction value and strain field prediction value are obtained.
[0012] Based on the predicted values of the deformation field and strain field, a composite cost function is constructed by combining the rigid body collision penalty cost of the robotic arm body, the flexible body dynamic deformation envelope collision penalty cost, the material strain energy penalty cost, and the path length cost. The candidate configuration points are comprehensively evaluated, and the candidate configuration points that satisfy the collision safety constraints and strain safety constraints are added to the random tree and neighborhood reconnection is performed. After iterative expansion, the joint configuration sequence from the initial configuration to the target configuration is obtained.
[0013] Then, using this joint configuration sequence as reference data points, B-spline curves are fitted to obtain a continuous and differentiable reference path;
[0014] Finally, using this reference path as the reference trajectory, the nonlinear model predictive control is used to solve the control sequence that satisfies the joint position limit constraints, acceleration constraints, and jump constraints in the prediction time domain. The solved control commands are then output to the robot controller for execution, driving the robotic arm to complete the transport of the flexible load.
[0015] The technical effects that this invention can achieve include:
[0016] First, the computational delay for predicting the deformation of flexible bodies is compressed from tens to hundreds of milliseconds in finite element numerical simulation to the sub-millisecond level required for neural network forward inference, meeting the real-time interpolation cycle requirements of the robot's underlying layer without relying on slow-down operation.
[0017] Secondly, the spatial relationship between the dynamic deformation envelope of the flexible body and obstacles in the workspace can be assessed during the path search stage, reducing the probability of on-site scraping and collision accidents.
[0018] Third, by constraining the internal force state of the flexible body during multi-arm collaborative handling through strain energy penalty, the probability of cable breakage and entanglement is reduced.
[0019] Fourth, after the generated motion commands are smoothed by B-spline fitting and nonlinear model predictive control, the jump and acceleration are clearly constrained and can be directly executed by existing industrial controllers, reducing the possibility of exciting high-frequency vibrations of flexible bodies. Moreover, the overall solution of this invention is compatible with existing robot operating system frameworks or programmable logic control architectures, and the engineering deployment and transformation costs are low. Attached Figure Description
[0020] Figure 1 This is a diagram of the industrial robot motion path planning system architecture of the present invention. Detailed Implementation
[0021] The following is in conjunction with the appendix Figure 1 The specific embodiments of the present invention will be further described below. It should be noted that the description of these embodiments is for the purpose of helping to understand the present invention, but does not constitute a limitation of the present invention.
[0022] Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.
[0023] I. System Overall Architecture
[0024] The system architecture upon which the industrial robot motion path planning method of this invention relies includes interconnected modules: a simulation data generation module, a latent space dynamic model training module, an online sensing module, a deformation-sensing path search module, a trajectory smoothing module, and a robot controller execution module. The simulation data generation module generates deformation datasets of the flexible body under different end-effector motion states using a physical simulation engine in an offline phase. The latent space dynamic model training module receives this deformation dataset and constructs and trains a latent space dynamic model of the flexible body, including an encoder network and a decoder network. The online sensing module acquires the current joint state of the robotic arm, the initial shape of the flexible body, and information on obstacles in the workspace. The path search module calls the trained flexible body latent space dynamic model to predict deformation and combines it with a composite cost function to complete the path search. The trajectory smoothing module includes a B-spline path fitting unit and a nonlinear model prediction control optimization unit, which fits and optimizes the discrete configuration sequence obtained from the path search. The robot controller execution module receives the joint position, velocity, and acceleration commands output by the trajectory smoothing module and drives the robot arm to move. The above modules are connected in sequence according to the data flow direction. After offline training is completed, the network parameters obtained by the latent space dynamic model training module are provided to the deformation-aware path search module for calling in a unidirectional transmission manner. In the online stage, the modules are executed in series in sequence.
