Electric tool anthropomorphized automatic test method based on six-dimensional force sensor

Abstract

The application relates to a kind of electric tool anthropomorphism automatic test method based on six-dimensional force sensor, method includes: obtaining the multimodal perception data in the process of mechanical arm clamping electric tool test;Based on the topographic features of target test surface are extracted from multimodal perception data, and the physical semantic graph of test scene is constructed in combination with preset physical rules;Physical semantic graph is converted into physical semantic feature vector, and physical semantic feature vector is input into pre-trained strategy model, and the action trajectory segment of mechanical arm is generated;Action trajectory segment is sent to force-position hybrid controller, and force closed loop correction is carried out based on force sensation data through force-position hybrid controller, and mechanical arm clamping electric tool is driven to execute automatic test action.The application has the advantages that the flexible operation ability of optimization automatic test lacks anthropomorphism, which causes the test condition to deviate from the real use working condition problem.

Description

Technical Field

[0001] This application relates to the technical field of industrial automation testing, and in particular to an anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor. Background Technology

[0002] Power tools such as electric hammers must undergo rigorous durability testing during the research and development phase to verify their lifespan and failure modes in real-world usage scenarios. This testing typically requires the tool to continuously drill into media such as concrete for hundreds or even thousands of hours.

[0003] Existing testing solutions mainly include manual testing and traditional non-standard automated testing. Manual testing, due to the high noise and dust generated at the testing site, poses a serious health risk to operators with prolonged exposure; furthermore, individual differences among operators lead to poor test consistency, long cycles, and high safety risks. To circumvent the shortcomings of manual testing, some companies use pre-programmed non-standard automated clamping equipment. However, such solutions lack flexibility because the equipment operates along a pre-planned fixed trajectory, failing to perceive and respond to the actual state of the test surface. Especially when encountering obstacles (such as reinforcing bars within concrete), the equipment is still forced to advance along a fixed path, causing the tool to be subjected to undue additional lateral forces and bending moments, resulting in premature tool failure, severely distorted test results, and difficulty in adapting to the differences in test media across different batches. Summary of the Invention

[0004] This application aims to provide an anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, in order to solve the technical problems of existing automated testing equipment lacking anthropomorphic flexible operation capabilities and easily leading to additional loads and failure mode distortion.

[0005] On the one hand, this application provides a human-like automatic testing method for power tools based on a six-dimensional force sensor, which adopts the following technical solution:

[0006] An anthropomorphic automated testing method for power tools based on a six-dimensional force sensor, comprising:

[0007] Multimodal perception data is acquired during the testing process of a robotic arm gripping a power tool. The multimodal perception data includes force data, depth image data, and pose data.

[0008] Based on the multimodal perception data, the terrain features of the target test surface are extracted, and a physical semantic map of the test scene is constructed in combination with preset physical rules.

[0009] The physical semantic graph is converted into a physical semantic feature vector, and the physical semantic feature vector is input into a pre-trained policy model to generate the motion trajectory fragment of the robotic arm; wherein, the policy model performs consistency correction on the motion trajectory fragment based on physical constraints during the training and generation process;

[0010] The motion trajectory segment is sent to the force-position hybrid controller, which performs force closed-loop correction based on the force data, and drives the robotic arm to clamp the power tool to perform automated testing actions.

[0011] In one embodiment: the physical semantic graph includes a set of nodes and a set of edges; the step of extracting terrain features of the target test surface based on the multimodal perception data and constructing a physical semantic graph of the test scene in combination with preset physical rules specifically includes:

[0012] The depth image data is mapped to the terrain features using a point cloud encoder;

[0013] Based on the pose data, force data, and terrain features, the node set is constructed, which includes tool chisel head nodes, test surface nodes, obstacle nodes, and robotic arm end effector nodes.

[0014] The edge set is constructed based on the contact force transmission relationship, kinematic constraint relationship and collision constraint relationship between the nodes in the node set.

[0015] In one embodiment: the step of converting the physical semantic graph into a physical semantic feature vector specifically includes:

[0016] The state vectors of each node in the node set are concatenated into a node feature sequence in a preset order, wherein the state vectors include at least one of the corresponding node's position, velocity, force, or curvature.

[0017] The attribute vectors of each edge in the edge set are concatenated into an edge feature sequence in a preset order, wherein the attribute vectors include at least one of the contact force, fixed offset or distance of the corresponding edge;

[0018] The node feature sequence and the edge feature sequence are concatenated to form the physical semantic feature vector, and the input of the policy model includes at least the physical semantic feature vector.

[0019] In one embodiment: the step of the policy model performing consistency correction on the motion trajectory segment based on physical constraints during training and generation specifically includes:

[0020] During the generation of the motion trajectory segment by the strategy model, physical constraint checks are performed in real time; the physical constraint checks include force-motion consistency constraints, energy conservation constraints, and contact stability constraints.

[0021] When the motion trajectory segment is detected to violate any of the physical constraint checks, the motion trajectory segment is physically consistent by projecting it into the direction that satisfies the constraint, scaling it, or filtering and smoothing it.

[0022] In one embodiment: the force-motion consistency constraint is checked by verifying the mechanical balance relationship between the generated velocity, acceleration and contact force based on the equivalent mass and equivalent damping coefficient of the power tool, wherein the equivalent mass includes the converted value of the end inertia of the robotic arm, and the equivalent damping coefficient is determined based on the dynamic characteristics of the robotic arm and the power tool.

[0023] The energy conservation constraint is checked by verifying the energy input and output balance relationship based on the input energy of the robotic arm, the impact energy input by the power tool itself, the kinetic energy of the power tool, the damping dissipation energy, and the breaking energy of the test medium, wherein the breaking energy is estimated based on the penetration depth and the hardness of the medium.

[0024] The contact stability constraint is checked by verifying that the generated normal contact force is positive and that the time change rate of the normal contact force is within a preset range.

[0025] In one embodiment: the training loss function of the policy model includes a basic generation loss and a physical constraint violation penalty term, wherein the physical constraint violation penalty term is a weighted sum of squares of the violation amount of each physical constraint and the corresponding tolerance threshold;

[0026] During training, the weights of each loss term are dynamically adjusted. In the early stage of training, the basic generation loss is used as the main factor. As training progresses, the weight of the physical constraint violation penalty term is gradually increased until the weights of the physical constraint violation penalty term and the basic generation loss reach a preset ratio.

[0027] When the strategy model uses a diffusion strategy as the backbone network, the physical constraint check and correction are performed after each denoising step; when the strategy model uses an action block transformer as the backbone network, the physical constraint check and correction are performed on the action blocks output by the action block transformer decoder.

[0028] In one embodiment: the policy model incorporates a physical intent discriminator for adversarial training during the training phase; the training steps of the physical intent discriminator specifically include:

[0029] Obtain the sequence of physical semantic feature vectors within a time window;

[0030] Based on the physical semantic feature vector sequence, the contact force-displacement stiffness curve, impedance parameter variation mode, and energy dissipation trajectory are calculated as physical intent features.

[0031] Based on the physical intent features, the style of the action trajectory fragment is determined to guide the parameter update of the policy model. The style includes the human expert demonstration style and the policy generation style.

[0032] In one embodiment: the physical intent discriminator is constructed using a temporal convolutional network superimposed with a self-attention mechanism, taking the physical semantic feature vector sequence as input and outputting the probability that the action trajectory segment belongs to the human expert demonstration style;

[0033] The physical intent discriminator uses manually generated demonstration data as positive samples and early policy-generated trajectories and violation trajectories constructed by applying physical perturbations to the demonstration data as negative samples. After offline training until convergence, the parameters are frozen and used as a trajectory scorer to give expert style scores to the action trajectory segments output by the policy model. The expert style scores are used as a scoring weighting term in the loss function of the policy model or as a screening condition for high-confidence training samples.

[0034] In one embodiment: the policy model employs a bootstrap data augmentation mechanism during training, specifically including:

[0035] In each training round, the policy model generates candidate action trajectory fragments based on the current policy parameters;

[0036] The physical intent discriminator is used to perform expert style scoring on the candidate action trajectory segments, and candidate trajectories with expert style scores higher than a preset threshold are selected as style qualified samples.

