Robot complex curved surface generation type constant force polishing method based on diffusion model

Abstract

The present application belongs to the technical field of intelligent polishing of industrial robots, and discloses a robot complex curved surface generated constant force polishing method based on a diffusion model. The method comprises the following steps: collecting multi-modal data to construct a training data set; modeling the polishing action sequence as a conditional reverse diffusion process, taking the current multi-modal observation as the input, training the noise prediction network to fit the expert operation distribution, and obtaining the conditional diffusion strategy model; during execution, the model receives the topography of the visual image of the area to be polished, the six-dimensional force, the end pose and the joint state in a rolling horizon manner, generates the future action sequence through denoising iteration, and inputs the impedance controller to convert into joint torque instructions; at the same time, a safety watchdog is set, the deviation between the expected force and the measured force is compared, and when the risk of overcutting or hard collision is determined, the control is immediately taken over and the rollback is executed. The above-mentioned method does not require an accurate geometric model, has strong adaptability, has small force control overshoot, and significantly improves the polishing quality of the robot.

Description

Technical Field

[0001] This invention relates to the field of intelligent grinding technology for industrial robots, and in particular to a constant force grinding method for complex curved surfaces generated by robots based on a diffusion model. Background Technology

[0002] Complex curved surface components such as water turbines and aero engines are prone to damage such as cavitation pits and cracks during service. Repaired weld areas often exhibit randomly distributed weld beads, pits, and abrupt hardness changes. Traditional offline programming methods based on CAD models fail due to the irregularity of the curved surface. Existing robotic grinding systems mostly employ PID force-position hybrid control or impedance control, but when faced with sudden changes in contact stiffness (such as the grinding head impacting a hard weld point), they often experience force oscillations, overcutting, or even tool breakage. Furthermore, the implicit knowledge of highly skilled technicians adjusting grinding angles and pressure through "feel" is difficult to quantify into traditional control code, resulting in robotic grinding quality relying on manual teaching, leading to low efficiency and poor consistency.

[0003] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0004] The main objective of this invention is to provide a robot-based method for generating constant force grinding on complex curved surfaces based on a diffusion model. This method aims to solve the technical problems of current robots being unable to achieve adaptive constant force grinding on unknown irregular curved surfaces, and the grinding quality relying on manual teaching, resulting in low efficiency and poor consistency.

[0005] To achieve the above objectives, this invention provides a robot-based method for generating constant force grinding of complex surfaces using a diffusion model. This method includes the following steps: Multimodal data of the polishing process is collected, and the multimodal data is preprocessed by time alignment and voxelization to construct a multimodal training dataset. The multimodal data includes RGB-D visual images, six-dimensional contact force data, robot end pose data, and joint velocity data. The generation process of the grinding action sequence is modeled as a conditional backdiffusion process. The multimodal observation vector at the current moment is used as the input condition. The noise-based prediction neural network is trained using the multimodal training dataset. The action distribution law of the expert operation is fitted by iterative optimization. After training, the conditional diffusion strategy model is obtained. The conditional diffusion strategy model supports the output of a grinding action sequence that conforms to the distribution of expert operation after multiple denoising iterations from Gaussian noise input. The grinding action sequence includes the end pose adjustment amount and normal contact force setting instructions for several future steps. The conditional diffusion strategy model is loaded, and the RGB-D topography data, six-dimensional contact force data, robot end pose and joint velocity data of the current area to be polished are received in real time through the conditional diffusion strategy model. These data are fused into a multimodal observation vector as the model input. Based on the input observation vector, a polishing action sequence of a future preset duration is generated through rolling temporal denoising iteration. The polishing action sequence is then input into an impedance controller and converted into joint torque commands to drive the robotic arm to perform the polishing operation. During the polishing process, the conditional diffusion strategy model dynamically adjusts the polishing parameters according to the observation data updated once at a preset frequency. The system synchronously acquires the expected normal contact force command output by the conditional diffusion strategy model and the six-dimensional contact force data collected during the grinding process. Based on the preset safety judgment rules, the system compares the deviation characteristics of the two sets of data. If it determines that there is a risk of overcutting or hard collision, it immediately takes over the control of the robot and executes deceleration, safe retreat, or emergency stop actions. The abnormal working condition data is synchronously transmitted back to the training library of the conditional diffusion strategy model for subsequent incremental optimization of the model.

[0006] In one embodiment, the denoising iterative process of the conditional diffusion strategy model is implemented by the following formula:

[0007] in, For the first The noise-enhanced action sequence tensor of the diffusion time step. This is the current multimodal observation vector, which incorporates RGB-D visual features. With historical sense sequence This is a Transformer-based neural network for noise prediction, used to estimate and remove noise. These are noise scheduling parameters that control the denoising step size. Injected Gaussian random noise is used to enhance the diversity and perturbation resistance of the strategy. The standard deviation parameter of the Gaussian noise scheduling corresponding to the k-th denoising iteration is denoised as follows.

