Robot shaft hole assembly search method and system based on similar feature space fusion
By constructing a visual-touch fusion platform and a similar feature space, and combining PID force control and near-end strategy optimization algorithms, the problem of insufficient precision in shaft and hole assembly of assembly robots was solved, and high-precision fully automated assembly was achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- SHANDONG UNIV
- Filing Date
- 2023-03-13
- Publication Date
- 2026-07-14
AI Technical Summary
Existing assembly robots lack sufficient search and positioning accuracy in shaft and hole assembly operations, especially in complex operations, resulting in low assembly success rates and making it difficult to achieve high-precision full automation.
A shaft-hole assembly platform based on visual-touch fusion is constructed. Image segmentation network and force sensor are used to extract shaft-hole assembly features. A similar feature space is constructed through transfer learning. Combined with PID force control and proximal strategy optimization algorithm, the generalization ability of the assembly strategy is improved.
It improves the generalization ability of robot assembly, adapts to various types of shaft and hole assembly, reduces human intervention and costs, and achieves high-precision fully automated assembly.
Smart Images
Figure CN116100555B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of assembly robot technology, and particularly relates to a robot shaft and hole assembly search method and system based on similar feature space fusion. Background Technology
[0002] The statements in this section are merely background information related to the present invention and do not necessarily constitute prior art.
[0003] Assembly robots are widely used in manufacturing, especially in shaft and hole assembly operations. However, achieving higher search and positioning accuracy remains a challenge for more complex operations. Further advancements in search strategies hold promise for improving the success rate of assembly operations, reducing human intervention, lowering costs, and ultimately achieving full automation of high-precision assembly.
[0004] Search strategies rely on a large amount of data from sensors. Therefore, the number and type of sensors, data processing algorithms, and the integration of sensors and search strategies are the main factors that determine the robustness, stability, and search accuracy of robot assembly. In addition, the types of shaft-hole assemblies are diverse, making it very difficult to find a universal search strategy. Summary of the Invention
[0005] To overcome the shortcomings of the prior art, this invention provides a robot shaft hole assembly search method and system based on similar feature space fusion. Under the framework of feature transfer learning, based on the constructed vision-touch fusion shaft hole assembly platform, the search feature space of different shaft hole assembly tasks is extracted. Based on the MMD algorithm, a similar feature space between the source task and the target task is constructed. Combined with force control and policy learning, it can adapt to various types of shaft hole assembly and electronic connector assembly, and improve the generalization ability of robot assembly.
[0006] To achieve the above objectives, one or more embodiments of the present invention provide the following technical solutions:
[0007] The first aspect of this invention provides a robot shaft hole assembly search method based on similar feature space fusion;
[0008] A robot shaft hole assembly search method based on similar feature space fusion includes:
[0009] A feature extraction model was constructed and trained using the collected shaft and hole assembly data;
[0010] Based on the trained feature extraction model, shaft hole assembly image features and shaft hole assembly position features are extracted from the source assembly task with existing search strategy and the target assembly task without search strategy, respectively, to obtain the source task feature space and the target task feature space.
[0011] Based on the maximum mean difference algorithm, the mean distance between the feature space of the source task and the feature space of the target task is calculated, a feature space mapping is constructed, and a similar feature space for different assembly tasks is built.
[0012] Based on the constructed similar feature space, a method combining PID force control and proximal strategy optimization algorithm is used to learn and optimize the search strategy to obtain the search strategy for the target assembly task.
[0013] Furthermore, the construction of the shaft hole assembly data is based on a vision-touch fusion shaft hole assembly platform;
[0014] The shaft hole assembly data includes shaft hole assembly status images captured by the camera, force information collected by the force sensor, and the end effector pose of the robotic arm.
[0015] Furthermore, the shaft hole assembly platform includes two cameras in the x and y directions, a force sensor, a robotic arm, and a shaft hole assembly worktable;
[0016] Furthermore, the feature extraction model takes the shaft-hole assembly state image, force information, and robotic arm end pose as input to generate and output shaft-hole assembly image features and shaft-hole assembly position features.
