3c flexible precision assembly robot skill learning method

Abstract

This invention relates to the field of intelligent manufacturing robot technology and discloses a skill learning method for 3C flexible precision assembly robots, including: constructing a multimodal digital twin simulation platform corresponding to a physical assembly production line and collecting multimodal teaching time-series data; extracting and fusing spatiotemporal features from the multimodal teaching time-series data, predicting the action category frame by frame for the fused features, and outputting standardized assembly action primitives; constructing a knowledge base and generating corresponding executable primitive action sequences in response to target task input by retrieving the knowledge base; constructing reinforcement learning policy networks for each assembly action primitive and training them independently; constructing a cross-domain feature encoding network, training the encoding network after applying random perturbations to the simulated visual image, and mapping the simulated image domain and the real image domain to a unified feature space; and deploying the trained primitive control models and encoding networks to the physical assembly robot to complete the transfer of simulation domain skills to the real production line.

Description

Technical Field

[0001] This invention relates to the field of intelligent manufacturing robot technology, and in particular to a skill learning method for a 3C flexible precision assembly robot. Background Technology

[0002] As 3C electronic products rapidly develop towards miniaturization and flexibility, processes such as FPC flexible cable bonding and precision assembly of mobile phone cameras place extremely high demands on assembly accuracy and operational smoothness. Currently, the industry still heavily relies on manual labor to complete such precision assembly operations.

[0003] Existing robot skill learning solutions have the following main shortcomings: First, teaching and training schemes based on physical robotic arms and real materials rely on repeated grasping and pressing operations, which can easily cause scratches or damage to precision flexible parts, resulting in high material loss costs. Furthermore, when changing products, a large amount of teaching data needs to be collected again, leading to long debugging cycles and low efficiency.

[0004] Second, pre-training schemes based on a single virtual simulation environment generally lack the ability to simulate multi-physics coupling of flexible contact forces, and cannot realistically reproduce the minute force interactions during the assembly process of flexible parts. At the same time, the simulation platform is usually equipped with only a single visual sensor, making it difficult to simultaneously collect multi-dimensional information such as tactile force and voice commands. It also lacks an immersive VR teaching channel, and the collected teaching trajectory deviates significantly from the actual operation.

[0005] Third, existing simulation learning schemes lack a closed-loop chain throughout the entire process, making it impossible to automatically break down long-term continuous assembly processes into standardized basic action units. They also lack structured knowledge graphs to store and reuse assembly operation experience. Reinforcement learning training usually has fixed precision constraints, resulting in slow convergence of high-precision micron-level assembly tasks. Furthermore, there are significant domain differences between simulation-rendered images and real industrial camera images in terms of texture and lighting, which leads to a severe drop in assembly success rate when simulation training strategies are directly deployed to physical robots.

[0006] Therefore, there is an urgent need for a full-process robot skill learning solution that can integrate high-fidelity multimodal digital twin simulation, automatic parsing of long-term tasks, hierarchical reinforcement learning strategy training, and adaptive transfer between simulation and real visual domains, in order to achieve low-cost, high-precision, and highly reusable autonomous learning and rapid production line deployment of 3C flexible precision assembly skills. Summary of the Invention

[0007] The technical problem that this invention aims to solve is the comprehensive technical challenge of high material loss due to reliance on physical teaching during the skill learning process of existing 3C flexible precision assembly robots, lack of flexible contact force modeling in a single simulation environment and difficulty in automatically splitting long-sequence tasks, slow convergence of high-precision assembly strategy training, and visual domain shift when migrating simulation strategies to real production lines.

[0008] To address the aforementioned technical problems, this invention provides a skill learning method for a 3C flexible precision assembly robot, comprising the following steps: Construct a multimodal digital twin simulation platform corresponding to the physical assembly production line. In the simulation platform, assembly teaching is performed through immersive interaction, and multimodal teaching timing data containing at least two modal sensor data is collected. The multimodal teaching time series data is subjected to spatiotemporal feature extraction and fusion processing. The time series action segmentation network is used to predict the action category of the fused features frame by frame, and consecutive frames of the same category are merged into complete action segments to output standardized assembly action primitives. A hierarchical structured skill knowledge base is constructed based on a graph database. In response to the input of a target task, the knowledge base is retrieved and a corresponding executable primitive action sequence is generated. The knowledge base includes a static knowledge layer for storing assembly static attributes and a dynamic knowledge layer for storing the assembly action primitives and their temporal relationships. For each assembly action primitive in the executable primitive action sequence, a reinforcement learning policy network is constructed and trained independently; during the training process, a course learning mechanism is introduced to dynamically adjust the accuracy constraint threshold to obtain the control model for each primitive. A cross-domain feature encoding network is constructed. The encoding network is trained after applying random perturbations to the simulated visual image, and the simulated image domain and the real image domain are mapped to a unified feature space. The trained primitive control models and the encoding network are deployed to a physical assembly robot to complete the transfer of simulation domain skills to the real production line.

[0009] Furthermore, the multimodal teaching timing data is encapsulated into multimodal data tuples in units of frames. Each frame of data contains at least three of the following: visual image data, tactile force data, robotic arm body pose data, voice command data, action label data, and timing label data.

[0010] Furthermore: The multimodal digital twin simulation platform is based on the Unity engine to build a three-dimensional solid model of the part, and coupled with the MuJoCo physics engine to build a flexible soft contact model to simulate the stress behavior of flexible part assembly; The immersive interaction is VR immersive interaction, in which operators perform assembly teaching through VR devices.

[0011] Furthermore, the spatiotemporal feature extraction and fusion processing of the multimodal teaching time-series data includes: Based on single-frame image data, an object relationship graph network is constructed with each part in the scene as graph nodes and the relative poses between parts in three-dimensional space as edge attributes to extract spatial features. Temporal event features are extracted by using a sliding window to capture temporal sensing data. The spatial features and the temporal event features are then concatenated and fused and input into the temporal action segmentation network.

[0012] Furthermore: The object relationship graph network is a 6D pose object relationship graph network; The temporal action segmentation network is a multi-level temporal convolutional network; The standardized assembly action primitives include at least two of the following: idle, move to target, pick up, place, press, and return to position.

[0013] Furthermore: The nodes and relationships of the hierarchical structured skills knowledge base are stored in the Neo4j graph database; The static knowledge layer includes at least one of a part attribute layer, an actuator attribute layer, and a device attribute layer, and is used to store at least one of the part geometric dimensions, robot arm kinematic parameters, and end effector fixed parameters. The dynamic knowledge layer includes at least a task layer and a skill layer, used to store at least one of the following: action primitive type, primitive temporal dependency, assembly target pose, and allowed assembly error threshold.