[0008] In one embodiment, the conditional diffusion strategy model continuously replans the action sequence for the next 2 seconds at a frequency of not less than 50Hz. When the normal contact force mutation amplitude is detected to exceed 15N, the conditional diffusion strategy model automatically reduces the trajectory feed speed and improves the smoothness of the action sequence in the output action sequence.

[0009] In one embodiment, the impedance controller converts the grinding action sequence output by the conditional diffusion strategy model into joint torque commands using the following formula:

[0010] Where τcmd is the final control torque command for each joint of the robot output by the impedance controller. Let Λ(q) be the transpose of the Jacobian matrix corresponding to the current joint angle q of the robot, where q is the actual angle vector of each joint of the robot at the current time, which is collected in real time by the robot joint encoder, and Λ(q) is the operational space inertial matrix under the current joint pose. d represents the desired end-effector acceleration vector in the grinding action sequence output by the conditional diffusion strategy model, and B represents the desired end-effector acceleration vector. d Let be the desired damping matrix output in real time by the conditional diffusion strategy model, and d be the desired end velocity vector in the grinding action sequence output by the conditional diffusion strategy model. K represents the actual velocity vector of the robot's end effector. d X is the desired stiffness matrix output in real time by the conditional diffusion strategy model. d Let X be the desired end-effector pose vector in the grinding action sequence output by the conditional diffusion strategy model, and F be the actual pose vector of the robot's end-effector. d The normal contact force feedforward term generated for the conditional diffusion strategy model, g(q) is the gravity compensation torque vector under the current joint pose, and C( , ) represents the Coriolis force and centrifugal force compensation torque vector under the current joint pose and velocity.

[0011] In one embodiment, when the conditional diffusion strategy model identifies the area to be ground as a high-hardness weld bead through the input RGB-D visual features, it automatically adjusts the feed speed to 5 mm / s and the normal contact force setting value to 25 N in the output action sequence. If the normal contact force change rate is detected to exceed 50 N / s during the grinding process, the impedance controller instantly reduces the stiffness parameter.

[0012] In one embodiment, security is determined using the following security scoring function:

[0013] in, For indicator functions, For the rate of change of force, The measured normal contact force value is collected in real time by a six-dimensional force sensor during the grinding process. This represents the desired normal contact force command value output in real time by the conditional diffusion strategy model. This is the rated maximum normal contact force of the grinding system. For safety attenuation coefficient, This is the allowable deviation threshold for normal contact force. It is a natural exponential function.

[0014] Furthermore, to achieve the above objectives, this invention also proposes a robot complex surface generation constant force grinding device based on a diffusion model. This device is applied to the robot complex surface generation constant force grinding method based on a diffusion model described above. The device includes: The data acquisition module is used to collect multimodal data during the polishing process. After time alignment and voxelization preprocessing, the multimodal data is used to construct a multimodal training dataset. The multimodal data includes RGB-D visual images, six-dimensional contact force data, robot end-effector pose data, and joint velocity data. The model training module is used to model the generation process of the grinding action sequence as a conditional backdiffusion process. It uses the multimodal observation vector at the current moment as the input condition, and trains a noise-based prediction neural network using the multimodal training dataset. Through iterative optimization, it fits the action distribution law of expert operations. After training, a conditional diffusion strategy model is obtained. The conditional diffusion strategy model supports outputting a grinding action sequence that conforms to the distribution of expert operations after multiple denoising iterations from Gaussian noise input. The grinding action sequence includes end pose adjustment amount and normal contact force setting instructions for several future steps. The model processing module is used to load the conditional diffusion strategy model. The conditional diffusion strategy model receives RGB-D topography data, six-dimensional contact force data, robot end pose and joint velocity data of the current area to be polished in real time, and fuses them into a multimodal observation vector as model input. Based on the input observation vector, a polishing action sequence of a future preset duration is generated through rolling temporal denoising iteration. The polishing action sequence is input to the impedance controller and converted into joint torque commands to drive the robotic arm to perform the polishing operation. During the polishing process, the conditional diffusion strategy model dynamically adjusts the polishing parameters according to the observation data updated once at a preset frequency. The anomaly detection module is used to synchronously acquire the expected normal contact force command output by the conditional diffusion strategy model and the six-dimensional contact force data collected during the grinding process in real time. Based on the preset safety judgment rules, it compares the deviation characteristics of the two sets of data. If it determines that there is a risk of overcutting or hard collision, it immediately takes over the control of the robot and executes deceleration, safe retreat or emergency stop actions. The abnormal working condition data is synchronously transmitted back to the training library of the conditional diffusion strategy model for subsequent incremental optimization of the model.