[0037] The physical constraint deviation of the style qualified samples is calculated based on the physical constraint conditions, and the style qualified samples whose physical constraint deviation meets the preset conditions are selected as high confidence samples.

[0038] The high-confidence samples are automatically added to the training dataset of the policy model, forming a bootstrapping loop of policy generation, dual screening, and data expansion, so as to continuously expand the training data under the condition of limited manual demonstration data.

[0039] In one embodiment: when the bootstrap loop iterates until the physical constraint violation rate on the validation set and the expert style score of the physical intent discriminator reach a preset standard and no longer improve, the training of the policy model is completed;

[0040] After training, the strategy model is deployed to the robotic arm and tested a preset number of times on a real test medium, while simultaneously collecting real physical feedback data of the robotic arm on the real test medium.

[0041] The strategy model is fine-tuned using the real physical feedback data at a learning rate lower than that used during the training phase, so that the parameters of the physical constraints in the strategy model are adapted to the actual physical characteristics of the real test medium.

[0042] In one embodiment: the step of sending the motion trajectory segment to a force-position hybrid controller, and using the force-position hybrid controller to perform force closed-loop correction based on the force data, and driving the robotic arm to grip the power tool to perform automated testing actions, specifically includes:

[0043] The motion trajectory segment is interpolated into a servo command, the desired contact force is tracked through the force control channel of the force-position hybrid controller, and the lateral motion trajectory is executed through the position control channel.

[0044] The system monitors forces or torques in all directions collected by the six-dimensional force sensor in real time. When an abnormal force exceeding the threshold of normal operating conditions is detected, a compliant response mechanism is triggered in stages according to the degree of abnormality.

[0045] In one embodiment: the tiered triggering of the compliant response mechanism based on the degree of anomaly includes:

[0046] When a Level 1 anomaly is detected, the stiffness parameter of the force-position hybrid controller is reduced without interrupting the test; the Level 1 anomaly is defined as exceeding the normal operating threshold by less than 20%.

[0047] When a level 2 anomaly is detected, the feed rate is reduced and admittance control is activated to generate elastic yield; the level 2 anomaly is defined as exceeding the normal operating threshold by 20% to 50%.

[0048] When a Level 3 anomaly is detected, the chiseling action is paused. After a yielding displacement is generated in the direction of the abnormal force and the tool posture is adjusted, an attempt is made to resume the operation at a reduced feed rate. The Level 3 anomaly is defined as exceeding the normal operating condition threshold by more than 50%.

[0049] Once the abnormal force is eliminated, the normal testing state will automatically resume.

[0050] In one embodiment: before each chisel strike, the normal vector of the test surface is sensed based on the depth image data and the slope angle is calculated. The attack angle of the end of the robotic arm is dynamically planned so that the angle between the tool axis and the normal of the test surface is kept between 0° and 15°. The chisel strike position is planned so that the distance between adjacent chisel strikes is not less than 20mm.

[0051] In one embodiment: the tool life distribution and failure mode distribution obtained from automated testing are statistically compared with the results of manual testing of the same batch of tools; when the mean deviation rate of life is not greater than 15% and the standard deviation rate of life is not greater than 20%, the life comparison is deemed qualified; when the maximum deviation of the proportion distribution of each failure mode is not greater than 10%, the failure mode comparison is deemed qualified; if the representativeness requirement is not met, supplementary manual demonstration data is used to retrain the strategy model.

[0052] In one embodiment: the force-position hybrid controller adopts a hierarchical frequency architecture: the upper layer runs the strategy model inference at a frequency of 20-50Hz and outputs the motion trajectory segment, and the lower layer runs the force-position hybrid controller at a frequency of 1kHz to perform interpolation tracking and force closed-loop correction on the motion trajectory segment.

[0053] On the other hand, the training method for an anthropomorphic automatic testing strategy model for power tools based on a six-dimensional force sensor provided in this application adopts the following technical solution:

[0054] A training method for an anthropomorphic automatic testing strategy model for power tools based on a six-dimensional force sensor, characterized by comprising:

[0055] Acquire multimodal perception data containing force data, depth image data, and pose data; extract terrain features of the target test surface based on the multimodal perception data; construct a physical semantic map of the test scene in combination with preset physical rules; and convert the physical semantic map into a physical semantic feature vector.

[0056] A physical intent discriminator is introduced to perform adversarial training on the policy model. The physical intent discriminator calculates physical intent features based on the sequence of physical semantic feature vectors and discriminates the style of action trajectory segments to guide the parameter update of the policy model. The styles include human expert demonstration style and policy generation style.

[0057] During training, a bootstrap data augmentation mechanism is used, in which the policy model generates candidate action trajectory fragments, which are then filtered by the physical intent discriminator and physical constraints to obtain high-confidence samples and expand them to the training dataset.

[0058] The training of the policy model is completed when the physical constraint violation rate on the validation set and the expert style score of the physical intent discriminator reach a preset standard and no longer improve.

[0059] In one embodiment: the physical intent discriminator is constructed using a temporal convolutional network with a self-attention mechanism, taking the physical semantic feature vector sequence as input and outputting the probability that the action trajectory segment belongs to the human expert demonstration style; the physical intent discriminator uses human demonstration data as positive samples, and early policy-generated trajectories and violation trajectories constructed by applying physical perturbations to the demonstration data as negative samples, and freezes the parameters after offline training until convergence, serving as a trajectory scorer to give an expert style score to the action trajectory segment output by the policy model.

[0060] In one embodiment: In the bootstrap data augmentation mechanism, the number of high-confidence samples added in each round of bootstrap does not exceed 5%-10% of the current training set size, and the policy performance indicators on the validation set are monitored. When the expert style score of the physical intent discriminator on the validation set continues to decline or the physical constraint violation rate continues to rise, the bootstrap data augmentation is paused.

[0061] In one embodiment: after training, the policy model is deployed to a robotic arm and tested on a real test medium, while real physical feedback data is collected simultaneously; the policy model is fine-tuned using the real physical feedback data at a learning rate lower than that during the training phase; after fine-tuning, verification is performed, and if the expert style score decreases by more than 5%, the model parameters are reverted to those before fine-tuning.

[0062] Thirdly, this application provides a human-like automatic testing system for power tools based on a six-dimensional force sensor, which adopts the following technical solution:

[0063] An anthropomorphic automatic testing system for power tools based on a six-dimensional force sensor, comprising:

[0064] Robotic arms are used to hold power tools and perform automated testing actions.

[0065] A six-dimensional force sensor is installed between the end flange and the gripper of the robotic arm for real-time force data acquisition;

[0066] Depth vision sensor, used to acquire depth image data of the test surface;

[0067] The upper-level planning unit is used to construct a physical semantic map of the test scenario based on multimodal perception data, convert the physical semantic map into a physical semantic feature vector and input it into a pre-trained policy model to generate the motion trajectory fragments of the robotic arm.

[0068] The lower-level execution unit includes a force-position hybrid controller, which is used to perform force closed-loop correction on the motion trajectory segment based on the force sensing data, and drive the robotic arm to perform the automated test action.

[0069] Fourthly, the electronic device provided in this application adopts the following technical solution:

[0070] An electronic device includes a processor and a memory, the memory storing a computer program, wherein the processor executes the computer program to implement the steps of the described anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, or to implement the steps of the described training method.

[0071] Fifthly, the electronic device provided in this application adopts the following technical solution:

[0072] A computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the steps of the described anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, or implements the steps of the described training method.

[0073] In summary, this application has the following beneficial effects:

[0074] 1. By reorganizing the raw sensor data into a structured representation that explicitly encodes the entity state and physical constraint relationship through physical semantic graphs, the strategy model gains generalization ability from understanding physical laws rather than memorizing data patterns; at the same time, in the action generation process, three types of physical constraints—force-motion consistency, energy conservation, and contact stability—are checked and corrected in real time to ensure the physical rationality of the generated trajectory from the source and avoid generating abnormal actions that violate the laws of mechanics.