[0025] II. Deformation Prediction Method Based on Dynamic Model of Latent Space of Flexible Body
[0026] The dynamic model of the latent space of a flexible body includes an encoder network, a latent variable propagation unit, and a decoder network. The input to the encoder network is the displacement field, velocity field, and external force field of the flexible body at the current moment, which are discrete sampling points. The external force field includes the gravitational component and the inertial force component caused by the acceleration of the robotic arm's end effector. The output of the encoder network is low-dimensional latent variables. The latent variable propagation unit receives the current latent variables and the motion state instructions (including position, velocity, and acceleration) of the robotic arm's end effector at several future moments, and predicts the sequence of latent variables at future moments. The decoder network reconstructs the latent variables into estimated values of the deformation field and strain field of the flexible body, which are then used for subsequent collision detection and strain assessment. In a specific embodiment, both the encoder network and the decoder network can be implemented using a multilayer perceptron structure. The dimension of the latent variables can be set to tens of dimensions according to the geometric complexity of the flexible body. The specific number of layers and the number of neurons in each layer of the encoder network and decoder network can be adjusted according to the number of discrete sampling points of the flexible body and the scale of the simulation data. Those skilled in the art can determine the specific network hyperparameters according to the actual type of flexible body after reading this specification, without affecting the implementation of the technical solution of this invention.
[0027] The relationship between latent variables and decoding reconstruction can be expressed as:
[0028] ;
[0029] ;
[0030] In the above formula, express Low-dimensional latent variables corresponding to the deformation state of the flexible body at any given time. The parameter is encoder network, express The displacement field, i.e., the deformation field, of discrete sampling points of the flexible body at any given time. This represents the corresponding velocity field. express The external force field acting on a flexible body at any given moment This represents the deformation field estimate obtained from the decoder network reconstruction. The parameter is The decoder network.
[0031] To ensure that the network output satisfies fundamental physical laws while fitting simulation data, a physical residual loss is introduced during training. This physical residual is constructed based on the elastic dynamics governing equations. Specifically, the second derivative of the deformation field estimate from the decoder network output with respect to time is used to obtain the acceleration field estimate. The product of the flexible material's mass density and this acceleration field estimate, the product of the material's damping coefficient and the deformation field velocity estimate, and the divergence term of the stress tensor determined by constitutive relations (such as the generalized Hooke's law) are added together. Then, the external load field applied to the flexible body at the current moment is subtracted to obtain the physical residual. Ideally, this physical residual should approach zero. The magnitude of the physical residual is used to measure the degree to which the network output satisfies the elastic dynamics governing equations. To clarify this relationship, the physical residual can be expressed in the following form:
[0032] ;
[0033] In the above formula, Representing latent variables The corresponding physical residuals, This indicates the mass density of the flexible material. This represents the second derivative of the reconstructed deformation field with respect to time, i.e., the estimated value of the acceleration field. Indicates the material damping coefficient. This represents the stress tensor determined by the constitutive relation, which is the reconstructed deformation field. The function, The divergence of the stress tensor represents the internal forces within the material. express An external load field is constantly applied to the flexible body.
[0034] The total training loss of the network consists of three parts: data fitting loss, physical residual loss, and latent variable temporal smoothing regularization term, which are weighted and summed using preset positive weighting coefficients. The data fitting loss measures the deviation between the network's reconstructed deformation field estimate and the actual deformation field label obtained from offline simulation. This deviation can be calculated using the mean square error method. The physical residual loss measures the overall magnitude of the aforementioned physical residuals on the training sample set. The smaller the physical residual value, the more the network output conforms to the constraints of the elastic dynamics control equation. The latent variable temporal smoothing regularization term constrains the variation amplitude of latent variables between consecutive time steps and suppresses non-physical jumps in latent variables. After summing the above three losses according to their respective weighting coefficients to obtain the total training loss, the parameters of the encoder network and decoder network are iteratively updated using a gradient descent-type optimization method until the total training loss converges to below a preset threshold. The network parameters after training are then saved.
[0035] The offline training process specifically includes: using a physical simulation engine, performing batch numerical simulations on the deformation response of the target flexible body under different end-effector motion trajectories, generating a training dataset containing deformation field, velocity field, external force field, and corresponding real deformation field labels. This physical simulation engine can be implemented based on a mass-spring model or a finite element model. An encoder network and a decoder network are constructed. The encoder network maps the deformation field, velocity field, and external force field in the simulation data into low-dimensional latent variables. The decoder network reconstructs and restores the latent variables to the deformation field estimate. Based on the reconstructed deformation field estimate and the elastic dynamics control equation, the physical residual is calculated, and the physical residual loss is constructed accordingly. The data fitting loss, physical residual loss, and latent variable smoothing loss are summed with weighted coefficients to obtain the total training loss. The gradient descent optimization algorithm is used to iteratively update the parameters of the encoder network and decoder network until the total loss converges to below a preset threshold. The trained encoder network parameters and decoder network parameters are saved as the model to be called in the online stage deformation prediction.