[0017] Furthermore, the feature extraction model specifically includes an image segmentation network and a location feature extraction module;
[0018] The image segmentation network, built on U-Net, performs image segmentation and feature extraction dimensionality reduction on the input shaft hole assembly state image to obtain shaft hole assembly image features;
[0019] The position feature extraction module extracts the end pose of the robotic arm and the contact force / torque generated during the shaft hole assembly process as the position features of the shaft hole assembly.
[0020] Furthermore, the feature space mapping is specifically as follows:
[0021]
[0022] Where, n S and n D Let G(R) represent the number of samples in the source and target domains, respectively. The distance calculated by G(R) is the distance in the source task feature space R. S and the target task feature space R D The maximum mean difference between them, and the source task feature space R S and the target task feature space R D satisfy:
[0023] R D =R S +G(R)
[0024] Furthermore, the method of combining PID force control with a proximal strategy optimization algorithm to learn and optimize the search strategy specifically involves:
[0025] Obtain the necessary actions for assembly based on the assembly status of the shaft and hole;
[0026] Then, the robotic arm is controlled by PID force control to search for the hole in the xy plane, ensuring that the contact force between the shaft and the edge of the hole remains constant during the search.
[0027] Determine if the current state meets the conditions for a successful / failed search to decide whether to end the search.
[0028] A second aspect of the present invention provides a robot shaft hole assembly search system based on similar feature space fusion.
[0029] A robot shaft hole assembly search system based on similar feature space fusion includes an extraction model construction module, a feature space extraction module, a similar space construction module, and a policy learning and optimization module.
[0030] The extraction model building module is configured to: build a feature extraction model and train the feature extraction model using the collected shaft hole assembly data;
[0031] The feature space extraction module is configured to: extract shaft hole assembly image features and shaft hole assembly position features from the source assembly task with an existing search strategy and the target assembly task without a search strategy, based on the trained feature extraction model, to obtain the source task feature space and the target task feature space.
[0032] The similarity space construction module is configured to: calculate the mean distance between the feature space of the source task and the feature space of the target task based on the maximum mean difference algorithm, construct the feature space mapping, and build a similar feature space for different assembly tasks;
[0033] The strategy learning and optimization module is configured to learn and optimize the search strategy based on the constructed similar feature space, using a combination of PID force control and proximal strategy optimization algorithm, to obtain the search strategy for the target assembly task.
[0034] A third aspect of the present invention provides a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps in the robot shaft hole assembly search method based on similar feature space fusion as described in the first aspect of the present invention.
[0035] The fourth aspect of the present invention provides an electronic device, including a memory, a processor, and a program stored in the memory and executable on the processor, wherein the processor executes the program to implement the steps in the robot shaft hole assembly search method based on similar feature space fusion as described in the first aspect of the present invention.
[0036] The above one or more technical solutions have the following beneficial effects:
[0037] This invention addresses the hole search problem in shaft-hole assembly tasks by constructing a shaft-hole assembly platform based on visual-touch fusion. A search feature space is built based on images, contact forces, and end-effector poses. The U-Net image segmentation network is used to extract shaft-hole features. Transfer learning is employed to construct a mapping relationship between the search feature spaces for different shaft-hole assembly tasks. Regarding search strategies, a strategy transfer method is used to learn strategies at different shaft-hole search stages. Combining force control and strategy learning, this approach can adapt to various types of shaft-hole assembly and electronic connector assembly, improving the robot's generalization ability in assembly.
[0038] Advantages of additional aspects of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description
[0039] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.
[0040] Figure 1 This is a method framework diagram for the first embodiment.
[0041] Figure 2 This is a structural diagram of the shaft hole assembly platform of the first embodiment.
[0042] Figure 3 This is a schematic diagram of the annotation results for the first embodiment.
[0043] Figure 4 This is a diagram of the U-Net network structure of the first embodiment.