[0014] Furthermore, the reinforcement learning policy network is a residual reinforcement learning policy network, and its policy output is expressed as:

[0015] In the formula, This is a priori control strategy used to output coarse positioning control commands; This is a residual compensation strategy used to output the contact error compensation amount for the coarse positioning control command; The prior control strategy is a proportional control strategy based on visual pose estimation; the residual compensation strategy is constructed based on the maximum entropy reinforcement learning algorithm, with the goal of maximizing the expected long-term cumulative reward, and adopts a gradient descent iterative update strategy for network weights.

[0016] Furthermore, the training process incorporates a course learning mechanism to dynamically adjust the accuracy constraint threshold, including: In the initial stage of training, a first error threshold is set, and as the number of iterations increases, the allowable error threshold is adjusted from the first error threshold to the target accuracy threshold.

[0017] Furthermore, the independent training sets a reward function for each type of action primitive. The reward function assigns a positive reward value when the target pose is reached, a penalty value when a collision occurs or the number of single primitive interaction steps exceeds a preset threshold, and a reward value associated with the error between the current end position and the target position in other states.

[0018] Furthermore: The cross-domain feature encoding network is an encoder-decoder structure; The random perturbation includes applying brightness perturbation, contrast perturbation, texture perturbation and / or local occlusion perturbation to the simulated visual image; The coding network is trained using frequency domain mean square error as an auxiliary loss for image reconstruction.

[0019] Furthermore, after the encoding network is trained, the physical assembly robot acquires real images through real vision sensors, maps them to the unified feature space through the encoding network, and then the executable primitive action sequence is executed sequentially by each primitive control model.

[0020] Furthermore, the simulation platform also integrates a physics simulation engine; The physical simulation engine transforms the contact force simulation of flexible parts into a convex optimization problem, and uses numerical inverse kinematics algorithm to solve the rotation angles of each joint of the virtual robotic arm; The VR device establishes a mapping relationship between the displacement and rotation angle of the control handle and the movement of the virtual robotic arm end effector, so as to collect teaching actions including grasping, moving, placing, pressing, and returning processes.

[0021] Furthermore, before extracting spatiotemporal features from the multimodal teaching time-series data, a preprocessing operation is performed, which includes: Standardization processing is performed on each modal sensing data to eliminate dimensional differences between different sensing data; Linear scaling is applied to the time series labels for time series alignment.

[0022] Compared with the prior art, the present invention has at least the following beneficial effects: On the one hand, by constructing a multimodal digital twin simulation platform that maps to the physical production line and using immersive interaction to complete the entire process of teaching and data collection, all training is moved to the virtual environment in advance, which fundamentally avoids the repeated damage and equipment collision risks caused by physical teaching to precision flexible materials and greatly shortens the production changeover and debugging cycle.

[0023] On the other hand, by performing spatiotemporal feature fusion and temporal action segmentation on multimodal teaching data such as vision and force perception, the long-term continuous assembly process is automatically decomposed into standardized action primitives. Combined with the structured storage and retrieval of assembly static attributes and primitive temporal associations by a hierarchical graph knowledge base, the rapid generation and cross-product reuse of task-level primitive sequences are realized. Furthermore, a reinforcement learning strategy network is independently constructed for each primitive and a course learning mechanism is introduced to dynamically adjust the accuracy threshold, which effectively solves the problem of slow training convergence of micron-level high-precision assembly tasks under fixed accuracy constraints.

[0024] Finally, by using a cross-domain feature encoding network with randomized augmentation training in the image domain, the simulation and real images are mapped to a unified feature space, eliminating the difference between virtual and real visual distribution. This allows the control model trained in simulation to be directly deployed to the physical production line without additional fine-tuning, thus forming a low-cost, highly reusable, high-precision assembly robot skill autonomous learning closed loop with simulation-real cross-domain transfer capability. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0026] Figure 1 This is a schematic diagram of the overall process disclosed in this invention.

[0027] Figure 2 This is a schematic diagram of the overall architecture of the Unity+MuJoCo coupled multimodal digital twin simulation platform disclosed in this invention.

[0028] Figure 3 This is a schematic diagram of the mapping relationship between the VR human-machine controller and the virtual robotic arm disclosed in this invention.

[0029] Figure 4 This is a diagram of the spatiotemporal event fusion multimodal skill analysis network structure disclosed in this invention.

[0030] Figure 5 This is a topological diagram of the five-layer hierarchical skill knowledge base of Neo4j disclosed in this invention.

[0031] Figure 6 This is a schematic diagram of the residual SAC+ dynamic error course learning primitive strategy training module disclosed in this invention.

[0032] Figure 7 This is a diagram of the encoder-decoder image domain randomized virtual-real migration network structure disclosed in this invention.

[0033] Figure 8 This is a line graph illustrating the ablation experiment results disclosed in this invention.

[0034] Figure 9 This is a bar chart comparing the accuracy of multimodal skill analysis disclosed in this invention. Detailed Implementation

[0035] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0036] This invention aims to provide a skill learning method for 3C flexible precision assembly robots based on multimodal digital twins. This method can be applied to 3C flexible precision assembly scenarios, such as FPC flexible cable assembly, mobile phone camera module bonding assembly, SIM card slot assembly, and coaxial cable assembly. The method of this invention will be described in detail below through specific embodiments and with reference to the accompanying drawings.

[0037] Please see Figure 1 The method comprises five core processing stages. The first phase is the construction of a multimodal digital twin simulation platform and the VR teaching data acquisition phase. This phase is in... Figure 1 The corresponding part S1 specifically includes building a multimodal digital twin simulation environment that maps one-to-one with the physical assembly production line, and collecting multimodal teaching timing data through VR immersive interactive devices.

[0038] The second stage is the multimodal fusion skill analysis stage, corresponding to... Figure 1 In S2, the spatiotemporal features of the collected multimodal teaching data are extracted and fused, and the long temporal assembly process is automatically broken down into standardized assembly action primitives using a temporal action segmentation network.

[0039] The third stage is the construction of the hierarchical graph knowledge base and the generation of primitive sequences, corresponding to... Figure 1 In S3, a five-layer hierarchical structured skill knowledge base is built based on a graph database. Static assembly attributes and dynamic action primitives are stored separately, and executable primitive action sequences are automatically retrieved and generated in response to the input task objective.

[0040] The fourth stage is the residual reinforcement learning and curriculum learning training stage, corresponding to Figure 1 In S4, a residual reinforcement learning policy network is independently constructed for each type of assembly action primitive. A dynamic error course learning mechanism is introduced for training to obtain the control model for each primitive.