[0015] In one embodiment, the conditional diffusion strategy model continuously replans the action sequence for the next 2 seconds at a frequency of not less than 50Hz. When the normal contact force mutation amplitude is detected to exceed 15N, the conditional diffusion strategy model automatically reduces the trajectory feed speed and improves the smoothness of the action sequence in the output action sequence.

[0016] Furthermore, to achieve the above objectives, the present invention also proposes a robot complex surface generation constant force grinding device based on a diffusion model. The robot complex surface generation constant force grinding device based on a diffusion model includes: a memory, a processor, and a robot complex surface generation constant force grinding program based on a diffusion model stored in the memory and executable on the processor. The robot complex surface generation constant force grinding program based on a diffusion model is configured to implement the steps of the robot complex surface generation constant force grinding method based on a diffusion model as described above.

[0017] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a robot complex surface generation constant force grinding program based on a diffusion model. When the robot complex surface generation constant force grinding program based on a diffusion model is executed by a processor, it implements the steps of the robot complex surface generation constant force grinding method based on a diffusion model as described above.

[0018] In this invention, multimodal data is collected to construct a training dataset. The generation of the grinding action sequence is modeled as a conditional backdiffusion process. Using the current multimodal observations as input, a noise prediction network is trained to fit the expert operation distribution, resulting in a conditional diffusion strategy model. During execution, the model receives the topography, six-dimensional forces, end-effector pose, and joint states of the visual image of the area to be ground in real time using a rolling time-domain approach. After denoising and iterative generation, the future action sequence is generated and input into an impedance controller to be converted into joint torque commands. Simultaneously, a safety watchdog is set up to compare the deviation between the expected force and the measured force. If an overcut or hard collision risk is detected, control is immediately taken over and a rollback is executed. This method does not require a precise geometric model, has strong adaptability, and small force control overshoot, significantly improving the grinding quality of the robot. Attached Figure Description

[0019] Figure 1 This is a flowchart illustrating the first embodiment of the robot-based complex surface generation constant force grinding method based on a diffusion model according to the present invention. Figure 2 This is a structural block diagram of the first embodiment of the robot complex surface generation constant force grinding device based on the diffusion model of the present invention.

[0020] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0021] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0022] This invention provides a robot-based constant-force grinding method for complex surfaces based on a diffusion model, referring to... Figure 1 , Figure 1This is a flowchart illustrating the first embodiment of a robot-based constant-force grinding method for complex curved surfaces based on a diffusion model, according to the present invention.

[0023] In this embodiment, the robot complex surface generative constant force grinding method based on diffusion model includes the following steps: Step S10: Collect multimodal data of the polishing process, and construct a multimodal training dataset after temporal alignment and voxelization preprocessing of the multimodal data.

[0024] In this embodiment, the executing entity is a robot complex surface generation constant force grinding device based on a diffusion model. This robot complex surface generation constant force grinding device based on a diffusion model has functions such as data processing, data communication, and program execution. The robot complex surface generation constant force grinding device based on a diffusion model can be a computer terminal device or other network device, or other devices with similar functions. This embodiment does not limit the scope of such devices.

[0025] It should be noted that complex curved surface components such as water turbines and aero engines are prone to damage such as cavitation pits and cracks during service. Repaired weld areas often exhibit randomly distributed weld beads, pits, and abrupt hardness changes. Traditional offline programming methods based on CAD models fail due to the irregularity of the curved surface. Current robotic grinding often employs PID force-position hybrid control or impedance control, but when faced with sudden changes in contact stiffness (such as the grinding head impacting a hard weld point), it frequently results in force oscillations, overcutting, or even tool breakage. Furthermore, the implicit knowledge of highly skilled technicians adjusting grinding angles and pressure through "feel" is difficult to quantify into traditional control code, leading to robotic grinding quality relying on manual teaching, resulting in low efficiency and poor consistency.

[0026] To address the aforementioned technical challenges, this embodiment collects multimodal data to construct a training dataset. The generation of the grinding action sequence is modeled as a conditional backdiffusion process. Using current multimodal observations as input, a noise prediction network is trained to fit the expert operation distribution, resulting in a conditional diffusion strategy model. During execution, the model receives the topography, six-dimensional forces, end-effector pose, and joint states of the visual image of the area to be ground in real-time using a rolling time-domain approach. After denoising and iterative generation, a future action sequence is generated and input into an impedance controller to convert it into joint torque commands. Simultaneously, a safety watchdog is set up to compare the deviation between the expected force and the measured force. If an overcut or hard collision risk is detected, control is immediately taken over and a rollback is executed. This method eliminates the need for a precise geometric model, exhibits strong adaptability, and has a small force control overshoot, significantly improving the robot's grinding quality.