[0075] 2. By implementing force control in the contact normal direction and position control in the tangential direction through the force-position hybrid controller, and triggering a three-level compliant response based on abnormal force levels, the robotic arm can reproduce the behavior pattern of adjusting the magnitude of applied force and posture according to resistance feedback in manual operation. This avoids the additional lateral force and bending moment caused by forced force when rigid equipment encounters obstacles, thereby ensuring that the failure mode is consistent with manual testing and improving the authenticity and repeatability of test results. Attached Figure Description

[0076] Figure 1 This is a flowchart illustrating an anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, as described in this embodiment. Detailed Implementation

[0077] To better understand the purpose, technical solutions, and advantages of this application, it has been described and illustrated below with reference to the accompanying drawings and embodiments. However, those skilled in the art should understand that this application can be implemented without these details. In some cases, to avoid obscuring various aspects of this application due to unnecessary description, well-known methods, processes, systems, components, and / or circuits already described at a higher level will not be elaborated upon. It will be apparent to those skilled in the art that various modifications can be made to the embodiments disclosed in this application, and the general principles defined in this application can be applied to other embodiments and application scenarios without departing from the principles and scope of this application. Therefore, this application is not limited to the illustrated embodiments, but conforms to the broadest scope consistent with the scope of protection claimed in this application.

[0078] It should be noted that the descriptions of these embodiments are for the purpose of aiding understanding the present invention, but do not constitute a limitation thereof. Furthermore, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

[0079] In the description of this application, "several" means one or more, "more than" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0080] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any one or more embodiments or examples.

[0081] See Figure 1 , Figure 1 This is a flowchart illustrating an anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, provided in an embodiment of this application. In this embodiment, the method includes:

[0082] Step S100: Acquire multimodal perception data during the power tool testing process. The multimodal perception data includes force data, depth image data, and pose data.

[0083] It should be noted that multimodal perception data refers to a data set that is synchronously collected by multiple types of sensors and can comprehensively reflect the state of the test scene. Specifically, in this embodiment, the multimodal perception data includes force data, depth image data, and pose data. The force data is collected by a six-dimensional force sensor, the depth image data is collected by a depth vision sensor, and the pose data is collected by a robotic arm encoder.

[0084] A six-dimensional force sensor is installed between the end flange of the robotic arm and a dedicated gripper, enabling real-time detection of the tool's movement in three translational directions (F). x F y F z ) and three rotational directions (M x M y M z The force and torque on the test surface are measured at a sampling frequency of no less than 1 kHz. A depth vision sensor is installed near the end effector of the robotic arm and can employ technologies such as structured light, time-of-flight (ToF), or binocular stereo vision to perceive the three-dimensional shape of the test surface in real time. The robotic arm has a built-in encoder that provides joint angle and end effector pose data.

[0085] Understandably, the synchronous acquisition of multimodal sensing data is fundamental to achieving human-like automated testing. By fusing force, vision, and pose data, the system can comprehensively perceive the physical state of the test scene, providing complete input information for subsequent physical semantic graph construction and strategy model reasoning. The three types of sensors describe the test scene from the dimensions of mechanics, geometry, and kinematics, respectively, complementing each other to form a multi-faceted understanding of the test environment.

[0086] In the specific implementation, the original state vector collected by the system at time t is defined as:

[0087] s rawt =(p t ,q t ,F t M t ,φ(I depth,t ))

[0088] Where, p t =[x t ,y t ,z t ] T q represents the three-dimensional position of the robotic arm's end effector (unit: mm). t =[α t ,β t ,γ t ] T F represents the Euler angles (in rad) of the robotic arm's end effector. t =[F x,t ,F y,t,F z,t ] T The force (in N) measured in three directions by a six-dimensional force sensor, M t =[M x,t M y,t M z,t ] T I represents the torque (in Nm) measured in three directions by a six-dimensional force sensor. depth,t φ(·) is the depth image acquired by the depth vision sensor, and φ(·) is the point cloud encoder that maps the depth image into a 128-dimensional terrain feature vector.

[0089] Step S200: Extract terrain features of the target test surface based on multimodal perception data, and construct a physical semantic map of the test scene in combination with preset physical rules.

[0090] It should be noted that a physical semantic graph is a structured state representation that reorganizes raw sensor data into a graph structure that explicitly encodes entity states and physical relationships between entities. These physical relationships, such as contact force transmission, kinematic constraints, and collision constraints, replace the traditional raw sensor state vectors as input to the policy model. Physical Semantic Graph G t =(V,E) consists of a set of nodes V and a set of edges E, where nodes represent physical entities in the test scenario and edges represent physical constraints between entities.

[0091] Understandably, traditional methods directly concatenate raw sensor data into a state vector and input it into the policy model, lacking explicit encoding of the constraint relationships between physical entities. This forces the model to implicitly learn physical laws from a large amount of training data. In contrast, physical semantic graphs explicitly encode force-motion constraint relationships, enabling the policy model to "understand" rather than "memorize" physical laws, thereby improving generalization ability and physical plausibility.

[0092] In the specific implementation, the physical semantic graph G t The node set V={v1,v2,v3,v4} of the array (V,E) includes: tool / chisel node v1, which contains the state. , where p tip The chisel head position is calculated by combining the positive kinematics of the robotic arm with the offset of the fixture. The chisel speed (determined by the chisel position p) tip (Differential and low-pass filtered to obtain), F tip The resultant force on the chisel head (taken from the six-dimensional force sensor reading after gravity and inertia compensation); test surface node v2, including state [p surf ,n surf ,κ,φ t ] T , where p surf Let n be the location of the contact point.surf The test surface normal vector (obtained by RANSAC plane segmentation), κ is the local curvature, and φ is the local curvature. t The terrain feature vector is extracted by the point cloud encoder; obstacle node v3 contains the state [p obs ,r obs ] T , where p obs r is the center position of the obstacle. obs The equivalent radius of the obstacle; the robotic arm's end effector node v4, containing the state [p end , ,τ joint ] T , where p end and Taken from the encoder reading of the robotic arm, τ joint The result is calculated using the robotic arm's dynamics model combined with joint current. When obstacle node v3 is absent, it is displayed as 0.

[0093] Edge set E={e 12 ,e 13 ,e 14 Includes: chisel head - test surface contact force transmission edge e 12 It always exists, connecting v1 and v2, with the edge attribute being contact force F. contact =F t ; Chisel head - obstacle collision constraint edge e 13 Established when obstacles exist, connecting v1 and v3, with the edge attribute being distance d = ||p obs -p tip The constraint prevents the chisel from entering the obstacle area; the robotic arm end effector-chisel motion constraint edge e 14 It always exists, connecting v1 and v4, with an edge attribute of fixed offset p. offset The constraint relationship is p end =p tip +R(q t )·p offset , where R(q) t ) is the Euler angle q of the final attitude. t A defined rotation matrix.

[0094] Step S300: Convert the physical semantic graph into a physical semantic feature vector, and input the physical semantic feature vector into a pre-trained policy model to generate motion trajectory segments of the robotic arm; wherein, the policy model performs consistency correction on the motion trajectory segments based on physical constraints during training and generation.

[0095] It should be noted that the physics semantic graph needs to be converted into a fixed-dimensional feature vector before it can be input into the neural network of the policy model. The policy model π(a|s) defines the mapping relationship from the state space S to the action space A. In this application, the actual input of the policy model is the physics semantic feature vector g. t Instead of the original state vector s rawt .

[0096] The action vector output by the policy model is a t =(Δp,Δq,F d M d ,K), where Δp=[Δx,Δy,Δz] T The end displacement increment (unit: mm / cycle, range ±5 mm) is given by Δq = [Δα, Δβ, Δγ]. T F represents the attitude adjustment (unit: rad / cycle, range ±0.05 rad). d =[F dx ,F dy ,F dz ] T The target contact force (unit: N, range 0-500N), M d =[M dx M dy M dz ] T For the target torque (unit: Nm, range: 0-50 Nm), K = [K x ,K y ,K z ,K rx ,K ry ,K rz ]T This is the impedance stiffness parameter (unit: N / m or Nm / rad, range 500-10000).

[0097] Understandably, the policy model, based on traditional imitation learning methods such as Behavior Cloning (BC), Dataset Aggregation (DAgger), Diffusion Strategy, and Action Block Transformer (ACT), adds physical semantic enhancement modules such as the Physical Semantic Encoder (PSE), Physical Consistency Generator (PCG), and Physical Intent Discriminator (PID), forming a training framework that combines a basic generator backbone with physical semantic enhancement. The PSE reconstructs raw sensor data into structured physical representations; the PCG adds physical constraint corrections during the original generation process to ensure the physical rationality of the generated trajectories; and the PID upgrades the original discriminator from discriminating trajectory distribution to discriminating physical intent. These three components enhance the original training framework from three dimensions: representation, generation, and discrimination.