[0041] The fifth stage is the image domain randomization virtual-real migration stage, corresponding to Figure 1 In S5, a cross-domain feature encoding network is constructed. Image domain randomization technology is used to eliminate the visual distribution difference between simulated images and real images, and the trained control model is deployed to the physical assembly robot.

[0042] From collecting teaching data to analyzing skills, generating primitive sequences in the knowledge base, training each primitive strategy, and finally completing virtual-to-real migration deployment, a complete closed-loop process is formed.

[0043] The following section elaborates on each of the above stages with specific implementation steps.

[0044] S1: Build a multimodal digital twin simulation platform and collect VR multimodal teaching timing data.

[0045] Please see Figure 2 The simulation platform adopts a layered architecture, including a physical entity layer, a data acquisition layer, a simulation engine layer, and an interaction layer.

[0046] The physical entity layer includes various physical equipment in the physical assembly production line, such as six-degree-of-freedom robotic arms, end effectors, flexible parts, carrier platforms, and conveying mechanisms.

[0047] The data acquisition layer includes various sensors deployed on the physical production line, such as global top-down cameras, eye-to-hand cameras, six-dimensional force sensors, and joint encoders. These sensors are used to collect the operating status data of the physical production line and transmit the data to the simulation engine layer through standardized communication protocols.

[0048] The simulation engine layer is the core of the platform, employing an architecture that couples the Unity engine with the MuJoCo physics engine. The Unity engine is responsible for constructing the 3D visual scene, including 3D models of parts, robotic arm models, workbench models, lighting environments, and texture rendering, providing high-fidelity visual presentation. The MuJoCo physics engine, as the backend for physical simulation, is responsible for simulating rigid body dynamics, soft body contact mechanics, collision detection, and friction effects, providing high-precision physical interaction simulation. The Unity engine and the MuJoCo physics engine exchange data in real time through a data interface. Unity transmits the pose information of each object in the scene to MuJoCo for physical calculations, while MuJoCo sends the calculated contact forces, deformations, and other information back to Unity for visual updates. The two work together to ensure that the simulation scene remains consistent with the real production line at both the visual and physical levels.

[0049] The interaction layer includes a VR headset and a handheld controller. Operators enter an immersive virtual environment through the VR device to perform assembly teaching operations within the virtual environment. The interaction layer also includes a virtual voice sensor to collect the operator's voice commands during the teaching process.

[0050] The various layers communicate with each other through standardized data interfaces. Specifically, sensor data from the physical entity layer is uploaded to the simulation engine layer via the data acquisition layer for calibrating and updating the digital twin model; operation commands from the interaction layer are processed by the simulation engine layer to drive the virtual robotic arm; simulation data generated by the simulation engine layer can also be transmitted back to the data acquisition layer for recording and storage.

[0051] In this embodiment, when constructing the multimodal digital twin simulation platform, 3D modeling software is first used to collect the dimensional parameters, texture information, and material properties of various parts in the physical assembly line to construct a high-precision 3D solid model. For flexible parts such as FPC cables and mobile phone cameras, in addition to constructing their geometric models, it is also necessary to collect their material elastic modulus, Poisson's ratio, bending stiffness, and other physical parameters for subsequent flexible contact mechanics simulation.

[0052] The completed 3D model is imported into the Unity engine, and the scene is replicated 1:1 according to the actual layout of the physical production line to ensure that the spatial position and posture of each object in the virtual scene are consistent with the physical production line.

[0053] After the Unity scene is built, the MuJoCo physics engine is coupled to construct a flexible soft contact model. The MuJoCo physics engine transforms the contact force simulation of flexible parts into a convex optimization problem. By constraining and optimizing physical quantities such as contact force, friction force, and deformation, it simulates the force behavior and deformation state of flexible parts during the assembly process.

[0054] In some implementations, for the insertion process of the FPC cable, MuJoCo calculates the contact force distribution and bending deformation of the cable body when the male connector contacts the female connector; for the bonding process of the camera module, MuJoCo calculates the contact pressure distribution of the bonding surface and the compressive deformation of the elastomer. This flexible soft contact model can realistically reproduce the minute force interactions of flexible parts during assembly, providing high-fidelity physical simulation data for subsequent force perception and strategy training.

[0055] In the simulation platform, the CCD numerical inverse kinematics algorithm is also used to solve the rotation angles of each joint of the six-DOF virtual robotic arm. The CCD numerical inverse kinematics algorithm optimizes the end effector joint by joint from the end effector towards the base, so that the end effector pose gradually approaches the target pose. This method has high computational efficiency and can handle multiple solutions and singular pose problems.

[0056] In this embodiment, multiple types of virtual sensors are deployed on the simulation platform, including but not limited to: Deploy a global overhead virtual vision camera, positioned above the work scene, to acquire global RGB and depth images of the assembly area, providing an overall perception of the assembly scene's situation.

[0057] A local eye-in-the-hand virtual vision camera is deployed at the end of the robotic arm. This camera is mounted near the end effector of the robotic arm and moves with the robotic arm to acquire detailed images of local areas of the assembly operation, providing high-resolution local visual information.

[0058] A virtual haptic force sensor array is deployed, which is distributed on the contact surface of the end effector to collect multi-dimensional contact force and torque information during the assembly process, including normal force, tangential force and torsional torque. The sampling frequency is set to 1000Hz to ensure accurate capture of rapidly changing contact forces.

[0059] A virtual sensor for voice acquisition is deployed. This sensor is used to collect voice commands issued by operators during the teaching process, such as natural language commands like "start grabbing", "place in target position", and "press and snap". The voice data is digitally recorded at a sampling rate of 16kHz.

[0060] The four types of virtual sensors mentioned above have their own independent data acquisition channels and timestamp recording mechanisms to ensure timing alignment during subsequent data fusion.

[0061] Please see Figure 3 The VR interactive device uses a head-mounted display and a matching six-degrees-of-freedom controller. After wearing the head-mounted display, the operator can gain an immersive first-person perspective in the virtual scene, observing the virtual workbench, virtual robotic arm, virtual parts, and the visual display interface of various sensors. The operator holds the controller, which is equipped with a pose tracking sensor and buttons / triggers.

[0062] The pose tracking sensor detects the handle's position coordinates and rotation angle in three-dimensional space in real time and transmits these motion parameters to the simulation platform via a wireless communication protocol. The simulation platform establishes a mapping relationship between the handle's displacement and rotation angle and the motion of the virtual robotic arm's end effector. The mapping scaling factor is set to 1, meaning that the amount of motion of the handle and the amount of motion of the virtual robotic arm's end effector maintain a 1:1 ratio. In this way, every movement of the operator can be accurately reproduced on the virtual robotic arm, making the movement of the virtual robotic arm highly consistent with the operator's natural hand movements.