[0036] The online inference process specifically includes: acquiring the current end-effector motion state (position, velocity, acceleration) of the robotic arm and the current observed deformation state of the flexible body. This observed deformation state can be obtained by torque sensor estimation or by visual initial calibration. The current deformation state, velocity state, and external force state are input into the trained encoder network to obtain the latent variables at the current moment. Combined with the candidate motion commands to be evaluated in the future, the latent variable propagation unit recursively predicts the sequence of latent variables at several future moments. The predicted latent variable sequences are input into the decoder network respectively to reconstruct the estimated values of the deformation field and strain field of the flexible body at the corresponding moment, which are then used by the subsequent collision detection and strain energy evaluation units. In a specific embodiment, the prediction step number of the latent variable propagation unit can be set to correspond to the motion duration on the order of hundreds of milliseconds to one second in the future. The specific prediction step number can be adjusted according to the robotic arm's motion speed and the obstacle distribution density in the workspace.
[0037] III. Composite Cost Function of Deformation Sensing and Improved Path Search Method
[0038] The calculation process of the deformation-sensing composite cost function includes the following sequential processing steps: The rigid body collision detection unit performs traditional geometric collision detection on the robotic arm body and end effector to obtain the rigid body collision penalty cost; the flexible body deformation prediction calling unit calls the previously trained flexible body latent space dynamic model to predict the deformation of the end effector motion state corresponding to the candidate path points; the dynamic envelope collision detection unit fits the flexible body spatial envelope surface based on the decoded deformation field and calculates the spatial distance between this envelope surface and the workspace obstacle model; when the spatial distance between the envelope surface and the obstacle is less than a preset safety distance threshold, a positive penalty cost related to this distance difference is generated; otherwise, the cost is zero. This yields the flexible body dynamic deformation envelope collision penalty cost, which is then evaluated by the strain energy assessment unit. The estimation unit calculates the strain energy penalty cost based on the deviation between the estimated strain field value obtained from decoding and the preset safe strain reference threshold, combined with the equivalent stiffness coefficient of the flexible material. The safe strain reference threshold is set according to the allowable tensile limit and allowable relaxation limit of the material. The cost aggregation unit weights and sums the rigid body collision penalty cost of the robot arm body, the dynamic deformation envelope collision penalty cost of the flexible body, the material strain energy penalty cost, and the path length cost according to their respective positive weighting coefficients, and outputs the comprehensive cost of the candidate path point. The weighted summation of the above four costs constitutes a composite cost function, where the path length cost is used to measure the motion energy consumption or stroke length of the joint spatial trajectory, and the other three costs correspond to the evaluation results of the robot arm body safety, the flexible body environmental safety, and the flexible body material internal force safety, respectively.
[0039] The improved fast expanding random tree star topology algorithm, based on the traditional fast expanding random tree star topology algorithm, expands the cost function from a single path length cost to a composite cost function that includes the dynamic collision cost and strain energy cost of the flexible body. Its execution process is as follows: Initialize the random tree, adding the robot arm's initial joint configuration as the root node to the tree structure, and setting the target joint configuration, workspace obstacle model, safety distance threshold, and safety strain threshold. Random sampling is performed in the robot arm's joint state space to obtain candidate configuration points. The nearest neighbor node search method is used to determine the existing node in the random tree that is spatially closest to the candidate configuration point as the expansion starting point. An expansion path segment is generated along the direction from the expansion starting point to the candidate configuration point at a preset step size. The aforementioned online inference process is then invoked to predict the latent variables and deformation field of the flexible body for the end-effector motion state sequence corresponding to the expansion path segment. Based on the prediction... The obtained deformation field and strain field are combined with the aforementioned composite cost function to calculate the comprehensive cost of the extended path segment. It is then determined whether the dynamic envelope collision cost and strain energy cost exceed the preset safety threshold. If the extended path segment satisfies the collision safety constraint and strain safety constraint, the candidate configuration point is added to the random tree as a new node, and the existing nodes in the neighborhood of the new node are reconnected. That is, the cost of reaching the neighborhood node through the new node is compared with the original cost of the neighborhood node. If the former is lower, the parent node of the neighborhood node is updated to achieve continuous optimization of the local path cost. It is then determined whether the new node has entered the neighborhood of the target configuration or whether the current iteration number has reached the preset value. If neither is satisfied, the sampling and expansion are continued. If the termination condition is met, the random tree is backtracked from the target node to the root node, and the corresponding joint configuration sequence is output as the initial path obtained by the search.