[0027] In the specific implementation, a force feedback teleoperation device is first used, with a senior technician demonstrating the polishing process, while simultaneously collecting multimodal data: RGB-D visual images (such as a 640×480 depth map output by Intel RealSense D435i), six-dimensional contact force / torque (such as ATI Gamma IP68, range 130N, accuracy 0.1N), robot end-effector pose (obtained by joint encoders), and joint velocity. The acquisition frequency is no less than 50Hz to ensure the temporal continuity of the action sequence. The collected multimodal data undergoes temporal alignment (linear interpolation to fill in missing frames based on a unified timestamp) and voxelization preprocessing (downsampling the point cloud to 0.5mm resolution) to construct a multimodal training dataset. Each sample contains the observation vector at the current moment (fusing RGB-D visual features and historical force perception sequences) and an expert action sequence A (including end-effector pose adjustment and normal contact force setting instructions) for the next k steps (e.g., k=16, corresponding to the next 0.32s). Timing alignment includes hardware synchronization triggering, such as providing synchronization signals to each sensor through the same clock source (e.g., PTP protocol or hardware trigger line) so that they start sampling at the same time; and software interpolation alignment, such as selecting the lowest sampling frequency (e.g., 50Hz) as the reference for sensors that cannot be hardware synchronized, and downsampling or sliding window averaging for other high-frequency data; and using linear interpolation or nearest neighbor interpolation for low-frequency data (e.g., depth images) to complete to each reference time point. Voxelization preprocessing specifically involves the following implementation methods: transforming the point cloud in the camera coordinate system to the robot base coordinate system or the grinding head tool coordinate system; and setting the voxel side length (e.g., 0.5mm, 1mm, or 2mm). Based on the bounding box of the area to be polished, an L×W×H grid is created. The voxel index of each point is calculated, and points falling within the same voxel are averaged (position, color), or the presence of a point within that voxel is counted (occupancy flag). Each voxel can store multiple channels: occupancy probability (0 or 1), average height (relative to the reference plane), and normal vector components (calculated via PCA), ultimately forming a 4D tensor (depth × height × width × channels) of size D×H×W×C. Finally, a voxel grid filter is used to remove isolated points and homogenize high-density areas. For example, with a side length of 1mm, a 120mm×80mm×10mm area will generate approximately 120×80×10=96000 voxels. For computationally demanding scenarios, this can be further coarsened to 2mm voxels (approximately 24000 voxels). Step S20: Model the generation process of the grinding action sequence as a conditional backdiffusion process. Using the multimodal observation vector at the current moment as the input condition, train a noise-based prediction neural network using the multimodal training dataset. Iteratively optimize and fit the action distribution law of expert operation. After training, the conditional diffusion strategy model is obtained.

[0028] In the specific implementation, the generation process of the polishing action sequence A is modeled as a conditional back-diffusion process. The diffusion model includes a forward diffusion process (gradually adding Gaussian noise to the data until it becomes pure noise) and a back-diffusion denoising process (gradually recovering the action sequence from the noise). The denoising iteration formula is as follows:

[0029] in, For the first The noise-enhanced action sequence tensor of the diffusion time step. This is the current multimodal observation vector, which incorporates RGB-D visual features. With historical sense sequence This is a Transformer-based neural network for noise prediction, used to estimate and remove noise. These are noise scheduling parameters that control the denoising step size. Injected Gaussian random noise is used to enhance the diversity and perturbation resistance of the strategy. The standard deviation parameter of the Gaussian noise scheduling corresponding to the k-th denoising iteration is denoised as follows.

[0030] In this embodiment, the diffusion step number is set to 100 steps. After training, the conditional diffusion strategy model supports outputting a grinding action sequence that conforms to the expert operation distribution after multiple denoising iterations from Gaussian noise input (accelerated sampling is usually used during inference, such as 16-step denoising). The grinding action sequence includes end pose adjustment and normal contact force setting instructions for several future steps (e.g., the next 2 seconds, a total of 100 control points).

[0031] In one embodiment, the denoising iteration process of the conditional diffusion strategy model is implemented through the above formula, and a denoising diffusion implicit model (DDIM) is used to accelerate sampling, with a single inference time of less than 0.1s.

[0032] Step S30: Load the conditional diffusion strategy model, and receive in real time the visual image of the area to be polished, six-dimensional contact force data, robot end pose and joint velocity data of the current area to be polished, and fuse them into a multimodal observation vector as the model input. Based on the input observation vector, generate a polishing action sequence of a future preset duration through rolling temporal denoising iteration, and input the polishing action sequence into the impedance controller to convert it into joint torque commands to drive the robotic arm to perform the polishing operation.