[0098] In its implementation, each inference iteration of the policy model outputs a motion trajectory segment containing a sequence of positions, forces, and stiffness parameters for a predetermined future time period. The upper low-frequency planning layer runs the policy model inference at a frequency of 20-50Hz, while the lower high-frequency execution layer runs the force-position hybrid controller at a frequency of 1kHz. The motion trajectory segment output by the upper layer is interpolated into fine-grained servo commands using either a zero-order hold or linear interpolation. During training and generation, the policy model performs consistency correction on the motion trajectory segment based on physical constraints to ensure that the generated trajectory conforms to physical laws.

[0099] In step S400, the motion trajectory segment is sent to the force-position hybrid controller. The force-position hybrid controller performs force closed-loop correction based on the force data and drives the robotic arm to hold the power tool to perform automated test actions.

[0100] It should be noted that the force-position hybrid controller is a crucial link between the strategy model output and the robotic arm execution. It uses a selection matrix S=diag(s1,...,s6) to decompose the task space into force control and position control channels, achieving decoupled execution of force control and position control. Force closed-loop correction refers to the process by which the controller tracks and corrects the desired contact force output by the strategy model based on real-time force feedback data from the six-dimensional force sensor.

[0101] Understandably, by using a force-position hybrid controller to transform the high-level planning of the strategy model into actual robotic arm motion commands, and by performing real-time force closed-loop correction based on force feedback data, the robotic arm can precisely control the contact force as if it were a human operator when performing automated testing. This avoids the additional lateral forces and bending moments caused by the lack of force feedback in rigid automated equipment, thereby ensuring the authenticity of the test results.

[0102] In its implementation, the force-position hybrid controller selects force control (s3=0) in the contact normal (Z-axis) and tracks the desired contact force F output by the strategy model. d Position control (s1=s2=1) is selected in the tangential direction (X, Y axes) to execute the lateral motion trajectory planned by the strategy model. Force control error e f =F d -F measured The PI controller converts the position correction value Δp. f u(t)=K p ·e f (t)+K i ·∫e f (τ)dτ, where K p The proportional gain indicates that for every 1N difference in force error, a corresponding K gain is generated. p The position correction amount; the larger the error, the faster the correction. K iThe integral gain represents the increase in Ks for every additional unit of accumulated force error (N·s). i The correction is used to eliminate steady-state deviations and prevent the actual contact force from consistently falling below the expected force. The attitude degrees of freedom are dynamically adjusted by integrating depth visual terrain information and six-dimensional torque feedback.

[0103] In one possible implementation, step S200 specifically includes:

[0104] S201. Map the depth image data into terrain features using a point cloud encoder;

[0105] S202. Based on pose data, force data and terrain features, construct a node set, which includes tool chisel head nodes, test surface nodes, obstacle nodes and robotic arm end effector nodes.

[0106] S203. Construct an edge set based on the contact force transmission relationship, kinematic constraint relationship and collision constraint relationship between nodes in the node set.

[0107] In the embodiments described in this specification, the point cloud encoder φ(·) converts the depth image I output by the depth vision sensor into a digital image. depth,t The terrain feature vector is mapped, and the encoder can be implemented using PointNet or a variant network. After the terrain features are extracted, four nodes of the physical semantic map are constructed based on the pose data of the robotic arm encoder, the force data of the six-dimensional force sensor, and the extracted terrain features: tool chisel node v1, test surface node v2, obstacle node v3, and robotic arm end effector node v4.

[0108] The chisel head node state is calculated using the forward kinematics of the robotic arm, the test surface node state is obtained through RANSAC plane segmentation and local curvature estimation, obstacle nodes are created when depth vision detects an obstacle, and the end-effector node state is directly taken from encoder readings. The edge set is constructed based on the physical relationships between nodes: e 12 This indicates the force transfer relationship between the chisel and the test surface, e 14 e represents the kinematic constraint relationship between the end effector and the chisel head. 13 This indicates the collision constraint relationship between the chisel and the obstacle.

[0109] Furthermore, the step of converting the physical semantic graph into a physical semantic feature vector in step S300 specifically includes:

[0110] S301. Concatenate the state vectors of each node in the node set into a node feature sequence according to a preset order, wherein the state vector includes at least one of the corresponding node's position, velocity, force or curvature.

[0111] S302. Concatenate the attribute vectors of each edge in the edge set into an edge feature sequence in a preset order, wherein the attribute vectors include at least one of the contact force, fixed offset or distance of the corresponding edge.

[0112] S303. Concatenate the node feature sequence and the edge feature sequence into a physical semantic feature vector. The input of the policy model must contain at least the physical semantic feature vector.

[0113] In the embodiments described in this specification, the physical semantic graph G t The vectorization process of =(V,E) is as follows: First, the state vectors of each node are concatenated into h in a fixed order. V =[h v1 ,h v2 ,h v3 ,h v4 ] T , where h vi For node v i The state vector; then the attributes of each edge are concatenated into h in a fixed order. E =[h e12 ,h e13 ,h e14 ] T Finally, h V and h E Concatenate to form the final physical semantic feature vector g t =[h V ,h E ] T ∈R dg , where d g It is approximately 160-180 dimensions. The input to the policy model generator must contain at least this physical semantic feature vector g. t It can independently replace the original state vector s rawt It can also be used with s rawt After being concatenated, they are input together. Through this structured vectorization method, the entity states and inter-entity constraints in the physical semantic graph are fully preserved and passed to the policy model.

[0114] In one implementation, step S300, which involves the policy model performing consistency correction on motion trajectory segments based on physical constraints during training and generation, specifically includes:

[0115] S311. During the generation of motion trajectory segments by the strategy model, physical constraint checks are performed in real time. Physical constraint checks include force-motion consistency constraints, energy conservation constraints, and contact stability constraints.

[0116] S312. When a motion trajectory segment is detected to violate any physical constraint check, the motion trajectory segment is physically consistent by projecting it into the direction that satisfies the constraint, scaling it, or filtering it to smooth it.

[0117] In the embodiments described in this specification, regardless of whether the strategy model uses BC, DAgger, diffusion strategy, or ACT backbone network, its input is always the physical semantic feature vector g. t Furthermore, physical constraint correction mechanisms are added during the generation process. Taking the diffusion strategy as an example, the diffusion strategy models action generation as an iterative denoising process from pure noise to the target action sequence. The forward diffusion process is as follows:

[0118]

[0119] Where a 0 This is a real action sequence. =Πα i α i Let N(0,I) be the cosine noise scheduling coefficient, and let N(0,I) be a standard normal distribution. This invention performs physical constraint checks and corrections on the currently predicted action sequence after each denoising step. Force-motion consistency constraint requirements:

[0120]

[0121] Where F arm (t) represents the main force applied to the tool by the end effector of the robotic arm, m tool Let F be the total mass of the power tool and fixture, g be the vector of gravitational acceleration, and F be the total mass of the power tool and fixture. contact (t) represents the contact force, m eff For tool equivalent quality, c eff Here, v(t) is the equivalent damping coefficient (N·s / m), and v(t) is the generation velocity. The generated acceleration. If violated, the generated velocity and acceleration will be corrected by projecting them towards the direction satisfying the above equation. Energy conservation constraint requirements:

[0122] |E input (t)-E kinetic (t)-E damping (t)-E fracture (t)|<ε energy

[0123] If the proportional scaling of velocity or force is violated to restore energy conservation, contact stability constraints must be met.

[0124] F n (t)>0 and |dF n / dt|<ε rate

[0125] If the normal force is negative, it is corrected to a small positive number (e.g., 1N); if the rate of change is too large, low-pass filtering is used for smoothing. For the ACT backbone, physical constraint checking and correction are performed uniformly on the action blocks output by the decoder; for the BC and DAgger backbones, physical constraint checking is performed on the actions output by the model during training, and a physical constraint violation penalty term is added to the loss function.