[0040] IV. Trajectory Smoothing Method Based on B-Spline Fitting and Nonlinear Model Predictive Control
[0041] The trajectory smoothing module includes a B-spline path fitting unit and a nonlinear model predictive control rolling optimization unit. The B-spline path fitting unit receives the discrete joint configuration sequence obtained by the improved path search, uses the discrete point sequence as the fitting data points, and uses the least squares fitting method to solve the control points of the B-spline curve to obtain a continuous and differentiable reference path represented by normalized path parameters. The reference path is then resampled at equal time intervals according to the robot control cycle to obtain a discretized reference trajectory sequence.
[0042] The nonlinear model predictive control rolling optimization unit uses the aforementioned reference trajectory sequence as the tracking target. In the prediction time domain, it constructs a rolling optimization objective function containing joint jump cost, joint acceleration cost, and reference trajectory tracking error cost. Joint position limit constraints, joint acceleration upper limit constraints, and joint jump upper limit constraints are applied. The objective function is solved using a sequential quadratic programming approach to obtain the control sequence that satisfies the constraints in the current prediction time domain. The joint position command, velocity command, and acceleration command corresponding to the current control cycle in the solved control sequence are output to the robot controller for execution. The solution results of the remaining prediction steps are only used for internal calculations in this rolling optimization process and are not directly output. After execution, in the next control cycle, the actual joint state of the robotic arm is reacquired as the new initial state, and the above rolling optimization process is repeated to form a closed-loop execution process of rolling time-domain optimization until the trajectory reaches the end of the reference path. In a specific embodiment, the predicted time-domain length can be set to the number of steps in the tens to hundreds of milliseconds according to the robot control cycle and the dynamic characteristics of the flexible body. The specific values of the jump upper limit constraint and the acceleration upper limit constraint can be calibrated and set according to the joint driving capability of the robotic arm and the allowable vibration frequency range of the flexible body. By applying the above constraints, the final output control command can meet the requirements of tracking the reference path while avoiding motion abrupt changes that excite high-frequency vibrations of the flexible body.
[0043] V. Extended Implementation Example of Dual-Robot Collaborative Handling
[0044] For applications involving the collaborative handling of flexible loads by two robotic arms, the end effectors of the two arms clamp the two ends of the flexible load, with the middle section of the load exhibiting a natural drooping shape. This invention constructs a unified joint state space by combining the joint states of the two robotic arms. The external force field input of the flexible body's implicit dynamic model simultaneously includes the inertial force contribution components corresponding to the motion states (position, velocity, acceleration) of the end effectors of the two robotic arms. The aforementioned composite cost function and the nonlinear model predictive control rolling optimization objective function are both constructed and solved uniformly on this joint state space. That is, during the path search process, the evaluation of candidate configuration points simultaneously considers the joint influence of the motion states of the two robotic arms on the deformation of the flexible body. During the trajectory smoothing process, the control sequences of the two robotic arms are jointly solved to ensure that the dynamic envelope collision cost and material strain energy penalty cost of the flexible body during the collaborative motion simultaneously meet the safety constraints, avoiding the risk of excessive stretching and fracture or loosening and entanglement of the flexible body due to improper coordination of the two robotic arms.
[0045] VI. Alternative Implementation Methods
[0046] In the construction of the latent space dynamic model of flexible bodies, in an alternative embodiment of the present invention, instead of using a network structure that combines encoder and decoder with physical residual loss, a recurrent neural network or attention mechanism network can be used to construct a purely data-driven latent space dynamic prediction model. In this case, only a physical consistency constraint term based on the elastic dynamic equation is added to the training loss function in a regularized form, without explicitly constructing a backpropagation path for the physical residual. This alternative embodiment can also achieve low-dimensional latent space prediction of flexible body deformation and maintain the online inference speed at the order of magnitude that meets the robot interpolation cycle requirements. It can be used as an alternative implementation method for the latent space dynamic model of flexible bodies in the present invention.