[0063] The trigger on the control handle corresponds to the gripping and releasing operations of the virtual robotic arm's end effector. When the operator pulls the trigger, the virtual end effector performs a closing gripping action; when the trigger is released, it performs an opening release action. Through this mapping mechanism, the operator can control the virtual robotic arm to complete the entire assembly operation process in a natural and intuitive way.

[0064] Operators wear VR headsets and hold control handles to perform a complete assembly teaching operation in a virtual environment. The teaching process includes all actions such as grasping, moving, placing, pressing, and returning to position. During the teaching process, the simulation platform synchronously collects data from various virtual sensors frame by frame, encapsulating the multimodal data of each frame into a seven-tuple data structure. The expression of this seven-tuple data structure is: In the formula, V represents visual image data, including RGB and depth images captured by the global top-down camera and the local eye-on camera, with an image resolution of 640×480 pixels and a frame rate of 30fps; F represents tactile contact force array data, including multi-dimensional contact force and torque information, with a dimension of six; J represents the robot arm's pose information, including the angle values ​​of each joint and the six-degree-of-freedom pose data of the end effector; L represents voice command data, which is a digitally encoded voice signal segment; A represents the action label of the current frame, which is automatically recorded by the system as the type of action currently performed by the operator, such as "grab," "move," "place," "press," etc.; T S This represents the system timestamp, used for timing alignment and synchronization between different modalities of data. Each data type in the 7-tuple is stored as an independent data stream and uniformly associated with the same timestamp index.

[0065] After completing a full assembly teaching session, the operator obtains a complete multimodal teaching timing data sequence. This sequence contains the seven-tuple data of each frame in the entire assembly process. The frames are arranged in time stamp order to form a time-aligned multimodal teaching data stream.

[0066] S2: Multimodal teaching data preprocessing and temporal skill analysis.

[0067] This step includes two sub-stages: data preprocessing and multimodal fusion parsing.

[0068] In the data preprocessing stage, data cleaning is first performed on all collected seven-tuple time-series data. Abnormal noise samples caused by sensor communication interruptions or environmental interference are removed. By setting reasonable threshold values ​​for each modality, abnormal frames exceeding the threshold range are marked and removed. Redundant frames with completely identical content between adjacent frames are removed. By calculating the structural similarity index of adjacent frame images and the Euclidean distance of force data, if the difference between two consecutive frames is less than a preset threshold, the latter frame is considered redundant and deleted to reduce data redundancy and storage overhead. After cleaning, the data is divided according to time order to form training, testing, and validation sets. The training set samples account for more than 70% of the total dataset to ensure that the time-series action segmentation network can fully learn the features of various assembly actions. The remaining samples are evenly divided into the testing and validation sets. A hierarchical random sampling strategy is used for data partitioning to ensure that the distribution ratio of various actions in the training and testing sets is consistent with the original dataset.

[0069] Then, standardization is performed on the sensor data for each modality to eliminate the influence of differences in dimensions and magnitudes between different sensor data. The standardization calculation formula is as follows: In the formula, The original sensor data values, This is the sample mean of the modal data. This represents the sample standard deviation of the modal data. These are the standardized values.

[0070] After standardization, the values ​​of each modality data are converted into a standard normal distribution with a mean of 0 and a standard deviation of 1, making the data of different modalities comparable in terms of numerical magnitude, which facilitates subsequent multimodal feature fusion processing.

[0071] For time-series timestamp data, a linear scaling process is performed. This scaling is only used for temporal alignment operations in the temporal convolutional network and does not participate in the training process of the feature extraction network, so as to ensure that the scaling of timestamps will not affect the model's extraction of action features.

[0072] In the multimodal fusion analysis stage, a spatiotemporal event fusion multimodal skill analysis framework is constructed, which includes three parallel sub-modules: spatial feature extraction channel, temporal event feature extraction channel, and fusion classification channel.

[0073] Please see Figure 4 The network structure consists of four main components: a spatial feature extraction branch, a temporal feature extraction branch, a feature fusion layer, and an MS-TCN++ classification layer.

[0074] Specifically: The spatial feature extraction branch takes a single-frame RGB-D image as input and extracts spatial features through a 6D pose object graph network. Figure 4 The left side of the diagram illustrates the specific structure of the 6D pose object relationship graph network: each part instance in the scene is treated as a node in the graph, and the relative pose relationships between parts in 3D space are used as edge attributes between nodes, constructing a complete graph structure. The diagram exemplarily labels nodes 1 to n, and the edge attributes include the relative position offsets (Δx, Δy, Δz) and relative rotations (Δroll, Δpitch, Δyaw) between two parts, i.e., six-degree-of-freedom relative pose information. The graph neural network performs multiple rounds of message passing and feature aggregation on each node. In each iteration, each node collects spatial pose information from its neighboring nodes and updates its own feature representation. After multiple iterations, the feature representation of each node incorporates the spatial layout information of other surrounding parts, ultimately converging all node features into a global spatial feature vector, thus encoding the spatial configuration relationships between parts in the assembly scene.

[0075] The temporal feature extraction branch uses a sliding window to extract tactile force data and voice command data from preceding and following time sequences. Figure 4 The lower-middle section illustrates the temporal capture method of the sliding window. Centered on the current frame t, the window radius d is set to 10 frames, meaning the window length is 2d+1 (21 frames in total), covering temporal data within a range of 10 frames before and after the current frame. For haptic force data, the six-dimensional force / torque values ​​of each frame within the window are extracted to form a haptic force temporal feature sequence. For voice command data, voice command fragments are extracted from the voice signal within the window using a voice activity detection algorithm, and then converted into text command vectors by a speech recognition model. The haptic force temporal features and voice command features are concatenated along the time axis and input into a temporal event feature encoding network to extract temporal event features reflecting the action process and operational intent.

[0076] The feature fusion layer concatenates and fuses the spatial feature vector output from the spatial feature extraction branch with the temporal event feature vector output from the temporal feature extraction branch, forming a fused multimodal feature vector. This fused feature vector simultaneously contains spatial configuration information at the current moment and action progress information from the preceding and following moments, providing rich decision-making basis for subsequent action classification.

[0077] MS-TCN++ classification layer is located at Figure 4 The right side of the diagram contains the core classification module of the network structure. MS-TCN++ is a multi-level temporal convolutional network that employs a pyramid-shaped temporal receptive field design. Figure 4In the diagram, the MS-TCN++ module is expanded to show its internal structure—containing multiple layers, each consisting of several temporal convolutional layers, with the kernel size gradually increasing across different layers. The first layer uses small kernels to capture rapid local action changes, the second and third layers use medium kernels to capture mid-term action patterns, and the last layer uses large kernels to capture global action trends.