[0126] Furthermore, the method for checking the force-motion consistency constraint is as follows: based on the equivalent mass and equivalent damping coefficient of the power tool, the mechanical balance relationship between the generated velocity, acceleration and contact force is verified, where the equivalent mass includes the converted value of the end inertia of the robotic arm, and the equivalent damping coefficient is determined based on the dynamic characteristics of the robotic arm and the power tool.

[0127] The energy conservation constraint is checked by verifying the energy input and output balance relationship based on the input energy of the robotic arm, the impact energy input of the power tool itself, the kinetic energy of the power tool, the damping dissipation energy, and the breaking energy of the test medium. The breaking energy is estimated based on the penetration depth and the hardness of the medium.

[0128] The contact stability constraint is checked by verifying that the normal contact force is positive and that the rate of change of the normal contact force over time is within a preset range.

[0129] In the embodiments of this specification, the specific method for checking the force-motion consistency constraint is as follows: taking the power tool, fixture, and robotic arm end effector as a combined rigid body, first calculate the operational space inertia matrix Λ of the robotic arm end effector. ee =(J ee ·M arm -1 ·J ee T ) -1 J ee M is the Jacobian matrix at the end effector of the robotic arm. arm Let J be the joint space inertia matrix of the robotic arm in the current configuration; when J ee When approaching singular configurations, damped least squares pseudo-inverse is used instead of the inversion operation in the above equation. The equivalent mass m in the contact normal direction... eff As a scalar, its calculation method is as follows:

[0130] m eff =m tool +1 / (n surf T ·Λ v -1 ·n surf )

[0131] Where m tool Let n be the total mass of the power tool and fixture being tested. surf Let Λ be the unit vector normal to the test surface.v For Λ ee The 3×3 submatrix corresponds to the three translational degrees of freedom. The equivalent damping coefficient c eff The following method was used to determine c: During the calibration phase before testing, the robotic arm's end effector was driven to move in free space at different speeds, and the steady-state force at each speed was recorded. c was then obtained through least-squares fitting. eff The value typically ranges from 50 to 200 N·s / m.

[0132] The formula for checking force-motion consistency constraints is:

[0133] |F arm +m tool ·gF contact -(m eff ·a+c eff ·v)|≤ε force

[0134] Where F arm The main force applied to the tool by the end effector of the robotic arm is obtained by the pseudo-inverse mapping of the joint torque via the transpose of the Jacobian matrix, i.e., F. arm =(J ee T ) + ·τ act , τ act The joint torque is calculated by subtracting the robot arm's own weight and inertia; g is the gravitational acceleration vector, and F... contact The resultant contact force is measured by a six-dimensional force sensor, where a is the generated acceleration, v is the generated velocity, and ε is the resultant force. force A force check tolerance threshold is set. When this constraint is violated, the velocity vector is projected back towards the force equilibrium direction for correction.

[0135] The specific method for checking the energy conservation constraint is as follows: take the calculation time window Δt. eval (Typically, a complete chiseling cycle is taken), define the input energy E of the robotic arm. input The impact energy E input by the power tool's own impact mechanism is the integral of the joint's active torque and joint angular velocity within this window. tool =E impact ·f impact ·Δt eval (where E) impact For a single impact energy, f impact (Impact frequency), tool kinetic energy change ΔE kinetic 0.5 m eff • (v2²-v1²) (v1 and v2 are the velocities of the tool along the chisel direction at the start and end times of the window), damping dissipation energy E damping For c eff • v² is the integral of v² over time within this window. Energy consumed by breaking down the structure, E.fracture Estimate using the following methods:

[0136] E fracture =k h ·A·d

[0137] Where k h denoted as the hardness coefficient of the medium (unit: Pa), A is the cross-sectional area of ​​the chisel, and d is the cumulative penetration depth of the chisel within the calculation window.

[0138] The energy conservation check formula is:

[0139] |E input +E tool -(ΔE kinetic +E damping +E fracture )|≤ε energy

[0140] Where ε energy This is the tolerance threshold for energy checks (in J). When this constraint is violated, the velocity or force is scaled up to restore energy conservation.

[0141] The specific method for checking contact stability constraints is as follows: Check the expected normal contact force F generated by the strategy model. n Does F satisfy? n >0, if F n If the value is ≤0, it should be corrected to a small positive number (e.g., 1N); at the same time, check whether the time rate of change of the normal contact force meets the following requirements:

[0142] |dF n / dt|≤ε rate

[0143] Where ε rate The force change rate tolerance threshold is determined based on the statistical distribution of the normal contact force change rate in the manual demonstration data, with a typical value being 1.5 times the 99th percentile of the normal contact force change rate in the manual demonstration data. If the change rate is too large, low-pass filtering is applied for smoothing, and the filter cutoff frequency is set based on the control cycle.

[0144] In one possible implementation, the training loss function of the policy model includes a basic generation loss and a physical constraint violation penalty term, wherein the physical constraint violation penalty term is a weighted sum of squares of the violation amount of each physical constraint and the corresponding tolerance threshold.

[0145] During training, the weights of each loss term are dynamically adjusted. In the early stage of training, the basic generation loss is mainly used. As training progresses, the weight of the physical constraint violation penalty term is gradually increased until the weights of the physical constraint violation penalty term and the basic generation loss reach a preset ratio.

[0146] When the policy model uses a diffusion strategy as the backbone network, physical constraint checks and corrections are performed after each denoising step; when the policy model uses an action block transformer as the backbone network, physical constraint checks and corrections are performed on the action blocks output by the decoder of the action block transformer.

[0147] In the embodiments described in this specification, the total loss function of the generator is:

[0148] L PCG =L base +λ phy ·L physics +λ adv ·L adversarial

[0149] Where L base The loss is generated based on the standard diffusion model denoising loss when using a diffusion strategy as the backbone network, and action regression loss when using behavior cloning. physics =Σw k ·[ReLU(φ k (â)-ε k )] 2 Let φ be the weighted sum of squares of the penalties for violating the three physical constraints. k (â) represents the violation amount of the k-th physical constraint, ε k L is the tolerance threshold. adversarial For the adversarial loss from the physical intent discriminator, w k Let λ be the weight coefficient of the k-th physical constraint. phy λ is the weighting coefficient for the physical constraint violation penalty term. adv Weighting coefficients to counteract losses.

[0150] During training, the weights of each loss term are dynamically adjusted, and the weight coefficient λ of the physical constraint violation penalty term is adjusted accordingly. phy (t) increases gradually with the number of training steps t according to the following formula:

[0151] λ phy (t)=λ min +(λ max -λ min )·min(1,t / T warmup )

[0152] Where λ min Let λ be the initial weight. max For the target weight, T warmup This refers to the number of preheating steps. Scheduling methods include, but are not limited to, linear scheduling; cosine annealing scheduling can also be used.

[0153] λ phy (t)=λ min +(λ max-λ min )·(1-cos(π·min(1,t / T warmup ))) / 2

[0154] In the early stages of training (t is much smaller than T) warmup The loss function generates a basic loss L. base The primary goal is to ensure that the policy model first learns to generate reasonable action sequences; as training progresses, the weight of the penalty term for violating physical constraints gradually increases, forcing the generated trajectories to gradually satisfy the physical constraints. Where L... base The loss is generated based on the standard diffusion model denoising loss or behavioral clonal regression loss, depending on the type of backbone network.

[0155] When the strategy model uses a diffusion strategy as the backbone network, physical constraint checks and corrections are performed immediately after each denoising step. Let the total number of denoising steps be T. denoise After denoising at step t, the current predicted action sequence is obtained. t Three types of physical constraint checks are performed on the sequence, and positions that violate the constraints are corrected. The corrected sequence is... t As input for the next step of denoising:

[0156] â t =DenoiseStep(â t-1 ,t), â t =PhysicsCorrect(â) t )

[0157] When the policy model uses the Action Block Transformer (ACT) as the backbone network, physical constraint checking and correction are performed uniformly across the entire action block output by the ACT decoder. Let the action block output by the ACT be A. chunk ={a1,a2,...,a K (K is the length of the action block), and perform physical constraint checks and corrections on each action step in the action block in sequence:

[0158] A' chunk ={PhysicsCorrect(a1),PhysicsCorrect(a2),...,PhysicsCorrect(a K )}

[0159] Meanwhile, during backpropagation, an additional physical penalty term is added to the total loss function for generated samples that violate physical constraints. This penalty term only calculates the gradient for action steps that violate the constraints. When the policy model uses Behavior Cloning (BC) or Dataset Aggregation (DAgger) as the backbone network, physical constraints are checked on the actions output by the model during training, and a physical constraint violation penalty term is added to the loss function. Physical constraint checks and corrections are performed after each action output during both the training and inference phases.

