Method for motion planning of rope-towing parallel robot based on guided diffusion model

Abstract

The application provides a rope-towing parallel robot motion planning method based on a guiding diffusion model, and relates to the technical field of rope-towing parallel robot motion planning. By acquiring structural parameters of a rope-towing parallel robot, a task starting state, a task target state and obstacle information, motion planning conditions are formed, an initial trajectory is generated and is subjected to smoothing processing and constraint optimization, and an expert trajectory sample is obtained through collision and tension inspection; the constructed conditional diffusion probability model is trained by using the expert trajectory sample and the motion planning conditions; based on the constructed rope kinematics model, a differentiable hybrid cost function is constructed; under new conditions, an initial trajectory sequence is randomly generated for reverse denoising, the denoising result is subjected to gradient-guided correction, and a candidate trajectory set is generated; a candidate trajectory in the candidate trajectory set is subjected to collision inspection and rope tension inspection to determine a target trajectory, and the rope-towing parallel robot is driven to move according to the target trajectory.

Description

Technical Field

[0001] This invention relates to the field of motion planning technology for rope-driven parallel robots, and more specifically, to a motion planning method for rope-driven parallel robots based on a guided diffusion model. Background Technology

[0002] Cable-Driven Parallel Robots (CDPRs) utilize multiple flexible cables to drive the end-effector platform, offering significant advantages such as a large workspace, lightweight structure, high load-bearing capacity, and convenient reconfiguration. The overall performance of a CDPR is highly dependent on its path planning capability—that is, its ability to autonomously generate a feasible trajectory that satisfies kinematic constraints and avoids collisions. Excellent path planning not only improves operational efficiency and energy utilization but also plays a decisive role in the overall performance of task execution. However, because the cables in a CDPR can only withstand tension, the end-effector platform must simultaneously satisfy multiple constraints during movement, including no end-effector collisions, no cable collisions, and cable tension remaining within boundary limits. This makes its motion planning problem far more complex than that of traditional serial or mobile robots.

[0003] Existing CDPR motion planning methods are mainly divided into two categories: sampling-based methods and optimization-based methods. Sampling-based planning methods include Probabilistic Roadmap (PRM), Rapidly-exploring Random Trees (RRT), and its improved algorithm RRT. Similar to Informed-RRT, this approach can search for initial trajectories from the starting state to the target state in complex obstacle environments, exhibiting probabilistic completeness. However, these methods primarily guarantee geometric connectivity, and the generated trajectories often suffer from insufficient smoothness, excessive rope tension, or interference between the rope and obstacles. Another class of optimization-based planning methods, including Covariant Hamiltonian Optimization for Motion Planning (CHOMP), Trajectory Optimization (TrajOpt), Gaussian Process Motion Planning (GPMP) / Improved Gaussian Process Motion Planning (GPMP2), and Stochastic Trajectory Optimization for Motion Planning (STOMP), can smooth and constrain the initial trajectory. However, the quality of the final solution highly depends on the initial distribution or prior assumptions. If the initial distribution does not approximate the true solution well, the optimization process is prone to getting trapped in local optima, making it difficult to obtain a globally optimal trajectory.

[0004] Conditional diffusion probability models (CDPR) possess excellent distribution modeling capabilities. By progressively generating candidate trajectories through multiple inverse denoising processes, they effectively capture the highly multimodal data distribution characteristics, maintaining a good balance between sample quality and diversity. Furthermore, CDPR can be guided by external likelihood functions, allowing the denoising process to smoothly migrate trajectory samples from the prior distribution to a high-likelihood region, thus achieving efficient and controllable constraint sampling. However, directly applying this model to robot motion planning still faces many challenges, especially for systems like CDPR that involve complex kinematics, dynamics, and multiple collision constraints. Currently, there is no mature solution for constructing an effective conditional guidance mechanism to smoothly migrate trajectory samples from the prior distribution to a high-likelihood region that satisfies all constraints during the denoising process.