[0033] In the specific implementation, the robot system loads a pre-trained conditional diffusion strategy model. During the actual polishing process, the robot collects RGB-D topographic data, six-dimensional contact force data, end-effector pose, and joint velocity of the area to be polished in real time at a frequency of 50Hz, fusing them to obtain the current observation vector. Using this observation as a condition, the model starts with random Gaussian noise and performs 16 denoising iterations to generate a polishing action sequence for the next 2 seconds (containing the desired pose and desired normal contact force at 100 discrete time points). Then, a Receding Horizon Control (RHC) strategy is adopted: each time a new sequence is generated, only the first control point (or the first 5 points) in the sequence is executed, and then the sequence is re-planned based on new observations, thereby achieving real-time response to environmental changes.

[0034] The impedance controller converts the grinding motion sequence output by the conditional diffusion strategy model into joint torque commands using the following formula:

[0035] Where τcmd is the final control torque command for each joint of the robot output by the impedance controller. Let Λ(q) be the transpose of the Jacobian matrix corresponding to the current joint angle q of the robot, where q is the actual angle vector of each joint of the robot at the current time, which is collected in real time by the robot joint encoder, and Λ(q) is the operational space inertial matrix under the current joint pose. d represents the desired end-effector acceleration vector in the grinding action sequence output by the conditional diffusion strategy model, and B represents the desired end-effector acceleration vector. d Let be the desired damping matrix output in real time by the conditional diffusion strategy model, and d be the desired end velocity vector in the grinding action sequence output by the conditional diffusion strategy model. K represents the actual velocity vector of the robot's end effector. d X is the desired stiffness matrix output in real time by the conditional diffusion strategy model. d Let X be the desired end-effector pose vector in the grinding action sequence output by the conditional diffusion strategy model, and F be the actual pose vector of the robot's end-effector. d The normal contact force feedforward term generated for the conditional diffusion strategy model, g(q) is the gravity compensation torque vector under the current joint pose, and C( , ) represents the Coriolis force and centrifugal force compensation torque vector under the current joint pose and velocity.

[0036] In one embodiment, the conditional diffusion strategy model continuously reprograms the action sequence for the next 2 seconds at a frequency of not less than 50 Hz. When the normal contact force abruptly changes by more than 15 N (e.g., the grinding head impacts a high-hardness weld bead), the model automatically reduces the trajectory feed rate in the output action sequence (e.g., from 10 mm / s to 5 mm / s) and improves the smoothness of the action (by increasing the curvature continuity constraint of the trajectory).

[0037] In one embodiment, when the conditional diffusion strategy model identifies a high-hardness weld bead (with drastic changes in the point cloud normal and abnormal grayscale) in the area to be ground using the input RGB-D visual features, it automatically adjusts the feed rate to 5 mm / s and the normal contact force setting to 25 N in the output action sequence. If the rate of change of the normal contact force exceeds 50 N / s during the grinding process, the impedance controller instantaneously reduces the stiffness parameter K. d (For example, reducing from 2000 N / m to 500 N / m) softens the grinding head and allows it to slide tangentially, preventing chipping.

[0038] Step S40: Real-time synchronous acquisition of the expected normal contact force command output by the conditional diffusion strategy model and the six-dimensional contact force data collected during the grinding process, and comparison of the deviation characteristics of the two sets of data based on the preset safety judgment rules. If it is determined that there is a risk of overcutting or hard collision, the robot control is immediately taken over and deceleration, safe retreat or emergency stop actions are executed simultaneously. The abnormal working condition data is synchronously transmitted back to the training library of the conditional diffusion strategy model for subsequent incremental optimization of the model.

[0039] In practical implementation, to prevent the generative model from generating "illusions" that could lead to overcutting or collisions, this embodiment sets up an independent physical constraint watchdog. This watchdog synchronously acquires the expected normal contact force command output by the model and the measured values ​​from the force sensor in real time at a monitoring frequency of 1kHz, and makes a judgment using the following safety scoring function:

[0040] in, For indicator functions, For the rate of change of force, The measured normal contact force value is collected in real time by a six-dimensional force sensor during the grinding process. This represents the desired normal contact force command value output in real time by the conditional diffusion strategy model. This is the rated maximum normal contact force of the grinding system. For safety attenuation coefficient, This is the allowable deviation threshold for normal contact force. It is a natural exponential function.

[0041] In one embodiment, the security determination is performed using the aforementioned security scoring function, and in conjunction with a hardware watchdog circuit, a protection action can be triggered within 1ms after an anomaly is detected, which is much faster than the response time of the main controller.

[0042] In this embodiment, multimodal data is collected to construct a training dataset. The generation of the grinding action sequence is modeled as a conditional backdiffusion process. Using the current multimodal observations as input, a noise prediction network is trained to fit the expert operation distribution, resulting in a conditional diffusion strategy model. During execution, the model receives the topography, six-dimensional forces, end-effector pose, and joint states of the visual image of the area to be ground in real time using a rolling time-domain approach. After denoising and iterative generation, the future action sequence is generated and input into an impedance controller to be converted into joint torque commands. Simultaneously, a safety watchdog is set up to compare the deviation between the expected force and the measured force. If an overcut or hard collision risk is detected, control is immediately taken over and a rollback is executed. This method does not require a precise geometric model, has strong adaptability, and small force control overshoot, significantly improving the grinding quality of the robot.