[0078] Multimodal fusion features from the feature fusion layer are input into the MS-TCN++ network in chronological order. Each layer performs parallel computation and finally summarizes and outputs the action category prediction results for each frame.

[0079] MS-TCN++ employs a multi-level prediction and progressive refinement strategy. The coarse-grained prediction results output at lower levels are refined and corrected at higher levels by incorporating contextual information from a larger temporal receptive field. This approach maintains sensitivity to local details while improving overall temporal consistency and the accuracy of segmentation boundaries.

[0080] The action category labels predicted frame by frame by MS-TCN++ are arranged in chronological order, and consecutive frames of the same category are merged to form a complete action segment. For example, 50 consecutive frames predicted as "move to target" are merged into a single "move to target" action segment. After merging, the original continuous temporal data is divided into a series of discrete, standardized assembly action primitives.

[0081] In this embodiment, the standardized assembly action primitives are categorized into six types: idle, move to target, pick up, place, press, and return. Idle indicates that the robotic arm is in standby mode with no operation; move to target indicates that the robotic arm's end effector moves from its current position to the target part position; pick up indicates that the end effector performs a gripping or suction operation to pick up the part; place indicates that the picked-up part is moved and placed at the target assembly position; press indicates that pressure is applied to the placed part to complete the fastening or fitting operation; and return indicates that the robotic arm returns to its initial or safe position after completing the assembly operation.

[0082] S3: Neo4j hierarchical graph structure skill knowledge base construction and task retrieval.

[0083] This step involves building a five-layer hierarchical graph structure skill knowledge base based on the Neo4j graph database, organizing and storing various types of assembly-related knowledge according to a hierarchical structure.

[0084] Please see Figure 5The knowledge base employs a five-layer hierarchical architecture, from top to bottom: Part Layer, Actuator Layer, Attribute Layer, Task Layer, and Skill Layer. The diagram illustrates the connections between layers using nodes and relational edges. Each rectangle represents a knowledge node, connecting lines represent semantic relationships between nodes, and labels on the connecting lines indicate the relationship type. The Part Layer, located at the top of the knowledge base, describes the attribute information of various parts involved in the assembly task, including part type nodes, part size nodes, and part material nodes. In this layer, each part type node stores corresponding geometric parameters, material properties, and functional descriptions. The Actuator Layer, located below the Part Layer, describes the attribute information of various execution devices used in the assembly process, including robotic arm nodes, end effector nodes, and gripper nodes. It stores information such as the kinematic parameters of the robotic arm, the type and specifications of the end effector, and the travel range and clamping force range of the gripper. The Attribute Layer, located below the Actuator Layer, stores various static configuration parameters, including the matching relationship between parts and actuators, the spatial coordinates of the assembly station, and the calibration parameters of various sensors. The component layer, actuator layer, and attribute layer together constitute the static knowledge layer. This knowledge is fixed after the production line is built and remains unchanged during product changeover and strategy retraining, without the need for repeated input.

[0085] The task layer, located below the attribute layer, stores dynamic information related to assembly tasks, including description nodes, task decomposition nodes, and task execution order nodes for various assembly tasks. The skill layer, located at the bottom of the knowledge base, stores dynamic information such as various standardized assembly action primitives parsed in step S2, temporal dependencies between primitives, 3D coordinates of the assembly target corresponding to each primitive, and allowable assembly error thresholds corresponding to each primitive.

[0086] The task layer and the skill layer together constitute the dynamic knowledge layer. The knowledge stored in this layer can be dynamically updated and expanded according to the task requirements each time a new assembly task is executed. When the product model is changed, only the content of the dynamic knowledge layer needs to be updated to reuse the original static knowledge layer.

[0087] When building the knowledge base, Neo4j's node creation and relationship creation statements are used to write the aforementioned types of knowledge nodes and their interrelationships into the graph database. Each knowledge node contains a node identifier, a node type label, and node attribute key-value pairs. Relationship edges contain relationship type labels and relationship attribute key-value pairs; the relationship attributes are used to describe the confidence weight and association strength between nodes.

[0088] In some specific implementations, when a target 3C assembly task requirement is received, such as "complete FPC cable assembly" or "complete mobile phone front camera bonding," the system first performs semantic parsing on the task requirement, extracting the task type identifier and target part identifier. Then, using the task type identifier and target part identifier as query conditions, it traverses the task layer nodes in the Neo4j knowledge graph, retrieving task description nodes that match the task. After finding a matching task node, it retrieves each skill node contained in the task, i.e., each assembly action primitive node, along the association edges between the task node and skill layer nodes. During the retrieval process, it simultaneously obtains the assembly target coordinate parameters and allowable assembly error threshold parameters associated with each skill node. All retrieved skill nodes are topologically sorted according to the direction of the temporal association edges, generating an ordered primitive action sequence. This primitive action sequence contains all the assembly action primitives required to complete the target task and their execution order, and each primitive is bound to the corresponding assembly target coordinates and precision threshold parameters. For example, for an FPC cable assembly task, the generated primitive action sequence might be: move to target 1 → pick up → move to target 2 → place → press → return to position. Each action primitive in the sequence comes with an independent precision constraint parameter. Different primitives are allowed to set different precision requirements. The allowable error for the pick-up action is ±2mm, and the allowable error for the press-up action is ±0.05mm.

[0089] S4: Residual reinforcement learning and dynamic error curriculum learning training primitive control strategy.

[0090] In this embodiment, step S4 involves independently constructing and training a reinforcement learning policy network for each type of action primitive in the executable primitive action sequence output by the knowledge base in step S3.

[0091] Please see Figure 6 The training module comprises a unified training framework that is independently instantiated for each class of action primitives. The framework consists of four main components: a policy network module, an environment interaction module, a course learning scheduling module, and a reward calculation module.

[0092] Specifically: The policy network module includes a priori control policy branches. and residual compensation strategy branch The two sub-modules operate in parallel, and their outputs are summed at the output of the policy network module to form the final policy output. .

[0093] Among them, the prior control strategy branch adopts a proportional control structure, which calculates coarse positioning control commands based on the 6D pose information of the target part output by the visual pose estimation algorithm HS-Pose and the current pose of the robotic arm end effector.

[0094] The residual compensation strategy branch adopts a flexible actor-judge reinforcement learning algorithm structure based on maximum entropy, comprising two sub-networks: an actor network and a judge network. The actor network is responsible for outputting residual compensation actions based on the current state, while the judge network is responsible for evaluating the value of the current state. Both sub-networks use a multi-layer fully connected neural network structure, including an input layer, two hidden layers, and an output layer. The number of hidden layer neurons is 256 and 128, respectively, and the ReLU activation function is used.