[0005] In view of the above, this application is hereby submitted. Summary of the Invention

[0006] This invention provides a rope-driven parallel robot planning method based on a guided diffusion probability model, comprising the following steps:

[0007] S1: Obtain the structural parameters, task start state, task target state, and obstacle information of the rope-driven parallel robot to form motion planning conditions; establish a rope kinematic model based on the rope exit point and end-platform position of the rope-driven parallel robot. Furthermore, the structural parameters include: rope exit point coordinates, end-platform geometry, number of ropes, rope tension boundaries, and workspace range; the task start state and task target state at least include the position and velocity of the end-platform. Furthermore, the rope kinematic model is used to calculate the length, direction vector, and rope tension of each rope at each discrete time node, and to establish the rope structure matrix to determine whether the rope tension at each discrete time node satisfies the tension boundary constraints.

[0008] S2: Based on the motion planning conditions, generate an initial trajectory connecting the task start state and the task target state, and perform smoothing and constraint optimization on the initial trajectory. Perform end-platform collision detection, rope collision detection and rope tension detection on the optimized trajectory, and use the trajectory that passes the detection as an expert trajectory sample. Preferably, the generation of the expert trajectory includes the following steps: Under the aforementioned motion planning conditions, the RRT-connect algorithm is used to generate the initial trajectory; The GPMP2 algorithm is used to smooth and constrain the initial trajectory to obtain the optimized trajectory. The optimized trajectory is subjected to end-platform collision test, rope collision test, and rope tension test. The trajectory that passes the test is used as the expert trajectory sample.

[0009] S3: Construct a conditional diffusion probability model and train the conditional diffusion probability model using the expert trajectory samples and corresponding motion planning conditions, so that it learns the mapping relationship between motion planning conditions and expert trajectories, thereby obtaining a conditional diffusion probability model that can generate trajectories according to new motion planning conditions. Preferably, training the conditional diffusion probability model specifically includes: Using expert trajectory samples and motion planning conditions as training data, U-Net is used as the denoising network for the conditional diffusion probability model. During training, Gaussian noise is added to the expert trajectory samples according to a preset diffusion time step to obtain a noisy trajectory; the noisy trajectory, the corresponding diffusion time step, and the motion planning conditions are then input into the U-Net denoising network. A loss function is constructed based on the prediction error, and the parameters of the U-Net denoising network are updated according to the loss function until the training of the conditional diffusion probability model is completed.

[0010] S4: Based on the aforementioned rope kinematics model, construct a differentiable hybrid cost function for gradient guidance in the inverse denoising process; Preferably, the differentiable hybrid cost function includes end-platform obstacle avoidance cost, rope obstacle avoidance cost, and rope tension boundary cost; Furthermore, the differentiable hybrid cost function includes tension robustness cost and energy consumption cost; The differentiable hybrid cost function is used to comprehensively evaluate the collision safety, rope tension feasibility, system robustness, and motion energy consumption of the trajectory, and provides a differentiable gradient during the inverse denoising process to guide the trajectory to continuously optimize in the direction of reducing the overall cost value.

[0011] S5: Under the new motion planning conditions, the randomly generated initial trajectory sequence is used as the input for the inverse denoising process. In each inverse denoising time step, the trained conditional diffusion probability model is used for inverse denoising, and the gradient of the current trajectory state is calculated according to the differentiable hybrid cost function. The denoising result is then corrected by gradient guidance to generate candidate trajectories. The inverse denoising and gradient guidance correction process is repeated for different randomly generated initial trajectory sequences to obtain a candidate trajectory set. Preferably, the gradient-guided correction includes: calculating the gradient corresponding to the current trajectory state based on the differentiable hybrid cost function, and using the gradient to correct the denoising result obtained in the current inverse denoising time step, so as to reduce the comprehensive cost value corresponding to the corrected trajectory.

[0012] S6: Perform collision detection and rope tension detection on the candidate trajectories in the candidate trajectory set, and determine the target trajectory based on the detection results, and drive the rope-traction parallel robot to move according to the target trajectory.

[0013] Preferably, the determination of the target trajectory includes: for candidate trajectories that pass the collision test and rope tension test, calculating the comprehensive cost value of each candidate trajectory according to the differentiable hybrid cost function, and sorting them according to the comprehensive cost value. Among the candidate trajectories that meet the constraints, the candidate trajectory with the smallest comprehensive cost value is determined as the target trajectory; wherein, meeting the constraints means that the candidate trajectory meets the end-platform collision constraint, rope collision constraint, and rope tension constraint.