[0047] In the path search stage, in an alternative embodiment of the present invention, the improved fast extended random tree star topology algorithm based on sampling can be abandoned, and instead an optimization-based path planning framework can be used. The deformation-aware composite cost function (including dynamic envelope collision penalty term and strain energy penalty term) described in the present invention is used as one of the objective functions or constraints of the optimization framework, and the subsequent trajectory smoothing module based on nonlinear model predictive control and B-spline remains unchanged. This alternative embodiment is basically the same as the aforementioned main embodiment in terms of overall system architecture and technical effect, and can also achieve the purpose of flexible body deformation-aware path planning. It can be used as an alternative implementation of the path search stage of the present invention.
[0048] In summary, this invention compresses the high-dimensional deformation prediction problem into a low-dimensional neural network forward inference problem by constructing a flexible body latent space dynamic model. It achieves simultaneous assessment of the dynamic safety of the flexible body by introducing deformation-aware dynamic collision penalties and strain energy penalties into the path search cost function. Furthermore, it ensures that the output commands meet the constraints for execution by the industrial controller by introducing nonlinear model predictive control and B-spline trajectory smoothing mechanisms. These three elements work together to achieve safe planning of the load transport path for the flexible body while meeting real-time requirements. The above embodiments are only for illustrating the technical solution of this invention and are not intended to limit the scope of protection of this invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.
Claims
1. A method for planning the motion path of an industrial robot, characterized in that, include: A latent space dynamic model of a flexible body is constructed, comprising an encoder network and a decoder network. The encoder network maps the displacement field, velocity field, and external force field of the flexible body at the current time discrete sampling points into low-dimensional latent variables. The decoder network reconstructs the low-dimensional latent variables into estimated values of the deformation field and strain field of the flexible body. During the training of the latent space dynamic model, a physical residual loss is constructed based on the estimated value of the deformation field output by the decoder network and the elastic dynamics control equation. The physical residual loss and the data fitting loss are weighted and summed to construct the total training loss. The parameters of the encoder network and the decoder network are iteratively updated based on the total training loss. Acquire the current joint status of the robotic arm, the current deformation status of the flexible body, and information on obstacles in the workspace; Based on the improved fast extended random tree star algorithm, random sampling is performed in the joint state space of the robotic arm to obtain candidate configuration points. For the end motion state corresponding to the candidate configuration points, the trained flexible body latent space dynamic model is called to perform deformation prediction, and the deformation field prediction value and strain field prediction value corresponding to the candidate configuration points are obtained. Based on the predicted values of the deformation field and the predicted values of the strain field, a composite cost function is constructed by combining the rigid body collision penalty cost of the robotic arm body, the flexible body dynamic deformation envelope collision penalty cost, the material strain energy penalty cost, and the path length cost. Candidate configuration points are evaluated based on the composite cost function. Candidate configuration points that satisfy the collision safety constraints and strain safety constraints are added to a random tree and neighborhood reconnection is performed. After iterative expansion, a joint configuration sequence from the initial configuration to the target configuration is obtained. Using the joint configuration sequence as reference data points, a B-spline curve is used to fit the joint configuration sequence to obtain a continuous and differentiable reference path. Using the reference path as a reference trajectory, a control sequence that satisfies joint position limit constraints, acceleration constraints, and jump constraints is solved in the prediction time domain through nonlinear model predictive control. The control sequence is then output to the robot controller for execution, driving the robotic arm to complete the transport of the flexible load.
2. The method according to claim 1, characterized in that, In the step of constructing the latent space dynamic model of the flexible body, the deformation response of the target flexible body under different end motion trajectory excitations is simulated in batches through a physical simulation engine, generating a training dataset containing deformation field, velocity field, external force field and corresponding real deformation field labels. The data fitting loss is constructed based on the deviation between the deformation field estimate obtained by the decoder network reconstruction and the real deformation field labels in the training dataset. The total training loss also includes a latent variable temporal smoothing regularization term, which is used to constrain the change amplitude of latent variables between consecutive time points.
3. The method according to claim 1, characterized in that, After acquiring the current joint state of the robotic arm, the current deformation state of the flexible body, and the obstacle information in the workspace, the method further includes: inputting the current deformation state, velocity state, and external force state into the trained encoder network to obtain the latent variables at the current moment; combining the candidate end-effector motion commands to be evaluated; recursively predicting the latent variable sequences at multiple future moments based on the latent variables at the current moment through the latent variable propagation unit; and inputting the predicted latent variable sequences into the decoder network to obtain the predicted deformation field and strain field values at the corresponding moments.