[0044] Figure 5 This is a schematic diagram of the image feature extraction results of the first embodiment.
[0045] Figure 6 The flowchart shows the near-end policy optimization algorithm of the first embodiment. Detailed Implementation
[0046] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0047] Example 1
[0048] This embodiment discloses a robot shaft hole assembly search method based on similar feature space fusion;
[0049] Figure 1 A framework diagram for the robot shaft hole assembly search method, such as...Figure 1 As shown, the robot shaft hole assembly search method based on similar feature space fusion includes:
[0050] Step S1: Construct a feature extraction model and train the feature extraction model using the collected shaft and hole assembly data.
[0051] This embodiment uses a vision-touch fusion-based shaft-hole assembly platform for data acquisition. Figure 2 This is a structural diagram of the shaft and hole assembly platform, such as... Figure 2 As shown, the shaft hole assembly platform includes two industrial cameras (one in the x-direction and one in the y-direction), a six-dimensional force sensor, a robotic arm, and a shaft hole assembly worktable. During the shaft hole assembly process, the industrial cameras acquire images of the shaft hole assembly status, and the force sensor acquires force information (i.e., contact force / torque) and the end-effector pose of the robotic arm.
[0052] The feature extraction model is used to extract shaft and hole assembly image features and shaft and hole assembly position features from shaft and hole assembly data composed of shaft and hole assembly state images, force information and robot arm end pose. Therefore, the feature extraction model includes an image segmentation network and a position feature extraction module.
[0053] Image segmentation networks, based on image segmentation and image feature extraction, specifically:
[0054] (1) A random search strategy is used to search for shaft holes in the source assembly task. Images of the shaft hole assembly status are acquired using an industrial camera, and the acquired images are labeled with shaft holes to construct a shaft hole dataset; the labeling results are as follows. Figure 3 As shown:
[0055] (2) Construct an image segmentation network (U-Net network), the network structure is as follows: Figure 4 As shown, the U-Net network is pre-trained based on the constructed axis-hole dataset to obtain a trained image segmentation model.
[0056] The U-Net neural network can be viewed as an encoder on the left and a decoder on the right. It extracts target features through four downsampling operations and then classifies each pixel individually through four upsampling operations. The encoder has four sub-modules, each containing two convolutional layers, followed by a downsampling layer implemented with max pooling. The decoder also contains four sub-modules, with the resolution increasing sequentially through upsampling operations until it matches the resolution of the input image. The network uses skip connections, concatenating the upsampling results with the outputs of sub-modules in the encoder with the same resolution, which then serve as the input to the next sub-module in the decoder.
[0057] (3) An industrial camera captures an image of the current shaft hole assembly status. The trained U-Net network extracts assembly image features I, where I is a one-dimensional array. The result of the image feature extraction is...Figure 5 As shown, the images are, in order: before image segmentation, the segmented holes, and the segmented axes.
[0058] When performing image segmentation for a target assembly task, the weight parameters in the U-Net network model obtained from the pre-training of the source assembly task are fixed, and only the weight parameters of the input and output layers are unrestricted for training. This enables the reuse of the U-Net network, improves data utilization efficiency, and reduces training costs.
[0059] The location feature extraction module, based on a six-dimensional force sensor, extracts location features, specifically as follows:
[0060] (1) Using a six-dimensional force sensor in the shaft-hole assembly platform, the end-effector pose P = (p_i) of the robotic arm during the shaft-hole assembly process is collected. x ,p y ,p z ,o x ,o y ,o z The resulting contact force / torque Γ = (F, M), where p and o are the translation and rotation components, respectively, x, y, z represent the coordinate axes of the basic coordinate system, F is the end contact force, and M is the contact torque.
[0061] (2) Standardize the end effector pose P of the robotic arm so that the pose and force / torque are of the same order of magnitude. The standardized pose and force / torque are position features. The standardization method is as follows:
[0062]
[0063] Among them, P min P is the minimum boundary value that the robotic arm can reach in space. max P' represents the maximum boundary value that the robotic arm can reach in space, and P' represents the processed end-effector pose of the robotic arm.