[0014] By adopting the above technical solution, the present invention can achieve the following technical effects: Generating expert trajectory samples that satisfy motion constraints improves the quality of training data and provides a reliable foundation for training the conditional diffusion probability model. A conditional diffusion probability model incorporating motion planning conditions is constructed, enabling the model to quickly generate trajectories that meet planning conditions based on different task requirements, thus improving trajectory planning efficiency. A differentiable hybrid cost function is constructed based on the rope kinematics model and introduced into the inverse denoising process of the conditional diffusion probability model. Gradient guidance is used to correct the trajectory generation process in real time, ensuring that the generated trajectory simultaneously satisfies end-platform collision constraints, rope collision constraints, and rope tension constraints, improving trajectory planning success rate and trajectory quality. By generating multiple candidate trajectories and filtering them based on comprehensive cost value, a target trajectory with superior overall performance can be obtained, improving the motion planning ability, tension robustness, and operational stability of the rope-traction parallel robot in complex obstacle environments. Using a random initial trajectory sequence as the denoising starting point, it does not rely on prior initial solutions, avoiding local optima. Multiple candidate trajectories can be generated simultaneously, adapting to complex scenarios with multiple feasible paths. Compared to optimization methods that rely solely on a single initial solution, this invention possesses stronger global exploration capabilities, improving the success rate of feasible trajectory generation in complex obstacle environments and its generalization ability to different task scenarios. Attached Figure Description

[0015] To more clearly illustrate the technical solution of the present invention, the accompanying drawings used in the specific embodiments of the present invention will be briefly introduced below. It should be understood that the following drawings only show some specific embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0016] Figure 1 This is a flowchart illustrating the planning method for rope-driven parallel robots based on a guided diffusion probability model.

[0017] Figure 2 This is a structural schematic diagram of a rope-driven parallel robot.

[0018] Figure 3 A schematic diagram illustrating the process of generating a feasible reference trajectory for the end-effector platform of a rope-driven parallel robot.

[0019] Figure 4 This is a comparison diagram of the rope tension distribution of trajectories generated by different planning methods in Example 2; Figure 5 This is a comparison chart of the distribution of condition number of the structure matrix in the XY plane and the trajectory generated by different planning methods in Example 2; Detailed Implementation The technical solutions of the present invention will now be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention.

[0020] Example 1 See Figure 1 The first embodiment of the present invention provides a rope-traction parallel robot planning method based on a guided diffusion probability model, comprising the following steps: S1: Obtain the structural parameters, task start state, task target state, and obstacle information of the rope-driven parallel robot to form motion planning conditions. ; Furthermore, the structural parameters include: rope exit point coordinates, end-platform geometry, number of ropes, rope tension boundaries, and workspace range; the task start state and task target state at least include the position and velocity of the end-platform. See Figure 2 A rope kinematic model is established based on the rope exit point and end-platform position of the rope-traction parallel robot: For the Let there be discrete time points, and the position of the end platform be denoted as . , No. The location of the rope exit point is as follows: Then the first A rope can be approximated as a connection and The line segment, the first The length of the rope and direction vector They are respectively: ; Establish the rope structure matrix based on the direction vectors of each rope. : ; The gravity, mission load, and external disturbances experienced by the end platform are combined into an external load. ; Furthermore, the gravity refers to the load generated by the weight of the end platform itself; the task load refers to the additional load generated by the end platform during the process of transporting workpieces or performing operations; and the external disturbance force refers to inertial disturbance or other environmental forces.

[0021] At each discrete time node, according to the rope structure matrix and external loads Solve for the tension of each rope. And determine the tension corresponding to each rope. Does it meet the preset tension safety boundary? ; In the formula, The minimum allowable tension of the rope; The maximum allowable tension of the rope; For example, the minimum tension The maximum tension It is understood that the specific value of the tension safety boundary can be adjusted according to the actual rope specifications, robot size and load requirements. The above method is only a specific example, and the implementation of the present invention is not limited thereto.

[0022] If the tension of all ropes at the current discrete time node satisfies the above-mentioned tension safety boundary constraint, then the current discrete time node is determined to be within the tension safety boundary; if the tension of any rope at the current discrete time node exceeds the preset tension safety boundary, then the discrete time node is determined not to satisfy the tension safety boundary constraint.