[0043] Furthermore, this embodiment of the invention also proposes a storage medium storing a robot complex surface generation constant force grinding program based on a diffusion model. When the robot complex surface generation constant force grinding program based on a diffusion model is executed by a processor, it implements the steps of the robot complex surface generation constant force grinding method based on a diffusion model as described above.

[0044] Reference Figure 2 , Figure 2 This is a structural block diagram of the first embodiment of the robot complex surface generation constant force grinding device based on the diffusion model of the present invention.

[0045] like Figure 2 As shown, the robotic complex surface generation constant force grinding device based on a diffusion model proposed in this embodiment of the invention includes: The data acquisition module 10 is used to acquire multimodal data of the polishing process. After the multimodal data is preprocessed by time alignment and voxelization, a multimodal training dataset is constructed. The multimodal data includes visual images, six-dimensional contact force data, robot end pose data, and joint velocity data. The model training module 20 is used to model the generation process of the grinding action sequence as a conditional backdiffusion process. It uses the multimodal observation vector at the current moment as the input condition, and trains a noise-based prediction neural network using the multimodal training dataset. It iteratively optimizes and fits the action distribution law of expert operations. After training, a conditional diffusion strategy model is obtained. The conditional diffusion strategy model supports outputting a grinding action sequence that conforms to the distribution of expert operations after multiple denoising iterations from Gaussian noise input. The grinding action sequence includes the end pose adjustment amount and normal contact force setting instructions for several future steps. The model processing module 30 is used to load the conditional diffusion strategy model. The conditional diffusion strategy model receives the visual image of the area to be polished, six-dimensional contact force data, robot end pose and joint velocity data in real time, and fuses them into a multimodal observation vector as model input. Based on the input observation vector, a polishing action sequence of a future preset duration is generated through rolling temporal denoising iteration. The polishing action sequence is input to the impedance controller and converted into joint torque commands to drive the robotic arm to perform the polishing operation. During the polishing process, the conditional diffusion strategy model dynamically adjusts the polishing parameters according to the observation data updated once at a preset frequency. The anomaly detection module 40 is used to synchronously acquire the expected normal contact force command output by the conditional diffusion strategy model and the six-dimensional contact force data collected during the grinding process in real time. Based on the preset safety judgment rules, it compares the deviation characteristics of the two sets of data. If it is determined that there is a risk of overcutting or hard collision, it immediately takes over the control of the robot and executes deceleration, safe retreat or emergency stop actions. The abnormal working condition data is synchronously transmitted back to the training library of the conditional diffusion strategy model for subsequent incremental optimization of the model.

[0046] In this embodiment, multimodal data is collected to construct a training dataset. The generation of the grinding action sequence is modeled as a conditional backdiffusion process. Using the current multimodal observations as input, a noise prediction network is trained to fit the expert operation distribution, resulting in a conditional diffusion strategy model. During execution, the model receives the topography, six-dimensional forces, end-effector pose, and joint states of the visual image of the area to be ground in real time using a rolling time-domain approach. After denoising and iterative generation, the future action sequence is generated and input into an impedance controller to be converted into joint torque commands. Simultaneously, a safety watchdog is set up to compare the deviation between the expected force and the measured force. If an overcut or hard collision risk is detected, control is immediately taken over and a rollback is executed. This method does not require a precise geometric model, has strong adaptability, and small force control overshoot, significantly improving the grinding quality of the robot.

[0047] This application also provides a robot complex surface generation constant force grinding device based on a diffusion model, including a processor, a communication interface, a memory, and a communication bus. The processor, communication interface, and memory communicate with each other through the communication bus. The memory is used to store the robot complex surface generation constant force grinding program based on the diffusion model. When the processor executes the program stored in the memory, it implements the above-mentioned robot complex surface generation constant force grinding method based on the diffusion model.

[0048] The communication bus mentioned in the aforementioned diffusion-based robotic complex surface generative constant force grinding equipment can be a Peripheral Component Interconnect (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. This communication bus can be divided into address bus, data bus, control bus, etc.

[0049] The communication interface is used for communication between the aforementioned diffusion-based robotic complex surface generative constant force grinding equipment and other devices.

[0050] The memory may include random access memory (RAM) or non-volatile memory (NVM), such as at least one disk storage device. Optionally, the memory may also be at least one storage device located remotely from the aforementioned processor.

[0051] The processors mentioned above can be general-purpose processors, including central processing units (CPUs), network processors (NPs), etc.; they can also be digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components.