4. The method according to claim 1, characterized in that, The collision penalty cost of the flexible body dynamic deformation envelope is constructed based on the spatial distance between the flexible body spatial envelope surface obtained by fitting the deformation field prediction value and the geometric set of obstacles in the work space. When the spatial distance is less than a preset safe distance threshold, a positive penalty cost is generated. The material strain energy penalty cost is constructed based on the deviation between the strain field prediction value and the preset safe strain reference threshold.
5. The method according to claim 1, characterized in that, The steps of obtaining candidate configuration points by random sampling in the joint state space of the robotic arm based on the improved fast expanding random tree star topology algorithm include: random sampling in the joint state space of the robotic arm; using a neighboring node search method to determine the existing node in the random tree that is spatially closest to the candidate configuration point as the expansion starting point; generating an expansion path segment along the direction from the expansion starting point to the candidate configuration point with a preset step size; for new nodes that satisfy collision safety constraints and strain safety constraints, comparing the cost of reaching a neighboring node through the new node with the original cost of the neighboring node; and updating the parent node pointing of the neighboring node according to the comparison result.
6. The method according to claim 1, characterized in that, The steps of using nonlinear model predictive control to solve for a control sequence that satisfies constraints in the prediction time domain include: constructing a rolling optimization objective function that includes a joint jump cost term, a joint acceleration cost term, and a reference trajectory tracking error cost term; obtaining the actual joint state of the robotic arm in the current control cycle as the initial state for rolling optimization; solving the rolling optimization objective function using a sequential quadratic programming method to obtain the control sequence in the current prediction time domain; outputting the joint position command, velocity command, and acceleration command corresponding to the current control cycle in the control sequence to the robot controller for execution; and in the next control cycle, re-obtaining the actual joint state of the robotic arm as a new initial state and repeating the rolling optimization process.
7. The method according to claim 1, characterized in that, In scenarios where two robotic arms collaboratively transport flexible loads, the joint states of the two robotic arms are jointly constructed into a joint state space. The external force field input of the flexible body latent space dynamic model simultaneously includes the inertial force contribution corresponding to the motion state of each end effector of the two robotic arms. The composite cost function and the rolling optimization objective function are both solved on the joint state space to constrain the collision cost of the flexible body's dynamic deformation envelope and the material strain energy penalty cost during the collaborative transport of the two robotic arms.
8. An industrial robot motion path planning system, characterized in that, include: The simulation data generation module is used to generate deformation datasets of flexible bodies under different end-effector motion states through a physical simulation engine. The latent space dynamic model training module, connected to the simulation data generation module, is used to receive the deformation dataset, construct and train a flexible body latent space dynamic model including an encoder network and a decoder network. The encoder network is used to map the displacement field, velocity field and external force field of the flexible body into low-dimensional latent variables. The decoder network is used to reconstruct the low-dimensional latent variables into deformation field estimates and strain field estimates. During the training process, physical residual loss is constructed based on the elastic dynamics control equation and the network parameters are iteratively updated. The online sensing module is used to acquire the current joint status of the robotic arm, the current deformation status of the flexible body, and information on obstacles in the workspace; The deformation-sensing path search module is connected to the latent space dynamic model training module and the online sensing module. It is used to sample candidate configuration points based on the improved fast extended random tree star algorithm, call the trained flexible body latent space dynamic model to predict deformation, construct a composite cost function based on the rigid body collision penalty cost of the robotic arm body, the flexible body dynamic deformation envelope collision penalty cost, the material strain energy penalty cost and the path length cost, evaluate the candidate configuration points, and output the joint configuration sequence. The trajectory smoothing module, connected to the deformation sensing path search module, includes a B-spline path fitting unit and a nonlinear model prediction control optimization unit, used to fit and optimize the joint configuration sequence and output control commands that satisfy acceleration constraints and jump constraints. The robot controller execution module, connected to the trajectory smoothing module, is used to receive the control commands and drive the robot arm body to move.
9. The system according to claim 8, characterized in that, The deformation sensing path search module includes a rigid body collision detection unit, a flexible body deformation prediction and invocation unit, a dynamic envelope collision detection unit, a strain energy evaluation unit, and a cost aggregation unit. The cost aggregation unit is used to weight and aggregate the costs output by the rigid body collision detection unit, the dynamic envelope collision detection unit, and the strain energy evaluation unit according to weight coefficients.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the industrial robot motion path planning method according to any one of claims 1 to 7.