[0023] S2: Based on the motion planning conditions The system generates an initial trajectory that connects the task start state and the task target state, and performs smoothing and constraint optimization on the initial trajectory. The optimized trajectory is then subjected to end-platform collision detection, rope collision detection, and rope tension detection. The trajectories that pass the tests are used as expert trajectory samples. Preferably, the generation of the expert trajectory includes the following steps: In the motion planning conditions The initial trajectory is then generated using the RRT-connect algorithm. The GPMP2 algorithm is used to smooth and constrain the initial trajectory to obtain the optimized trajectory. The optimized trajectory is subjected to end-platform collision test, rope collision test, and rope tension test. The trajectory that passes the test is used as the expert trajectory sample. The expert trajectory sample A sequence representation consisting of states at several discrete time points: ; In the formula, For the first The trajectory contains a number of discrete time nodes, where H represents the total number of discrete time nodes. Furthermore, the trajectory state at the j-th discrete time node is represented as: ; In the formula, For the location of the end platform, For the end-platform speed; The RRT-connect algorithm can quickly generate feasible paths connecting the initial state and the target state. The GPMP2 algorithm is used to further improve the smoothness and continuity of the trajectory and make the trajectory meet the motion constraints, thereby obtaining high-quality expert trajectory samples and providing a data foundation for training the conditional diffusion probability model.

[0024] S3: Construct a conditional diffusion probability model and utilize the expert trajectory samples. and corresponding motion planning conditions The conditional diffusion probability model is trained to learn the mapping relationship between motion planning conditions and expert trajectories, thereby obtaining a conditional diffusion probability model that can generate trajectories based on new motion planning conditions. Preferably, training the conditional diffusion probability model specifically includes: Using expert trajectory samples and sports planning conditions U-Net is used as the denoising network for the conditional diffusion probability model for training data; During training, Gaussian noise is added to the expert trajectory samples according to a preset diffusion time step, so that the original expert trajectory is gradually transformed into a noisy trajectory. Noisy trajectory corresponding to each diffusion time step for; ; In the formula, This is the cumulative noise scheduling coefficient during the preset diffusion process; This represents random noise that follows a standard Gaussian distribution N(0,I); N(0,I) represents a Gaussian distribution with a mean of 0 and a covariance matrix of identity.

[0025] The noisy trajectory , corresponding diffusion time step and sports planning conditions The input is a U-Net denoising network, which extracts the noisy trajectory. The characteristics, combined with the motion planning conditions The mapping relationship between the learning trajectory and the motion planning conditions is used to predict the noise information corresponding to the current diffusion time step, and the predicted noise is output. ; In the formula, These are the parameters for the U-Net denoising network.

[0026] Construct a loss function based on prediction error. : ; The loss function is used to measure the predicted noise of the U-Net denoising network. With actual added random noise The prediction error between them.

[0027] The parameters of the U-Net denoising network are continuously updated based on the loss function. This allows the predicted noise to gradually approach the actual random noise added, until the loss function converges or reaches the preset number of training rounds, thereby completing the training of the conditional diffusion probability model and obtaining the trained conditional diffusion probability model.

[0028] S4: Based on the aforementioned rope kinematics model, construct a differentiable hybrid cost function for gradient guidance in the inverse denoising process; Preferably, the differentiable hybrid cost function includes end-platform obstacle avoidance cost, rope obstacle avoidance cost, and rope tension boundary cost; Furthermore, the differentiable hybrid cost function also includes tension robustness cost and energy consumption cost; The expression for the differentiable hybrid cost function is: ; In the formula, Representing the trajectory Under the conditions of sports planning The comprehensive value of the following; The obstacle avoidance cost of the end platform at the j-th discrete time node; The cost of rope obstacle avoidance at the j-th discrete time node; The boundary cost of the rope tension at the j-th discrete time node; The stress robustness cost at the j-th discrete time node; The energy consumption cost at the j-th discrete time node; , , , and These represent the weight coefficients of the corresponding cost terms, used to adjust the degree of influence of each evaluation index on the total cost function.