[0052] In the above embodiments, implementation can be achieved entirely or partially through software, hardware, firmware, or any combination thereof. When implemented using software, it can be implemented entirely or partially in the form of a computer program product. The computer program product includes one or more computer instructions. When the computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that integrates one or more available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium (e.g., solid state disk (SSD)).

[0053] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0054] The various embodiments in this specification are described in a related manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions of the method embodiments.

[0055] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

[0056] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0057] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0058] In addition, for technical details not described in detail in this embodiment, please refer to the robot complex surface generation constant force grinding method based on diffusion model provided in any embodiment of the present invention, which will not be repeated here.

[0059] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0060] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0061] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0062] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

[0063] It is understood that the system provided in the embodiments of the present invention corresponds to the method provided in the embodiments of the present invention, and the explanation, examples and beneficial effects of the relevant content can be referred to the corresponding parts of the above methods.

Claims

1. A robot-based constant-force grinding method for complex curved surfaces based on a diffusion model, characterized in that, The robot-based complex surface generative constant force grinding method based on the diffusion model includes: Multimodal data of the polishing process is collected, and the multimodal data is preprocessed by time alignment and voxelization to construct a multimodal training dataset. The multimodal data includes visual images, six-dimensional contact force data, robot end-effector pose data, and joint velocity data. The generation process of the grinding action sequence is modeled as a conditional backdiffusion process. The multimodal observation vector at the current moment is used as the input condition. The noise-based prediction neural network is trained using the multimodal training dataset. The action distribution law of the expert operation is fitted by iterative optimization. After training, the conditional diffusion strategy model is obtained. The conditional diffusion strategy model supports the output of a grinding action sequence that conforms to the distribution of expert operation after multiple denoising iterations from Gaussian noise input. The grinding action sequence includes the end pose adjustment amount and normal contact force setting instructions for several future steps. The conditional diffusion strategy model is loaded, and the visual image of the area to be polished, six-dimensional contact force data, robot end pose and joint velocity data are received in real time through the conditional diffusion strategy model. These data are fused into a multimodal observation vector as the model input. Based on the input observation vector, a polishing action sequence of a future preset duration is generated through rolling temporal denoising iteration. The polishing action sequence is then input into an impedance controller and converted into joint torque commands to drive the robotic arm to perform the polishing operation. During the polishing process, the conditional diffusion strategy model dynamically adjusts the polishing parameters according to the observation data updated once at a preset frequency. The system synchronously acquires the expected normal contact force command output by the conditional diffusion strategy model and the six-dimensional contact force data collected during the grinding process. Based on the preset safety judgment rules, the system compares the deviation characteristics of the two sets of data. If it determines that there is a risk of overcutting or hard collision, it immediately takes over the control of the robot and executes deceleration, safe retreat, or emergency stop actions. The abnormal working condition data is synchronously transmitted back to the training library of the conditional diffusion strategy model for subsequent incremental optimization of the model.

2. The robot complex surface generative constant force grinding method based on diffusion model as described in claim 1, characterized in that, The denoising iterative process of the conditional diffusion strategy model is implemented through the following formula: in, For the first The noise-enhanced action sequence tensor of the diffusion time step. This is the current multimodal observation vector, which incorporates RGB-D visual features. With historical sense sequence This is a Transformer-based neural network for noise prediction, used to estimate and remove noise. These are noise scheduling parameters that control the denoising step size. Injected Gaussian random noise is used to enhance the diversity and perturbation resistance of the strategy. The standard deviation parameter of the Gaussian noise scheduling corresponding to the k-th denoising iteration is denoised as follows.

3. The robot-based constant-force grinding method for complex curved surfaces based on a diffusion model as described in claim 2, characterized in that, The conditional diffusion strategy model continuously replans the action sequence for the next 2 seconds at a frequency of no less than 50Hz. When the normal contact force mutation amplitude exceeds 15N, the conditional diffusion strategy model automatically reduces the trajectory feed speed and improves the smoothness of the action sequence in the output action sequence.

4. The robot complex surface generative constant force grinding method based on diffusion model as described in claim 1, characterized in that, The impedance controller converts the grinding action sequence output by the conditional diffusion strategy model into joint torque commands using the following formula: Where τcmd is the final control torque command for each joint of the robot output by the impedance controller. Let Λ(q) be the transpose of the Jacobian matrix corresponding to the current joint angle q of the robot, where q is the actual angle vector of each joint of the robot at the current time, which is collected in real time by the robot joint encoder, and Λ(q) is the operational space inertial matrix under the current joint pose. d represents the desired end-effector acceleration vector in the grinding action sequence output by the conditional diffusion strategy model, and B represents the desired end-effector acceleration vector. d Let be the desired damping matrix output in real time by the conditional diffusion strategy model, and d be the desired end velocity vector in the grinding action sequence output by the conditional diffusion strategy model. K represents the actual velocity vector of the robot's end effector. d X is the desired stiffness matrix output in real time by the conditional diffusion strategy model. d Let X be the desired end-effector pose vector in the grinding action sequence output by the conditional diffusion strategy model, and F be the actual pose vector of the robot's end-effector. d The normal contact force feedforward term generated for the conditional diffusion strategy model, g(q) is the gravity compensation torque vector under the current joint pose, and C( , ) represents the Coriolis force and centrifugal force compensation torque vector under the current joint pose and velocity.