[0160] In one possible implementation, the policy model incorporates a physical intent discriminator for adversarial training during the training phase; the training steps for the physical intent discriminator specifically include:

[0161] Obtain the sequence of physical semantic feature vectors within a time window;

[0162] Based on the physical semantic feature vector sequence, the contact force-displacement stiffness curve, impedance parameter variation mode and energy dissipation trajectory are calculated as physical intent features;

[0163] Based on physical intent characteristics, the style of action trajectory segments is determined to guide the parameter update of the policy model. The styles include human expert demonstration style and policy generation style.

[0164] In the embodiments described in this specification, the Physical Intent Discriminator (PID) upgrades the discriminator in traditional GAIL from discriminating trajectory distribution to discriminating physical intent. The input to the PID is a time window T. win =The sequence of physical semantic feature vectors within 1 second {g t-Twin ,...,g t The output is the binary classification probability P(expert|g)∈[0,1] that the sequence belongs to "expert style" or "strategy style". The PID discrimination criteria include three physical intent features: contact force-displacement stiffness curve, and real-time calculation of incremental stiffness K. inc =ΔF contact / Δδ penetration , where ΔF contact Δδ is the change in contact force. penetration This curve represents the change in chisel penetration depth, reflecting the operator's "hard / soft" control attitude when facing different resistances; the impedance parameter variation pattern estimates the equivalent stiffness in real time based on the force-motion relationship. and damping The time-varying trajectory; the energy dissipation trajectory, and the real-time calculation of power P(t)=F contact (t)·v penetration (t) and dissipation rate η(t) = P(t) / (F contact (t)·v max ), where v maxThis represents the maximum feed rate. The loss function for PID control is the binary cross-entropy: L PID =-E τ~πE [logP(expert|g τ )]-E τ~πθ [log(1-P(expert|g τ ))], where π E For expert strategies, π θ This represents the current strategy to be trained.

[0165] Furthermore, the physical intent discriminator is constructed using a temporal convolutional network superimposed with a self-attention mechanism. It takes the sequence of physical semantic feature vectors as input and outputs the probability that the action trajectory fragment belongs to the human expert demonstration style.

[0166] The physical intent discriminator uses human-generated demonstration data as positive samples and early policy-generated trajectories and violation trajectories constructed by applying physical perturbations to the demonstration data as negative samples. After offline training until convergence, the parameters are frozen and used as a trajectory scorer to give expert style scores to the action trajectory segments output by the policy model. The expert style scores are used as a scoring weighting term in the policy model loss function or as a selection criterion for high-confidence training samples.

[0167] In the embodiments described in this specification, the Physical Intent Discriminator (PID) is a neural network consisting of a Temporal Convolutional Network (TCN) with a self-attention mechanism. The coupling between the PID and the generator includes two implementation methods:

[0168] 1. In the online adversarial mode, the PID and generator are updated alternately, and the discriminative output of the PID is used as the style adversarial term in the total loss of the generator;

[0169] 2. Offline scoring method: This method is preferred. The PID uses manually generated demonstration data as positive samples and trajectories generated by the early strategy and violation trajectories constructed by applying physical perturbations to the demonstration data as negative samples. After offline training until convergence, the parameters are frozen and used as a trajectory scorer. The frozen PID provides an expert-style score for the generator's output trajectory. This score is used as: a score weighting term in the generator's total loss (no longer used to update the PID in reverse); and as a selection criterion for high-confidence samples in bootstrapping data augmentation.

[0170] In one possible implementation, the policy model employs a bootstrap data augmentation mechanism during training, specifically including:

[0171] In each training round, the policy model generates candidate action trajectory fragments based on the current policy parameters;

[0172] A physical intent discriminator is used to perform expert style scoring on candidate motion trajectory segments, and candidate trajectories with expert style scores higher than a preset threshold are selected as style qualified samples.

[0173] The physical constraint deviation of style qualified samples is calculated based on physical constraints, and style qualified samples whose physical constraint deviation meets the preset conditions are selected as high confidence samples.

[0174] High-confidence samples are automatically added to the training dataset of the policy model, forming a bootstrapping loop of policy generation, dual screening, and data expansion, so as to continuously expand the training data under the condition of limited manual demonstration data.

[0175] In the embodiments of this specification, the specific implementation method of bootstrap data augmentation is as follows:

[0176] Step S1: Candidate Trajectory Generation. In each training round, the generator of the policy model generates a set of candidate action trajectory fragments based on the current policy parameters and the state sequence sampled from the environment, where each candidate trajectory contains a state-action pair sequence.

[0177] Step S2: Physical Intent Style Filtering. The frozen Physical Intent Discriminator (PID) is used to score the expert style of each candidate trajectory. The calculation method is as follows: input the physical semantic feature vector sequence corresponding to the trajectory into the PID to obtain the probability S that the trajectory belongs to the human expert demonstration style. PID The filtering criteria are: S PID >θ style , where θ style The style selection threshold is used. Candidate trajectories that pass the selection are labeled as the style-qualified sample set D. style .

[0178] Step S3: Physical constraint deviation screening. For the style-qualified sample set D style For each trajectory, its physical constraint deviation vector is calculated based on three types of physical constraints: δφ=[δφ1,δφ2,δφ3], where δφ1, δφ2, and δφ3 are the force-motion consistency violation, energy conservation violation, and contact stability violation, respectively. The calculation method is the absolute value of the left-side deviation from the right-side in each constraint check formula. The screening criterion is: max k (δφ k / ε k )<θ phys , where ε k Let θ be the tolerance threshold for the k-th physical constraint. phys The physical deviation screening threshold is used. Trajectories that simultaneously meet both style screening and physical constraint deviation screening are marked as high-confidence samples, forming the high-confidence sample set D. HQ .

[0179] Step S4: Data Expansion. Expand the high-confidence sample set D... HQ Automatically add the training dataset D of the policy modeltrain In the updated training dataset, D is: train =D train ∪D HQ In the next training epoch, the policy model continues training based on the expanded training dataset. The above steps form a bootstrapping loop of policy generation, PID style selection, physical constraint bias selection, and data expansion. This allows the policy model to continuously generate and select high-quality training data under the condition of limited manually generated demonstration data (usually only 20-50 demonstration trajectories), gradually improving the expert style consistency and physical rationality of the policy.

[0180] Step S5: Bootstrap Quality Control. To avoid distribution shifts during the bootstrap process, an upper limit is set on the number of high-confidence samples added in each round of bootstrap (typically 5%-10% of the current training set size), and policy performance metrics on the validation set are monitored. When the PID score on the validation set continues to decline or the physical constraint violation rate continues to rise, bootstrap data augmentation is paused, and training is performed using only the original human-generated demo data.

[0181] Furthermore, the training of the policy model is completed when the physical constraint violation rate and the expert style score of the physical intent discriminator on the validation set reach the preset standard and no longer improve.

[0182] After training, the strategy model is deployed to the robotic arm and tested a preset number of times on a real test medium, while simultaneously collecting real physical feedback data of the robotic arm on the real test medium.

[0183] By using a learning rate lower than that during the training phase, the policy model is fine-tuned using real physical feedback data, so that the parameters of the physical constraints in the policy model are adapted to the actual physical characteristics of the real test medium.

[0184] In the embodiments of this specification, for the online adversarial method, the termination condition is that the discrimination accuracy of the discriminator is close to the random level, that is, the binary classification accuracy is between 45% and 55%, indicating that the trajectory generated by the strategy is difficult to distinguish from the manual demonstration in the physical intent feature space.

[0185] For offline scoring methods, the termination condition must simultaneously meet the following three indicators:

[0186] (a) Physical constraint violation rate V phy Below the preset violation rate threshold ε violate :

[0187] V phy =|{τ i :max k (δφ k / ε k )>1}| / |D val|<ε violate

[0188] Where D val For the validation set, typical ε violate Values ​​range from 0.05 to 0.10;

[0189] (b) The mean score of the PID expert style score is higher than the preset score threshold θ. quality :

[0190] mean(S PID (τ i ))>θ quality

[0191] Where τ i Belongs to D val ;

[0192] (c) The above two indicators are in continuous N patience N training rounds (typically N) patience The value (10-20) will not improve further. The decision strategy model is considered complete when all three conditions are met.