[0029] For non-smooth terms caused by distance, tension, or condition number thresholds, this embodiment uses a smoothing penalty function or a differentiable approximation function to process them, so that the entire cost function remains differentiable, thereby enabling optimization using gradient calculation methods.

[0030] Furthermore, the obstacle avoidance cost of the end-point platform is used to constrain the end-point platform to maintain a safe distance from obstacles, and its expression is: ; In the formula, For the first An obstacle, For the end platform in the The spatial occupied area corresponding to each discrete time node. For the smoothing penalty function, For the first The pose of the end platform at each discrete time point. This is the minimum distance between the end platform and the obstacle. Preset a safe distance threshold for the end platform; Furthermore, the rope obstacle avoidance cost is used to constrain each rope to maintain a safe distance from the obstacle, and its expression is: ; In the formula, Let i be the spatial line segment corresponding to the i-th rope at the j-th discrete time node; This indicates the minimum distance between the rope and the obstacle; The rope safety distance threshold; Furthermore, the rope tension boundary cost is used to ensure that each rope maintains a safe stress state during robot operation, and its expression is: ; Furthermore, the tension robustness cost is used to improve the robot's stress stability under different configurations, and its expression is: ; In the formula, The condition number of the structure matrix. and The structure matrix is ​​respectively Maximum singular value and minimum singular value, To prevent small positive quantities with a denominator of zero, The preset threshold number of conditions; When the condition number of the structure matrix is ​​large, it indicates that the robot configuration is close to a singular state, and the corresponding cost increases. This helps guide the trajectory to avoid configurations with poor tension distribution performance and improves the tension robustness of the trajectory. Furthermore, the energy consumption cost is used to characterize the load level of the robot's rope drive system at the corresponding discrete time node, and can be determined according to the rope tension magnitude. The tension vector composed of the rope tensions at the j-th discrete time node can be expressed as: ; Preferably, the energy consumption cost can be calculated based on the L2 norm of the tension vector, that is, the greater the tension, the higher the corresponding driving load, and the greater the energy consumption cost; conversely, the lower the tension, the smaller the energy consumption cost. In summary, the differentiable hybrid cost function is used to comprehensively evaluate the collision safety, rope tension feasibility, system robustness, and motion energy consumption of the trajectory, and provides a differentiable gradient during the inverse denoising process to guide the trajectory to continuously optimize in the direction of reducing the overall cost value.

[0031] S5: Under the new motion planning conditions, candidate trajectories that meet the motion planning requirements are generated using the trained conditional diffusion probability model, specifically: Obtain new motion planning conditions c, using the randomly generated initial trajectory sequence as input for the inverse denoising process; Furthermore, the initial trajectory sequence can be composed of random Gaussian noise that satisfies a preset trajectory dimension; In each inverse denoising time step, the trained conditional diffusion probability model is used to perform inverse denoising. Starting from the preset maximum diffusion time step, it iterates step by step towards time zero. In each inverse denoising time step, the current trajectory is denoised to obtain the trajectory denoising result corresponding to the current time step. And calculate the gradient of the current trajectory state based on the differentiable hybrid cost function: ; In the formula, This indicates the gradient guidance direction corresponding to the current inverse denoising time step. The differentiable hybrid cost function with respect to the current trajectory state The gradient; Based on the gradient, the denoising result is corrected by gradient guidance to obtain the trajectory after correction at the current inverse denoising time step; ; In the formula, The denoising result of the current trajectory output by the U-Net denoising network. To preset the guidance intensity, The result of trajectory denoising after gradient guidance; Furthermore, the preset guiding intensity It can be set to a fixed value or dynamically changed with the time step of the reverse denoising process. For example, a larger gradient guiding weight is used in the early stage of reverse denoising to quickly guide the trajectory to move towards the region that meets the constraints. As the reverse denoising process gradually approaches the end, the gradient guiding weight is gradually reduced to ensure the stability of the trajectory generation process. The corrected trajectory is used as the input for the next inverse denoising time step, and the inverse denoising and gradient-guided correction process is continued until all inverse denoising time steps are completed, generating candidate trajectories. Repeatedly perform the inverse denoising and gradient-guided correction process on different randomly generated initial trajectory sequences to obtain a candidate trajectory set. ; Preferably, the gradient-guided correction includes: calculating the gradient corresponding to the current trajectory state based on the differentiable hybrid cost function, and using the gradient to correct the denoising result obtained in the current inverse denoising time step, so as to reduce the comprehensive cost value corresponding to the corrected trajectory.