[0064] Step S2: Based on the trained feature extraction model, extract shaft hole assembly image features and shaft hole assembly position features from the source assembly task with existing search strategy and the target assembly task without search strategy, respectively, to obtain the source task feature space and the target task feature space.
[0065] Based on image segmentation, the feature space F of the shaft hole assembly image is extracted. Simultaneously, the force sensor and the robotic arm end-effector pose form the feature space P of the shaft hole assembly position. Together, they constitute the feature space R in the shaft hole assembly, expressed by the formula:
[0066] R = F × P
[0067] The image feature space F is a one-dimensional array obtained by feature extraction and dimensionality reduction of the original image after processing by the U-Net network.
[0068] Step S3: Based on the maximum mean difference algorithm, calculate the mean distance between the feature space of the source task and the feature space of the target task, construct the feature space mapping, and build a similar feature space for different assembly tasks.
[0069] The processed image features and positional features constitute the feature space R required for assembly. The feature space is different for different shaft and hole assembly tasks. A similar feature space is constructed based on the robotic arm's working space and the shaft and hole assembly search process. The search strategy for the source assembly task is learned in the similar feature space.
[0070] Source assembly task X based on existing search strategies S And target assembly task X without search strategy D And construct the following feature map G(R), with the formula:
[0071]
[0072] Where, n S and n D Let G(R) represent the number of samples in the source and target domains, respectively. The distance calculated by G(R) is the distance in the source task feature space R. S and the target task feature space R D The maximum mean difference (MMD) between them. And the source task feature space R S and the target task feature space R D satisfy:
[0073] R D =R S +G(R)
[0074] The source assembly task and the target assembly task establish a connection between their feature spaces through feature mapping, thereby realizing feature transfer from the source domain to the target domain.
[0075] Step S4: Based on the constructed similar feature space, the search strategy is learned and optimized by combining PID force control and proximal strategy optimization algorithm to obtain the search strategy for the target assembly task.
[0076] The Proximal Policy Optimization (PPO) algorithm is used to obtain the shaft hole search strategy. As an online update strategy, PPO requires the agent to continuously interact with the shaft hole assembly environment. The decision sequence τ = (s0, a0, r0, s1, a1, r1, ..., s) generated by the interaction is used. T ,a T ,r T This algorithm updates the PPO policy learning model, where s, a, and r represent the state, action, and reward value, respectively, and T is the maximum time step. Algorithm 1 provides the update process. Figure 6 This is the near-end policy optimization algorithm flow, s t= (I,P,Γ) represents the shaft-hole assembly state at time t, I∈F is the image feature after processing by the U-Net network, P∈P is the end-effector pose, and Γ=(F,M)∈P is the contact force / torque generated during the shaft-hole assembly process. The specific steps are as follows:
[0077] (1) The Actor network is based on state s t Obtain the action a required for assembly t =(Δ x ,Δ y ), Δ x ,Δ y Let x be the displacement of the robotic arm in the xy plane.
[0078] (2) The robotic arm is then controlled by PID force control to search for the hole in the xy plane, ensuring that the contact force between the shaft and the edge of the hole remains constant during the search.
[0079] (3) Determine whether the current state meets the conditions for a successful / failed search to decide whether to end the search.
[0080] Conditions for determining a successful search: The contact force changes abruptly within the set maximum search steps; the average area of the holes obtained from the two cameras is less than the minimum area; and the offsets in the x and y directions are both within the gap range (0, 1 mm).
[0081] The conditions for determining search failure are: exceeding the set maximum search steps or the xy-direction offset exceeding the initial offset by 2cm.