[0193] The detailed implementation method for physical consistency fine-tuning after training (before deployment) is as follows:

[0194] Step S1: Deployment and Trial Run. Deploy the trained strategy model to the industrial control computer and establish a communication connection with the real robotic arm. Conduct a small-scale trial run on a real test medium (concrete test block), with 20-50 chiseling cycles. During the trial run, the system synchronously collects real physical feedback data, including: force / torque time series from the six-dimensional force sensor, pose time series from the robotic arm encoder, and depth image series from the depth vision sensor.

[0195] Step S2: Preprocessing of Real Physical Feedback Data. The collected real physical feedback data is constructed into a sequence of physical semantic graphs using a physical semantic encoder method, and then quantized into a sequence of physical semantic feature vectors. This sequence, along with the corresponding real robotic arm motion sequences, forms a fine-tuning dataset D. finetune .

[0196] Step S3: Fine-tuning training. Fine-tune the policy model using a learning rate lower than the learning rate used during the training phase. The fine-tuning learning rate is lr. finetune The learning rate lr during the training phase train 1 / 10, that is: lr finetune =lr train / 10. During fine-tuning, the parameters of the Physical Semantic Encoder (PSE) and Physical Intent Discriminator (PID) of the policy model are frozen, and only the generator parameters are updated to minimize the deviation of the generator's output motion trajectory from the physical constraints on the real medium. The fine-tuning loss function is:

[0197] L finetune =L base +Σ k w k ·max(0,δφ kreal -ε k )²

[0198] Where δφ k `real` represents the physical constraint violation rate calculated based on real physical feedback data. The number of fine-tuning training epochs is limited to 5-20 epochs to prevent overfitting to the specific characteristics of the real medium and thus losing cross-medium generalization ability.

[0199] Step S4: Verification of Fine-Tuning Effect. After fine-tuning, perform 100-200 chiseling cycles on a real test medium to verify the results. Compare the physical constraint violation rate and PID expert style score before and after fine-tuning. If the physical constraint violation rate decreases after fine-tuning and the PID score does not decrease significantly (the decrease does not exceed 5%), the fine-tuning is confirmed to be effective; otherwise, revert to the model parameters before fine-tuning.

[0200] In one possible implementation, the step of sending motion trajectory segments to a force-position hybrid controller, and then using the force-position hybrid controller to perform force closed-loop correction based on force data to drive a robotic arm to grip a power tool and perform automated testing actions specifically includes:

[0201] The motion trajectory segment is interpolated into a servo command, the desired contact force is tracked through the force control channel of the force-position hybrid controller, and the lateral motion trajectory is executed through the position control channel; the force or torque in each direction collected by the six-dimensional force sensor is monitored in real time, and when an abnormal force exceeding the normal working condition threshold is detected, the compliant response mechanism is triggered in stages according to the degree of abnormality.

[0202] In the embodiments described in this specification, the force-position hybrid controller adopts a hierarchical control architecture: the upper low-frequency planning layer runs the strategy model inference at 20-50Hz and outputs motion trajectory segments; the lower high-frequency execution layer runs the force-position hybrid controller at 1kHz to perform interpolation tracking and force closed-loop correction on the motion trajectory segments. Force control is selected in the contact normal direction to track the desired contact force output by the strategy model, achieving a human-like force distribution; position control is selected in the tangential direction to execute the lateral motion trajectory planned by the strategy model. Based on the force / torque statistical distribution recorded by the six-dimensional force sensor in the artificial demonstration data, a normal operating range threshold [μ] is set for each direction. i -3σ i ,μ i +3σ iThe system monitors readings from each channel in real time. When an abnormal force is detected, it triggers a compliant response mechanism based on the severity of the abnormality. Without waiting for the upper-level model to re-infer, the lower-level controller can trigger an emergency yield response within 1ms.

[0203] Furthermore, an adaptive response mechanism is triggered in stages based on the severity of the anomaly, including:

[0204] When a Level 1 anomaly is detected, the stiffness parameter of the force-position hybrid controller is reduced without interrupting the test; a Level 1 anomaly is defined as exceeding the normal operating threshold by less than 20%.

[0205] When a level 2 anomaly is detected, the feed rate is reduced and admittance control is activated to generate elastic yield; a level 2 anomaly is defined as exceeding the normal operating threshold by 20% to 50%.

[0206] When a Level 3 anomaly is detected, the chiseling action is paused. After the tool posture is adjusted and a yielding displacement is generated in the direction of the abnormal force, an attempt is made to resume the operation at a reduced feed rate. A Level 3 anomaly is defined as exceeding the normal operating threshold by more than 50%.

[0207] Once the abnormal force is eliminated, the system will automatically return to normal testing status.

[0208] The above-mentioned normal operating condition threshold is obtained based on the historical data collected by the six-dimensional force sensor and is defined as the μ±3σ range, where μ is the mean of force / torque in each direction and σ is the standard deviation. The deviation range is classified according to the relative amount by which the measured value exceeds the threshold boundary, i.e., |measured value - threshold boundary| / |threshold boundary|.

[0209] In the embodiments of this specification, the specific execution method of the three-level compliance response is as follows: In the first-level abnormal response, the stiffness K is reduced by 30-50% without interrupting the test, so that the robotic arm produces a slight compliance when the contact force is slightly abnormal; in the second-level abnormal response, the speed is reduced to 50%, and admittance control is initiated to produce elastic yielding. The admittance control equation is:

[0210]

[0211] Among them, M adm For virtual inertia, B adm For virtual damping, K adm For the admittance stiffness (dynamically output by the strategy model), when the yield displacement reaches steady state, this equation degenerates into Δp=F. ext / K admIn the Level 3 anomaly response, the chiseling action is paused, a yield displacement of no less than 5mm is generated in the direction of the abnormal force, and the tool posture is adjusted by 10°-15° before attempting to recover at 50% feed rate. The anomaly is cleared when all channel readings return to the threshold and remain stable for 2 seconds. Normal testing is automatically resumed after the anomaly is cleared. This three-level compliant response mechanism allows the system to flexibly respond to anomalies of varying severity, much like a human operator, avoiding premature tool failure caused by rigid automated equipment applying excessive force when encountering obstacles. (F...) ext The abnormal external force detected by the six-dimensional force sensor.

[0212] Optionally, embodiments of this application provide a complete joint training process, including the following steps:

[0213] Step 1: Convert the collected demonstration data into a sequence of physical semantic feature vectors {g} using a Physical Semantic Encoder (PSE). tE}, and the corresponding expert action {a tE The training dataset is constructed from the data collected by experienced electric hammer operators at a real test site. The system simultaneously collects pose data, force / torque data, depth image sequences, and tool status data. The collected raw data is then converted into a physical semantic graph sequence {G}. tE}, then vectorized into a sequence of physical semantic feature vectors {g tE}

[0214] Step 2: Initialize the generator backbone network (choose one from BC / DAgger / diffusion strategy / ACT), whose input is the physical semantic feature vector g. t .

[0215] Step 3: Construct the PID. In Implementation A, initialization is randomized; in Implementation B, offline training with positive and negative samples is performed as described above until convergence and then the PID is frozen.

[0216] Step 4: Generator Update. Implementation A: Update the PID first, then the generator in each round; Implementation B: Update only the generator, with the frozen PID providing trajectory scoring.

[0217] Step 5: If DAgger is used as the backbone, after each round of training, the current policy collects new states in the test environment, the actions are re-labeled by experts, and the new data is added to the training set for retraining.

[0218] Step 6: If a diffusion strategy or ACT is used, the aforementioned physical constraint corrections are forcibly applied during the generation process.

[0219] Step 7: Bootstrap data augmentation. High-confidence trajectories generated by the generator (samples whose PID judgment is higher than the preset threshold and whose physical constraint deviation is the smallest) are automatically added to the training set to expand the training data.

[0220] Step 8: Termination condition determination. Implementation method A: Iterate until the discrimination accuracy approaches a random level; Implementation method B: Iterate until the physical constraint violation rate and PID score on the validation set reach a preset standard and no longer improve.