[0032] Furthermore, if no candidate trajectory satisfies the constraints, a new random initial trajectory is generated and the planning continues. Furthermore, a maximum number of retries can be set, for example, 20 times. If no feasible trajectory is found after multiple retries exceeding the maximum number of retries, a planning failure alarm will be output.

[0033] S6: For the candidate trajectory set Each candidate trajectory in the test is subjected to collision testing and rope tension testing. Furthermore, the collision test and rope tension test specifically refer to: For each discrete time node in the trajectory, the end-platform collision test, rope collision test, and rope tension boundary test are performed sequentially. The end-platform collision test is used to determine whether the end platform collides with environmental obstacles during the entire motion process; the rope collision test is used to determine whether each rope collides with obstacles during the motion process; and the rope tension boundary test is used to determine whether the tension of each rope at each discrete time node meets the preset tension boundary. A candidate trajectory is considered feasible only if it satisfies the above constraints at all discrete time points. For feasible trajectories that pass the collision test and rope tension test, the comprehensive cost value of each candidate trajectory is calculated according to the differentiable hybrid cost function, and the trajectories are ranked according to the comprehensive cost value. The feasible trajectory with the smallest comprehensive cost value is determined as the target trajectory. : ; In the formula, As a feasibility indicator function, when a candidate trajectory passes the collision test and the rope tension test, ,otherwise ; For example, the process of generating the target trajectory of a rope-driven parallel robot is as follows: Figure 3 As shown in the figure, j represents the inverse denoising time step in the inverse denoising process of the conditional diffusion probability model. At the beginning of inverse sampling (corresponding to a larger inverse denoising time step, j=20), the trajectory is initialized with random Gaussian noise, and the trajectory points are randomly distributed and have no obvious correlation with the target trajectory.

[0034] As the reverse denoising process continues (j=15 and j=10), the conditional diffusion probability model gradually removes noise from the trajectory under the constraints of motion planning conditions. Simultaneously, it uses the gradient information provided by the differentiable hybrid cost function to guide and correct the denoising results, causing the trajectory to gradually converge towards the region that meets the motion planning requirements. At this point, the trajectory begins to move away from the obstacle and gradually forms a continuous motion path from the initial state of the task to the target state.

[0035] As the reverse denoising time step is further reduced (j=5), the randomness of the trajectory is further reduced, the trajectory shape tends to be stable, the obstacle avoidance constraints of the end platform, the obstacle avoidance constraints of the rope, and the rope tension constraints are gradually satisfied, and the smoothness of the trajectory is further improved.

[0036] When the reverse denoising process ends (j=0), a candidate trajectory set is obtained. The candidate trajectories in the candidate trajectory set connect the task's initial state and the task's target state, avoiding environmental obstacles. The candidate trajectories undergo collision testing and rope tension testing, and their comprehensive cost value is calculated. The trajectory with the minimum comprehensive cost value is the target trajectory for the rope-driven parallel robot to execute motion planning.

[0037] Based on the target trajectory, the rope-driven parallel robot is driven to move smoothly along the target trajectory.

[0038] Compared with the prior art, the present invention has the following beneficial effects: This invention combines data-driven planning and model constraint capabilities, fully considering the unique rope constraint characteristics of rope-driven parallel robots. It avoids the pitfalls of traditional trajectory planning methods that only plan the end-effector's motion while ignoring collision risks and tension instability caused by rope state changes, thus improving the executability and safety of the planned trajectory. It can adaptively adjust based on environmental obstacles, task objectives, and rope constraints. Compared to traditional methods based on random search or sampling, it reduces invalid trajectory searches and improves trajectory generation efficiency and planning success rate under complex constraints. By performing constraint checks and comprehensive evaluations on multiple candidate trajectories, it selects the target trajectory that satisfies the constraints and has the optimal overall cost, improving the robustness of robot motion planning and achieving efficient and safe motion planning for rope-driven parallel robots in complex environments.