5. The robot complex surface generative constant force grinding method based on diffusion model as described in claim 1, characterized in that, When the conditional diffusion strategy model identifies the area to be ground as a high-hardness weld bead through the input RGB-D visual features, it automatically adjusts the feed speed to 5mm / s and the normal contact force setting to 25N in the output action sequence. If the normal contact force change rate is detected to exceed 50N / s during the grinding process, the impedance controller instantly reduces the stiffness parameter.

6. The robot complex surface generative constant force grinding method based on diffusion model as described in claim 1, characterized in that, Security is determined using the following security scoring function: in, For indicator functions, For the rate of change of force, The measured normal contact force value is collected in real time by a six-dimensional force sensor during the grinding process. This represents the desired normal contact force command value output in real time by the conditional diffusion strategy model. This is the rated maximum normal contact force of the grinding system. For safety attenuation coefficient, This is the allowable deviation threshold for normal contact force. It is a natural exponential function.

7. A robotic complex surface generation constant force grinding device based on a diffusion model, characterized in that, The diffusion-model-based robot complex surface generative constant force grinding device is applied to the diffusion-model-based robot complex surface generative constant force grinding method as described in any one of claims 1 to 6, wherein the device comprises: The data acquisition module is used to collect multimodal data of the polishing process. After the multimodal data is preprocessed by time alignment and voxelization, a multimodal training dataset is constructed. The multimodal data includes visual images, six-dimensional contact force data, robot end pose data, and joint velocity data. The model training module is used to model the generation process of the grinding action sequence as a conditional backdiffusion process. It uses the multimodal observation vector at the current moment as the input condition, and trains a noise-based prediction neural network using the multimodal training dataset. Through iterative optimization, it fits the action distribution law of expert operations. After training, a conditional diffusion strategy model is obtained. The conditional diffusion strategy model supports outputting a grinding action sequence that conforms to the distribution of expert operations after multiple denoising iterations from Gaussian noise input. The grinding action sequence includes end pose adjustment amount and normal contact force setting instructions for several future steps. The model processing module is used to load the conditional diffusion strategy model. The conditional diffusion strategy model receives the visual image of the area to be polished, six-dimensional contact force data, robot end pose and joint velocity data in real time, and fuses them into a multimodal observation vector as the model input. Based on the input observation vector, a polishing action sequence of a future preset duration is generated through rolling temporal denoising iteration. The polishing action sequence is input to the impedance controller and converted into joint torque commands to drive the robotic arm to perform the polishing operation. During the polishing process, the conditional diffusion strategy model dynamically adjusts the polishing parameters according to the observation data updated once at a preset frequency. The anomaly detection module is used to synchronously acquire the expected normal contact force command output by the conditional diffusion strategy model and the six-dimensional contact force data collected during the grinding process in real time. Based on the preset safety judgment rules, it compares the deviation characteristics of the two sets of data. If it determines that there is a risk of overcutting or hard collision, it immediately takes over the control of the robot and executes deceleration, safe retreat or emergency stop actions. The abnormal working condition data is synchronously transmitted back to the training library of the conditional diffusion strategy model for subsequent incremental optimization of the model.

8. The robotic complex surface generation constant force grinding device based on a diffusion model as described in claim 7, characterized in that, The conditional diffusion strategy model continuously replans the action sequence for the next 2 seconds at a frequency of no less than 50Hz. When the normal contact force mutation amplitude exceeds 15N, the conditional diffusion strategy model automatically reduces the trajectory feed speed and improves the smoothness of the action sequence in the output action sequence.

9. A robotic complex surface generation constant force grinding device based on a diffusion model, characterized in that, The diffusion model-based robot complex surface generation constant force grinding device includes: a memory, a processor, and a diffusion model-based robot complex surface generation constant force grinding program stored in the memory and executable on the processor. The diffusion model-based robot complex surface generation constant force grinding program is configured to implement the steps of the diffusion model-based robot complex surface generation constant force grinding method as described in any one of claims 1 to 6.

10. A storage medium, characterized in that, The storage medium stores a robot complex surface generation constant force grinding program based on a diffusion model. When the robot complex surface generation constant force grinding program based on a diffusion model is executed by the processor, it implements the steps of the robot complex surface generation constant force grinding method based on a diffusion model as described in any one of claims 1 to 6.

Images