[0082] Algorithm 1 Update Process
[0083]
[0084]
[0085] Decision sequence τ=(s0,a0,r0,s1,a1,r1,…,s T ,a T ,r T r in ) t To evaluate the reward value of the current state, the reward function is set to a continuous function to reduce assembly time and speed up assembly, specifically expressed as follows:
[0086] r1=-k / k max
[0087] r2 = 1 - (Ar1 + Ar2) / 2Ar max
[0088] Where r1 represents a negative reward when the ratio of the number of assembly steps used to the maximum number of assembly steps is negative, and k is the current number of assembly steps used. maxThe maximum number of assembly steps is set; r2 represents a positive reward as the ratio of the detected hole area to the area between the largest shaft holes is used; Ar1 and Ar2 are the hole areas obtained after image segmentation from the two cameras, respectively. max This represents the maximum area of the hole detected by the camera.
[0089] To improve data utilization efficiency, the policy trajectory τ = (s0, a0, r0, s1, a1, r1, ..., s) generated during the learning and optimization process will be used. T ,a T ,r T The data is stored in a pre-set experience pool, where T is the number of data points required for each update. Based on the data added to the experience pool, the interaction data and experience data are combined to obtain a new policy trajectory τ' = (s0, a0, r0, ..., s n ,a n ,r n ,…,s T ,a T ,r T ), where the ratio of interactive data to empirical data is ν = Tn / n, and the specific process is as follows: Figure 6 As shown; based on the policy trajectory obtained from the combination, importance sampling is used to update the current policy, where the weight B t (φ) is calculated using the following formula:
[0090] Where, π φ (a t |s t ), The new and old Actor networks are respectively based on the current state s t The output follows a normal distribution. To control the update magnitude of the policy, the PPO algorithm uses a truncation method to process the objective function, as shown in the formula:
[0091]
[0092] Among them, clip(B t (φ), 1-ε, 1+ε) will be the importance sampling weight B t (φ) is constrained within the range (1-ε, 1+ε), where ε is a hyperparameter, and A t The dominance function is as follows:
[0093]
[0094]
[0095] in, For time steps At that time, the Critic network evaluates the model, and γ is a decay factor used to reduce the impact of the evaluation from the previous time step on the current evaluation.
[0096] Example 2
[0097] This embodiment discloses a robot shaft hole assembly search system based on similar feature space fusion;
[0098] A robot shaft hole assembly search system based on similar feature space fusion includes an extraction model construction module, a feature space extraction module, a similar space construction module, and a policy learning and optimization module.
[0099] The extraction model building module is configured to: build a feature extraction model and train the feature extraction model using the collected shaft hole assembly data;
[0100] The feature space extraction module is configured to: extract shaft hole assembly image features and shaft hole assembly position features from the source assembly task with an existing search strategy and the target assembly task without a search strategy, based on the trained feature extraction model, to obtain the source task feature space and the target task feature space.
[0101] The similarity space construction module is configured to: calculate the mean distance between the feature space of the source task and the feature space of the target task based on the maximum mean difference algorithm, construct the feature space mapping, and build a similar feature space for different assembly tasks;
[0102] The strategy learning and optimization module is configured to: learn and optimize the search strategy based on the constructed similar feature space by combining PID force control and proximal strategy optimization algorithm to obtain the search strategy for the target assembly task.
[0103] Example 3
[0104] The purpose of this embodiment is to provide a computer-readable storage medium.
[0105] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps in the robot shaft hole assembly search method based on similar feature space fusion as described in Embodiment 1 of this disclosure.
[0106] Example 4
[0107] The purpose of this embodiment is to provide an electronic device.
[0108] An electronic device includes a memory, a processor, and a program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the robot shaft hole assembly search method based on similar feature space fusion as described in Embodiment 1 of this disclosure.