[0039] Example 2 Referring to Embodiment 1, the second embodiment of the present invention employs RRT. The bidirectional RRT connectivity algorithm + GPMP2 trajectory optimization library (Rapidly-exploring Random Trees Connect+Gaussian Process Motion Planner2, RRTC+GPMP2), Conditional Variational Autoencoder (CVAE), Denoising Diffusion Probabilistic Model (DDPM), and the method of this invention (GDPM) were used to plan the trajectory of a rope-driven parallel robot. The effectiveness of the method of this invention was verified by comparing the results.

[0040] See Figure 4 The rope tension distribution at all discrete time points in the trajectories generated by each method was statistically analyzed. It can be seen that RRT... While most of the tension in the trajectories generated by planning methods such as RRTC+GPMP2, CVAE, and DDPM is within the allowable range, some trajectories still have rope tension close to the tension boundary. This indicates that some planning results are near the tension safety boundary and are easily affected by model errors, control errors, or external disturbances, thereby reducing the stability and robustness of the trajectory execution process. In contrast, the rope tension corresponding to the trajectory generated by the method of this invention is concentrated in the middle region of the preset tension boundary, with a large safety margin from both the minimum and maximum tension boundaries. The tension distribution is more concentrated and uniform, and extreme tension values ​​are significantly reduced.

[0041] See Figure 5 The planned trajectories are projected onto the XY plane for comparison. The condition number of the structure matrix corresponding to different positions of the rope-driven parallel robot is used as the background for visualization, where color represents the magnitude of the condition number; the more red the color, the larger the condition number, and the more blue the color, the smaller the condition number. It can be seen that RRT... The planned trajectory deviates from the safe zone overall, passing through areas with high condition numbers during movement, indicating a tendency for the robot to approach a singular configuration; although the trajectories generated by RRTC+GPMP2 and CVAE are better than those generated by RRT... While there has been some improvement, some trajectories still pass through regions with high condition numbers. Trajectories generated by DDPM can avoid high condition number regions to some extent, but lack guidance from robot kinematic constraints, and their trajectories do not fully consider tension distribution performance. In contrast, the trajectory generated by the method of this invention (GDPM) is located in regions with low condition numbers, the trajectory is smoother, and while ensuring obstacle avoidance performance, it keeps the robot away from strange configurations.

[0042] Example 3 The third embodiment of the present invention provides a motion planning system for a rope-driven parallel robot based on a guided diffusion model, comprising: The condition acquisition module is used to acquire the structural parameters, task start state, task target state and obstacle information of the rope-pulled parallel robot, form motion planning conditions, and establish a rope kinematic model based on the rope exit point and end platform position of the rope-pulled parallel robot. The trajectory sample generation module is used to generate an initial trajectory connecting the task start state and the task target state based on the motion planning conditions, and to perform smoothing and constraint optimization on the initial trajectory. The optimized trajectory is then subjected to end-platform collision detection, rope collision detection, and rope tension detection. The trajectory that passes the detection is used as an expert trajectory sample. The model training module is used to construct a conditional diffusion probability model. The expert trajectory samples and the motion planning conditions are used as training data to train the conditional diffusion probability model and obtain the trained conditional diffusion probability model. The cost function construction module is used to construct a differentiable hybrid cost function for gradient guidance in the inverse denoising process based on the rope kinematics model. The candidate trajectory generation module is used to take a randomly generated initial trajectory sequence as input to the inverse denoising process under new motion planning conditions. At each inverse denoising time step, it performs inverse denoising using the trained conditional diffusion probability model, calculates the gradient of the current trajectory state according to the differentiable hybrid cost function, and performs gradient-guided correction on the denoising result to generate candidate trajectories. The inverse denoising and gradient-guided correction process is repeated for different randomly generated initial trajectory sequences to obtain a candidate trajectory set. The target trajectory determination module is used to perform collision detection and rope tension detection on the candidate trajectories in the candidate trajectory set, and determine the target trajectory based on the detection results, thereby driving the rope-traction parallel robot to move according to the target trajectory.

[0043] Example 4 The fourth embodiment of the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method described in the first embodiment.

[0044] Example 5 The fifth embodiment of the present invention provides a computer program product, including a computer program that, when run on a processor, causes the processor to perform the method described in the first embodiment.