[0095] The environment interaction module is responsible for executing the action commands output by the policy network and returning the new state and reward value after execution. During training, the environment interaction module communicates with the MuJoCo physics simulation engine, sending action commands to the simulation engine to drive the virtual robotic arm's movement, and obtaining state updates and physical feedback data after movement from the simulation engine.

[0096] The course learning scheduling module is responsible for dynamically adjusting the accuracy constraint threshold of the training task according to the preset course learning strategy. This module includes an error threshold scheduler, which calculates the current allowable error threshold based on the current number of training iterations.

[0097] The reward calculation module is responsible for calculating the instant reward value based on the status returned by the environment interaction module.

[0098] In this embodiment, regarding policy decomposition, the output expression of the residual reinforcement learning policy network is as follows:

[0099] In the formula, This is a priori control strategy used to output coarse positioning control commands. Based on the coarse positioning results output by the HS-Pose target 6D pose estimation algorithm, the priori control strategy generates proportional prior control commands to complete the coarse positioning stage of the assembly action. Specifically, the HS-Pose target 6D pose estimation algorithm detects key points and regresses the pose of the target part in the visual image, outputting a six-degree-of-freedom pose estimate of the target part in the robot arm base coordinate system, including three-dimensional spatial position coordinates (x, y, z) and three-dimensional pose angles (roll, pitch, yaw).

[0100] This is a maximum entropy residual compensation strategy used to output contact error compensation for coarse positioning control commands. Based on the error state between the current robotic arm end-effector pose and the target pose, this strategy outputs micron-level control compensation values ​​to compensate for minute contact errors and assembly gaps that cannot be eliminated by prior control strategies. The residual compensation strategy is trained using a maximum entropy reinforcement learning algorithm framework. This framework maximizes the policy's entropy value while optimizing the cumulative reward, encouraging the policy to maintain sufficient exploratory behavior during training and avoiding premature convergence to local optima. The expression for the maximum entropy cumulative reward objective function is:

[0101] In the formula, Let t be the state vector at step t, which contains information such as the robot arm end pose, joint angles, and contact forces. The action vector executed at step t includes the end-effector displacement increment and attitude adjustment. In the state Execute action The instant reward value obtained afterward; For the strategy in the state The entropy value is used to measure the uncertainty of the action distribution of the policy in this state; This is the entropy weighting coefficient, used to regulate the balance between maximizing cumulative rewards and the degree of strategy exploration. The larger the value, the more exploratory the strategy. The smaller the value, the more the strategy tends to utilize existing experience; Indicating in strategy The mathematical expectation on the sampling trajectory.

[0102] In this embodiment, the custom assembly reward function is designed as follows: When the end effector of the robotic arm reaches the assembly target position and the pose error is within the allowable range.

[0103] When the robotic arm collides or the number of single primitive interaction steps exceeds 200, the following occurs.

[0104] In other states. The coordinates of the current position of the robotic arm's end effector in three-dimensional space. The three-dimensional spatial coordinates of the target assembly position. This represents the Euclidean distance between the current endpoint position and the target position.

[0105] A positive reward of 100 is awarded upon reaching the target position to incentivize the policy network to learn the behavior of accurately reaching the target. A penalty of -30 is awarded when a collision or timeout occurs to constrain the policy network to avoid unsafe behavior. In other states, the reward value is the negative of the Euclidean distance between the current endpoint and the target position; that is, the closer the distance, the higher the reward, and the farther the distance, the lower the reward, thus guiding the policy network to gradually approach the target position. During training, the optimization objective is to maximize the long-term cumulative reward expectation, and the gradient descent method is used to iteratively update the weight parameters of the residual policy network.

[0106] Regarding the course learning mechanism, the allowable distance error is dynamically adjusted according to the number of training iterations. It satisfies the expression:

[0107] In the formula, This represents the current training iteration step. This is the preset total number of iterations for the course learning. Set to 5000 steps; The initial allowable position error threshold is set to ±5mm. This relatively large error threshold makes the assembly task in the early stage of training relatively simple, and the policy network can more easily obtain positive rewards, thus accelerating the learning progress in the early stage of training. The final target accuracy threshold is set at ±0.05mm according to the industrial requirements of 3C precision assembly. This accuracy threshold corresponds to the micron-level precision assembly requirements.

[0108] In the early stages of training much smaller hour, near It has a large error tolerance, and the policy network can learn basic assembly trajectories and action patterns under relatively relaxed conditions. With the increase in iteration steps... Gradually increase In a linear manner from Towards The transition involves a gradual increase in precision requirements and a gradual increase in task difficulty. Exceed back, Keep as Without changing the target accuracy, the policy network is finely optimized. This progressive learning strategy, from easy to difficult, avoids the convergence difficulties caused by directly facing high-precision constraints in the early stages of training, effectively accelerating the overall convergence speed of the training process.

[0109] For each type of action primitive (move to target, pick up, place, press, return to position, etc.) output by the knowledge base, the above training process is executed independently to train a dedicated residual reinforcement learning policy network for each type of primitive.

[0110] Each primitive policy network is independent of the others and does not share parameters. Each network is specifically optimized for its corresponding action primitive type.

[0111] After training all primitive policy networks, the model parameters of each primitive policy network are saved to form a complete primitive control model library.

[0112] S5: Image Domain Randomized Encoder-Decoder Network Training and Virtual-Real Policy Deployment.

[0113] In this embodiment, step S5 constructs a cross-domain feature coding network and eliminates the visual distribution difference between simulated images and real images through image domain randomization technology, thereby realizing the transfer and deployment of simulation training strategies to physical robots.

[0114] Please see Figure 7 The cross-domain feature encoding network adopts an encoder-decoder symmetric structure. Figure 7 The left side is the encoder module, the middle is the bottleneck feature layer, and the right side is the decoder module.

[0115] The encoder module consists of multiple stacked convolutional layers. Figure 7 The diagram illustrates the sequential structure of convolutional layers 1 through n, each followed by a batch normalization layer and a ReLU activation function. The encoder module takes an RGB simulated image (224×224×3) as input from the simulation platform. After layer-by-layer convolution and downsampling, a low-dimensional latent space feature vector is output at the bottleneck feature layer. This feature vector has a dimension of 256. The bottleneck feature layer is the core of the network; it maps the high-dimensional image space to a low-dimensional feature space, and this mapping is invariant to domain differences. That is, whether it's a simulated image or a real image, as long as the semantic content is the same, their feature representations in the latent space should be similar.

[0116] The decoder module consists of multiple deconvolutional layers stacked together. Figure 7 The diagram illustrates the sequential structure of deconvolutional layers 1 through m, with each deconvolutional layer followed by a batch normalization layer and a ReLU activation function. The decoder module takes the latent space feature vector from the bottleneck feature layer as input and reconstructs an RGB image of the same size as the original input through layer-by-layer deconvolution and upsampling operations.