[0109] 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 robot shaft hole assembly search method based on similar feature space fusion, characterized in that, include: A feature extraction model was constructed and trained using the collected shaft and hole assembly data; Based on the trained feature extraction model, shaft hole assembly image features and shaft hole assembly position features are extracted from the source assembly task with existing search strategy and the target assembly task without search strategy, respectively, to obtain the source task feature space and the target task feature space. Based on the maximum mean difference algorithm, the mean distance between the feature space of the source task and the feature space of the target task is calculated, a feature space mapping is constructed, and a similar feature space for different assembly tasks is built. Based on the constructed similar feature space, the search strategy is learned and optimized by combining PID force control and proximal strategy optimization algorithm to obtain the search strategy for the target assembly task. Specifically, the feature space mapping is as follows: in, and These represent the number of samples in the source domain and the target domain, respectively. The distance sought is the source task feature space. and target task feature space The maximum mean difference between them, and the source task feature space and the feature space of the target task satisfy: 。 2. The robot shaft hole assembly search method based on similar feature space fusion as described in claim 1, characterized in that, The construction of the shaft hole assembly data is based on a vision-touch fusion shaft hole assembly platform; The shaft hole assembly data includes shaft hole assembly status images captured by the camera, force information collected by the force sensor, and the end effector pose of the robotic arm.
3. The robot shaft hole assembly search method based on similar feature space fusion as described in claim 2, characterized in that, The shaft hole assembly platform includes two cameras in the x and y directions, a force sensor, a robotic arm, and a shaft hole assembly worktable.
4. The robot shaft hole assembly search method based on similar feature space fusion as described in claim 2, characterized in that, The feature extraction model takes the shaft-hole assembly state image, force information, and robotic arm end pose as inputs to generate and output shaft-hole assembly image features and shaft-hole assembly position features.
5. The robot shaft hole assembly search method based on similar feature space fusion as described in claim 1, characterized in that, The feature extraction model specifically includes an image segmentation network and a location feature extraction module; The image segmentation network, built on U-Net, performs image segmentation and feature extraction dimensionality reduction on the input shaft hole assembly state image to obtain shaft hole assembly image features; The position feature extraction module extracts the end pose of the robotic arm and the contact force / torque generated during the shaft hole assembly process as the position features of the shaft hole assembly.
6. The robot shaft hole assembly search method based on similar feature space fusion as described in claim 1, characterized in that, The method of learning and optimizing the search strategy by combining PID force control and proximal strategy optimization algorithm is as follows: Obtain the necessary actions for assembly based on the assembly status of the shaft and hole; Then, the robotic arm is controlled by PID force control to search for the hole in the xy plane, ensuring that the contact force between the shaft and the edge of the hole remains constant during the search. Determine if the current state meets the conditions for a successful / failed search to decide whether to end the search.
7. A robot shaft hole assembly search system based on similar feature space fusion, characterized in that, The robot shaft hole assembly search method based on similar feature space fusion as described in any one of claims 1-6 includes an extraction model construction module, a feature space extraction module, a similar space construction module, and a strategy learning and optimization module; The extraction model construction module is configured to: construct a feature extraction model and train the feature extraction model using the collected shaft hole assembly data; The feature space extraction module is configured to: based on the trained feature extraction model, extract shaft hole assembly image features and shaft hole assembly position features from the source assembly task with existing search strategy and the target assembly task without search strategy, respectively, to obtain the source task feature space and the target task feature space. The similarity space construction module is configured to: calculate the mean distance between the source task feature space and the target task feature space based on the maximum mean difference algorithm, construct a feature space mapping, and build a similar feature space for different assembly tasks; The strategy learning and optimization module is configured to: learn and optimize the search strategy based on the constructed similar feature space by using a combination of PID force control and proximal strategy optimization algorithm to obtain the search strategy for the target assembly task.
8. An electronic device, characterized in that it comprises: Memory is used to store computer-readable instructions in a non-transitory manner. as well as Processor, for executing the computer-readable instructions, When the computer-readable instructions are executed by the processor, they perform the method described in any one of claims 1-6.
9. A storage medium, characterized in that, The computer-readable instructions are stored non-temporarily, wherein when the computer-readable instructions are executed by a computer, the instructions of the method according to any one of claims 1-6 are executed.