[0045] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A planning method for rope-driven parallel robots based on a guided diffusion probability model, characterized in that, Includes the following steps: S1: Obtain the structural parameters, task start state, task target state, and obstacle information of the rope-driven parallel robot to form motion planning conditions; A rope kinematic model is established based on the rope exit point and end-platform position of the rope-traction parallel robot. S2: Based on the motion planning conditions, generate an initial trajectory connecting the task start state and the task target state, and perform smoothing and constraint optimization on the initial trajectory. Perform end-platform collision detection, rope collision detection and rope tension detection on the optimized trajectory, and use the trajectory that passes the detection as an expert trajectory sample. S3: Construct a conditional diffusion probability model, using the expert trajectory samples and the motion planning conditions as training data to train the conditional diffusion probability model and obtain the trained conditional diffusion probability model. S4: Based on the aforementioned rope kinematics model, construct a differentiable hybrid cost function for gradient guidance in the inverse denoising process; S5: Under the new motion planning conditions, the randomly generated initial trajectory sequence is used as the input for the inverse denoising process. In each inverse denoising time step, the trained conditional diffusion probability model is used for inverse denoising, and the gradient of the current trajectory state is calculated according to the differentiable hybrid cost function. The denoising result is then corrected by gradient guidance to generate candidate trajectories. The inverse denoising and gradient guidance correction process is repeated for different randomly generated initial trajectory sequences to obtain a candidate trajectory set. S6: Perform collision detection and rope tension detection on the candidate trajectories in the candidate trajectory set, and determine the target trajectory based on the detection results, and drive the rope-traction parallel robot to move according to the target trajectory.

2. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 1, characterized in that, The structural parameters in step S1 include: rope exit point coordinates, end platform geometry, number of ropes, rope tension boundary, and workspace range.

3. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 1, characterized in that, The task start state and task target state in step S1 include at least the position and speed of the end platform.

4. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 1, characterized in that, The rope kinematic model in step S1 is used to calculate the length, direction vector, and rope tension of each rope at each discrete time node, and to establish a rope structure matrix to determine whether the rope tension corresponding to each discrete time node satisfies the tension boundary constraint.

5. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 1, characterized in that, The generation of the expert trajectory in step S2 includes the following steps: Under the aforementioned motion planning conditions, the RRT-connect algorithm is used to generate the initial trajectory; The GPMP2 algorithm is used to smooth and constrain the initial trajectory to obtain the optimized trajectory. The optimized trajectory is subjected to end-platform collision test, rope collision test, and rope tension test. The trajectory that passes the test is used as the expert trajectory sample.

6. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 1, characterized in that, The training of the conditional diffusion probability model in step S3 specifically includes: Using expert trajectory samples and motion planning conditions as training data, U-Net is used as the denoising network for the conditional diffusion probability model. During training, Gaussian noise is added to the expert trajectory samples according to a preset diffusion time step to obtain a noisy trajectory; the noisy trajectory, the corresponding diffusion time step, and the motion planning conditions are then input into the U-Net denoising network. A loss function is constructed based on the prediction error, and the parameters of the U-Net denoising network are updated according to the loss function until the conditional diffusion probability model training is completed.

7. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 1, characterized in that, The differentiable hybrid cost function in step S4 includes the end-platform obstacle avoidance cost, the rope obstacle avoidance cost, and the rope tension boundary cost.

8. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 7, characterized in that, The differentiable hybrid cost function also includes tension robustness cost and energy consumption cost.

9. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 8, characterized in that, The gradient-guided correction in step S5 includes: calculating the gradient corresponding to the current trajectory state based on the differentiable hybrid cost function, and using the gradient to correct the denoising result obtained in the current inverse denoising time step, so that the value of the differentiable hybrid cost function corresponding to the corrected trajectory is reduced.

10. The rope-traction parallel robot planning method based on a guided diffusion probability model according to claim 8, characterized in that, The determination of the target trajectory in step S6 includes: calculating the comprehensive cost value of each candidate trajectory based on the differentiable hybrid cost function for the candidate trajectories that have passed the collision test and rope tension test, sorting them according to the comprehensive cost value, and determining the candidate trajectory with the smallest comprehensive cost value as the target trajectory.

Images