[0117] During the training phase, random perturbations are applied to the original RGB simulation images output by the simulation platform to augment the samples.

[0118] Random perturbations include the following types: Brightness perturbation is achieved by superimposing a random gain coefficient uniformly distributed within the interval [-0.3, 0.3] onto the original image brightness value.

[0119] Contrast perturbation is achieved by multiplying the original image contrast value by a random scaling factor that is uniformly distributed within the range of [0.6, 1.4].

[0120] Texture perturbation applies a Gaussian blur of random intensity to the original image, with the standard deviation of the Gaussian kernel randomly sampled within the range of [0.5, 2.0] pixels.

[0121] Local occlusion perturbation involves randomly selecting a rectangular region in the original image and setting the pixel values ​​of that region to random or zero values. The area of ​​the rectangular region varies randomly between 5% and 20% of the total image area.

[0122] By applying a combination of the above-mentioned random perturbations, a massive number of image augmentation samples are generated, enabling the coding network to be exposed to simulated images under various visual change conditions during the training process.

[0123] The encoder-decoder network is trained using the frequency domain mean square error as an auxiliary loss for image reconstruction. The expression for the frequency domain mean square error loss function is:

[0124] In the formula, The batch size is set to 32 to represent the number of image samples in the training batch. For the first The original, unperturbed version of the image, i.e., the original image directly output by the simulation platform, is used as the ground truth for reconstruction; I_Ri is the nth image. The image is a perturbation version after applying random perturbation; For encoder-decoder networks to handle perturbed images The reconstruction prediction output; This represents all trainable parameters of the encoder-decoder network, including the kernel weights and biases of each convolutional layer in the encoder and the kernel weights and biases of each deconvolutional layer in the decoder.

[0125] loss function The network's ability to reconstruct perturbed images was measured. When the network can accurately reconstruct the original image from a severely perturbed image, it indicates that the encoder has learned the core features related to the semantic content of the image and has invariance to domain variations such as brightness, contrast, texture, and occlusion.

[0126] In this embodiment, the Adam optimizer is used to optimize the loss function using gradient descent, with the learning rate set to 0.0001 and the number of training rounds set to 100, until the loss function value converges.

[0127] After training, the encoder module is extracted and used independently. This encoder module has the ability to map any input image to a unified latent feature space. Whether it is a simulated image output from a simulation platform or a real image captured by a real camera on a physical production line, after being mapped by this encoder, they are all distributed in the same 256-dimensional latent feature space, and the distribution differences are significantly reduced. This mapping relationship achieves semantic alignment between the simulated image domain and the real image domain, eliminating visual domain offsets such as texture, lighting, and color.

[0128] During the deployment phase, all trained primitive control policy network models and encoder modules are distributed to the controller of the physical assembly robot. The industrial control computer of the physical assembly robot is equipped with inference acceleration hardware, enabling real-time forward inference computation of the encoder module and policy network. The physical assembly robot reads real visual images captured by the RealSense depth vision camera deployed on-site, and inputs the captured RGB images into the encoder module in real time for feature mapping, obtaining the feature representation of the real image in the latent space. This feature representation is in the same feature space as the feature representation output by the encoder during simulation training; therefore, the subsequent primitive control policy network can directly use this feature representation as input without any additional domain adaptation processing. Then, the executable primitive action sequences output by the knowledge base are executed sequentially. After the previous primitive is executed and successfully determined, the next primitive in the sequence is automatically entered until all primitives are executed, completing the entire assembly task. Throughout the execution process, the inference output of each primitive control policy network continuously generates control commands at a frequency of 30Hz, driving the robotic arm to adjust its motion trajectory in real time, achieving high-precision automated precision assembly.

[0129] To verify the effectiveness of the technical solution of the present invention, relevant experiments were conducted.

[0130] Please see Figure 8 The horizontal axis represents different experimental configuration groups, and the vertical axis represents the assembly success rate under the corresponding configuration. Figure 8 The comparison results of six experimental configurations are shown below: The first group was trained using only visual monomodal data (with tactile and speech modalities removed), corresponding to the first data point in the line graph, with a success rate of 62.3%.

[0131] The second group uses multimodal data but does not perform temporal action segmentation (the complete temporal sequence is directly input into the policy network), corresponding to the second data point, with a success rate of 68.7%.

[0132] The third group, which uses multimodal data and temporal segmentation but does not build a knowledge base (the action sequence is regenerated each time), corresponds to the third data point and has a success rate of 73.1%.

[0133] The fourth group used the first three methods but did not employ residual reinforcement learning (only standard reinforcement learning), corresponding to the fourth data point, with a success rate of 78.5%.

[0134] The fifth group used the first four methods but did not employ a course learning mechanism (fixed precision constraints throughout), corresponding to the fifth data point, with a success rate of 74.8%.

[0135] The sixth group represents the complete solution of this invention, containing all five core modules, corresponding to the sixth data point, with a success rate of 90.7%.

[0136] from Figure 8 It is clear that as each core module is added sequentially, the assembly success rate increases in a stepwise manner. Multimodal fusion brings an improvement of approximately 6.4 percentage points, temporal action segmentation brings an improvement of approximately 4.4 percentage points, knowledge base reuse brings an improvement of approximately 5.4 percentage points, residual reinforcement learning brings an improvement of approximately 5.4 percentage points, and the curriculum learning mechanism brings an improvement of approximately 12.2 percentage points. These data indicate that the five core modules of this invention have a synergistic effect; the combined use of these modules can significantly improve the overall assembly success rate, while the absence of any single module will lead to a significant performance decrease.

[0137] Please see Figure 9 The horizontal axis represents different action categories, including six standardized action primitives: idle, move to target, pick up, place, press, and return to position, as well as the overall average accuracy of the six action categories; the vertical axis represents the recognition accuracy of each action category, expressed as a percentage. Figure 9 For each action category, three bar charts are displayed side by side: the first bar represents the recognition accuracy based solely on the visual modality, the second bar represents the recognition accuracy based on the visual plus force bimodality, and the third bar represents the recognition accuracy based on the visual, force, and speech trimodality (i.e., the complete solution of this invention).

[0138] from Figure 9The data shows that for all six types of action primitives, the trimodal complete solution has significantly higher recognition accuracy than the monomodal and bimodal solutions. Specifically, the trimodal accuracy is 98.2% for idle actions, 96.7% for moving to a target, 95.9% for picking up, 94.3% for placing, 97.1% for pressing, and 96.5% for returning to position, with an overall average accuracy of 96.57%. In contrast, the average accuracy of the monomodal solution is only 83.22%, and the average accuracy of the bimodal solution is 90.15%. This comparison clearly demonstrates that the visual modality provides spatial configuration information, the force modality provides contact state information, and the speech modality provides operational intent information. The information from these three modalities complements each other, and their fusion significantly improves the accuracy of action recognition and segmentation. This is especially true for actions like pressing, which are sensitive to changes in contact force; the addition of the force modality brings the most significant accuracy improvement.

[0139] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A skill learning method for a 3C flexible precision assembly robot, characterized in that, Includes the following steps: Construct a multimodal digital twin simulation platform corresponding to the physical assembly production line. In the simulation platform, assembly teaching is performed through immersive interaction, and multimodal teaching timing data containing at least two modal sensor data is collected. The multimodal teaching time series data is subjected to spatiotemporal feature extraction and fusion processing. The time series action segmentation network is used to predict the action category of the fused features frame by frame, and consecutive frames of the same category are merged into complete action segments to output standardized assembly action primitives. A hierarchical structured skill knowledge base is constructed based on a graph database. In response to the input of a target task, the knowledge base is retrieved and a corresponding executable primitive action sequence is generated. The knowledge base includes a static knowledge layer for storing assembly static attributes and a dynamic knowledge layer for storing the assembly action primitives and their temporal relationships. For each assembly action primitive in the executable primitive action sequence, a reinforcement learning policy network is constructed and trained independently; during the training process, a course learning mechanism is introduced to dynamically adjust the accuracy constraint threshold to obtain the control model for each primitive. A cross-domain feature encoding network is constructed. The encoding network is trained after applying random perturbations to the simulated visual image, and the simulated image domain and the real image domain are mapped to a unified feature space. The trained primitive control models and the encoding network are deployed to a physical assembly robot to complete the transfer of simulation domain skills to the real production line.

2. The skill learning method for 3C flexible precision assembly robots according to claim 1, characterized in that, The multimodal teaching timing data is encapsulated into multimodal data tuples in units of frames. Each frame of data contains at least three of the following: visual image data, tactile force data, robotic arm body pose data, voice command data, action label data, and timing label data.

3. The skill learning method for 3C flexible precision assembly robots according to claim 1, characterized in that: The multimodal digital twin simulation platform is based on the Unity engine to build a three-dimensional solid model of the part, and coupled with the MuJoCo physics engine to build a flexible soft contact model to simulate the stress behavior of flexible part assembly; The immersive interaction is VR immersive interaction, in which operators perform assembly teaching through VR devices.

4. The skill learning method for 3C flexible precision assembly robots according to claim 1, characterized in that, The spatiotemporal feature extraction and fusion processing of the multimodal teaching time-series data includes: Based on single-frame image data, an object relationship graph network is constructed with each part in the scene as graph nodes and the relative poses between parts in three-dimensional space as edge attributes to extract spatial features. Temporal event features are extracted by using a sliding window to capture temporal sensing data. The spatial features and the temporal event features are then concatenated and fused and input into the temporal action segmentation network.

5. The skill learning method for 3C flexible precision assembly robots according to claim 4, characterized in that: The object relationship graph network is a 6D pose object relationship graph network; The temporal action segmentation network is a multi-level temporal convolutional network; The standardized assembly action primitives include at least two of the following: idle, move to target, pick up, place, press, and return to position.

6. The skill learning method for 3C flexible precision assembly robots according to claim 1, characterized in that: The nodes and relationships of the hierarchical structured skills knowledge base are stored in the Neo4j graph database; The static knowledge layer includes at least one of a part attribute layer, an actuator attribute layer, and a device attribute layer, and is used to store at least one of the part geometric dimensions, robot arm kinematic parameters, and end effector fixed parameters. The dynamic knowledge layer includes at least a task layer and a skill layer, used to store at least one of the following: action primitive type, primitive temporal dependency, assembly target pose, and allowed assembly error threshold.

7. The skill learning method for 3C flexible precision assembly robots according to claim 1, characterized in that, The reinforcement learning policy network is a residual reinforcement learning policy network, and its policy output is expressed as follows: In the formula, This is a priori control strategy used to output coarse positioning control commands; This is a residual compensation strategy used to output the contact error compensation amount for the coarse positioning control command; The prior control strategy is a proportional control strategy based on visual pose estimation; the residual compensation strategy is constructed based on the maximum entropy reinforcement learning algorithm, with the goal of maximizing the expected long-term cumulative reward, and adopts a gradient descent iterative update strategy for network weights.

8. The skill learning method for a 3C flexible precision assembly robot according to claim 1, characterized in that, The training process incorporates a course learning mechanism to dynamically adjust the accuracy constraint threshold, including: In the initial stage of training, a first error threshold is set, and as the number of iterations increases, the allowable error threshold is adjusted from the first error threshold to the target accuracy threshold.

9. The skill learning method for a 3C flexible precision assembly robot according to claim 1, characterized in that, The independent training sets a reward function for each type of action primitive. The reward function assigns a positive reward value when the target pose is reached, a penalty value when a collision occurs or the number of single primitive interaction steps exceeds a preset threshold, and a reward value associated with the error between the current end position and the target position in other states.

10. The skill learning method for a 3C flexible precision assembly robot according to claim 1, characterized in that: The cross-domain feature encoding network is an encoder-decoder structure; The random perturbation includes applying brightness perturbation, contrast perturbation, texture perturbation and / or local occlusion perturbation to the simulated visual image; The coding network is trained using frequency domain mean square error as an auxiliary loss for image reconstruction.

11. The skill learning method for a 3C flexible precision assembly robot according to claim 1, characterized in that, After the encoding network is trained, the physical assembly robot acquires real images through real vision sensors, maps them to the unified feature space through the encoding network, and then the executable primitive action sequence is executed sequentially by each primitive control model.

12. The skill learning method for a 3C flexible precision assembly robot according to claim 3, characterized in that, The simulation platform also integrates a physics simulation engine; The physical simulation engine transforms the contact force simulation of flexible parts into a convex optimization problem, and uses numerical inverse kinematics algorithm to solve the rotation angles of each joint of the virtual robotic arm; The VR device establishes a mapping relationship between the displacement and rotation angle of the control handle and the movement of the virtual robotic arm end effector, so as to collect teaching actions including grasping, moving, placing, pressing, and returning processes.

13. The skill learning method for a 3C flexible precision assembly robot according to claim 1, characterized in that, Before extracting spatiotemporal features from the multimodal teaching time series data, a preprocessing operation is performed, which includes: Standardization processing is performed on each modal sensing data to eliminate dimensional differences between different sensing data; Linear scaling is applied to the time series labels for time series alignment.

Images