[0221] Optionally, after fine-tuning, the system enters fully automatic test mode, where real-time inference of the strategy model is executed at a frequency of 20-50Hz, and the following closed loop is executed in each control cycle:

[0222] (1) Status acquisition: Simultaneously read real-time data from the six-dimensional force sensor (1kHz), depth vision sensor (30fps), and robotic arm encoder (1kHz);

[0223] (2) Terrain feature extraction: The depth image output by the depth vision sensor is extracted into a terrain feature vector φ(I) by the point cloud encoder φ(·). depth,t );

[0224] (3) Construction of the physical semantic graph: (p t ,q t ,F t M t ,φ t The physical semantic encoder method is used to construct a physical semantic graph and quantize it into g. t ;

[0225] (4) Model reasoning: using g t As input, the generator performs one forward propagation and outputs an action vector a. t =(Δp,Δq,F d M d ,K);

[0226] (5) Action execution: The action vector is sent to the lower-level force-position hybrid controller (1kHz), and after interpolation tracking and force closed-loop correction, it drives the robotic arm to execute;

[0227] (6) Status update: The robotic arm performs an action, the sensor data is updated, and the next cycle begins.

[0228] Optionally, before each strike, a depth vision sensor detects the test surface normal vector *n* and the obstacle position. The generator calculates the slope angle based on the test surface normal vector and, combined with the obstacle position information, dynamically plans the attack angle of the robotic arm end effector (keeping the angle between the tool axis and the test surface normal between 0° and 15°) and the strike point position (avoiding obstacles, with a spacing ≥ 20mm). This ensures that the tool always contacts the test surface at a natural angle, replicating the behavior of adjusting posture according to ground conditions during manual operation. The slope angle is calculated as follows: the degree of inclination of the test surface is determined based on the angle between the test surface normal vector *n* and the direction of gravity. When the slope angle exceeds a preset threshold, the generator adjusts the attack angle and strike force accordingly to adapt to the force characteristics of the inclined test surface.

[0229] Optionally, the tool lifetime distribution and failure mode distribution obtained from automated testing are statistically compared with the results of manual testing of the same batch of tools: Regarding lifetime comparison, the lifetime mean deviation rate ε... μ =|μ auto -μ manual | / μ manual ≤15% and lifetime standard deviation ε σ =|σ auto -σ manual | / σ manual A failure rate of ≤20% is considered acceptable; regarding the comparison of failure modes, the maximum deviation δ of the proportion distribution of each failure mode is... max =max k |p auto,k -p manual,k A representativeness level of ≤10% is considered acceptable. If the representativeness requirement is not met, supplementary manually generated demonstration data is provided, and the policy model is retrained. Where μ auto and μ manual The lifetime values ​​for automated testing and manual testing are σ, respectively. auto and σ manual The lifetime standard deviations for both are p, respectively. auto,k and p manual,k δ represents the proportion of the k-th failure mode in automated testing and manual testing, respectively. max This represents the maximum deviation in the proportion distribution of each failure mode.

[0230] In one possible implementation, a human-like automatic testing system for power tools based on a six-dimensional force sensor is provided, the system comprising:

[0231] Robotic arms are used to hold power tools and perform automated testing actions.

[0232] A six-dimensional force sensor is installed between the end flange and the gripper of the robotic arm for real-time force data acquisition;

[0233] Depth vision sensor, used to acquire depth image data of the test surface;

[0234] The upper-level planning unit is used to construct a physical semantic map of the test scenario based on multimodal perception data, convert the physical semantic map into a physical semantic feature vector and input it into a pre-trained policy model to generate the motion trajectory fragments of the robotic arm.

[0235] The lower-level execution unit includes a force-position hybrid controller, which is used to perform force closed-loop correction on the motion trajectory segment based on the force sensing data, and drive the robotic arm to perform the automated test action.

[0236] It should be noted that the upper-level planning unit and the lower-level execution unit constitute a hierarchical control architecture, corresponding to the policy inference layer and the force control execution layer, respectively. The upper-level planning unit receives multimodal perception data from a six-dimensional force sensor, a depth vision sensor, and a robotic arm encoder. It constructs a physical semantic graph through a physical semantic encoder and converts it into a physical semantic feature vector. This physical semantic feature vector is then input into a pre-trained policy model, outputting a motion trajectory segment. The force-position hybrid controller of the lower-level execution unit operates at a higher frequency, interpolating and tracking the motion trajectory segment and performing force closed-loop correction before driving the robotic arm to execute.

[0237] It is understood that the system assigns each step in the aforementioned method embodiment to the corresponding hardware unit for execution, realizing a complete closed loop from perception and planning to execution. The robotic arm serves as the execution subject, the six-dimensional force sensor serves as the force perception core, the depth vision sensor serves as the visual perception core, the upper-level planning unit serves as the strategy reasoning core, and the lower-level execution unit serves as the real-time control core. The five components work together to achieve human-like automated testing.

[0238] In practical implementation, the robotic arm can be a six-axis industrial robot. Its end effector flange is fitted with a fixture via a specialized adapter. A six-dimensional force sensor is installed between the fixture and the end effector flange, ensuring that all contact forces and torques during tool testing are transmitted and collected through the six-dimensional force sensor. A depth vision sensor is installed near the end effector of the robotic arm or at a fixed position in the work area, covering the surface to be tested. The upper-level planning unit can be implemented by an industrial control computer or embedded computing platform, performing computational tasks such as strategy model inference and physical semantic graph construction. The lower-level execution unit can be implemented by a real-time controller, running a force-position hybrid controller at a frequency of 1kHz or higher, receiving motion trajectory segments output from the upper layer and performing force closed-loop correction. The upper-level planning unit and the lower-level execution unit exchange data via shared memory or a real-time communication bus, ensuring low-latency instruction issuance and status feedback.

[0239] Furthermore, the upper-level planning unit is also configured to: calculate the slope angle based on the test surface normal vector perceived by the depth vision sensor before each chisel strike, dynamically plan the attack angle of the end of the robotic arm so that the angle between the tool axis and the test surface normal is maintained at 0°-15°, and plan the chisel strike position so that the distance between adjacent chisel strikes is not less than 20mm.

[0240] In the embodiments described in this specification, before each chiseling action, the upper-level planning unit acquires a depth image of the current test surface from a depth vision sensor, extracts the contact area normal vector nsurf through RANSAC plane segmentation, and calculates the slope angle as the angle between the normal vector and the opposite direction of gravity. Based on the slope angle, the upper-level planning unit adjusts the tool posture increment Δq in the motion trajectory segment output by the strategy model, so that the angle between the tool axis and the test surface normal is kept within the range of 0°-15°, replicating the behavior of adjusting the chiseling posture according to the ground tilt in manual operation. At the same time, the upper-level planning unit maintains the record of the landing point position of the chiseling that has been performed. When planning a new landing point, it calculates the Euclidean distance to the nearest neighboring existing landing point. When the distance is less than 20mm, it re-searches for a position that meets the spacing constraint around the new landing point, avoiding test result deviation caused by repeated chiseling in the same area.

[0241] In one possible implementation, this application embodiment also provides an electronic device, including a processor and a memory, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps of the above-described anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, or to implement the steps of the above-described training method.

[0242] It should be noted that the electronic device can be a hardware platform with data processing and program execution capabilities, such as an industrial control computer, server, cloud platform computing node, or embedded computing device. The processor can be a general-purpose processor such as a CPU, or a dedicated accelerated processor such as a GPU, NPU, or FPGA, used to perform computational tasks such as policy model inference, physical semantic graph construction, physical constraint checking and correction, and force-position hybrid controller operations. The memory is used to store policy model parameters, training datasets, physical constraint configuration parameters, and computer program code for executing the aforementioned method steps.

[0243] It is understandable that when the electronic device executes the steps of the above-mentioned automated testing method, it acts as a computing platform on the online testing execution side, running real-time tasks such as policy model inference and force-position hybrid control; when executing the steps of the above-mentioned training method, it acts as a computing platform on the offline training side, running batch computing tasks such as policy model training, physical intent discriminator training, and bootstrapping data augmentation. Both execution modes can be deployed on the same electronic device or on different electronic devices.

[0244] In a specific implementation, the memory of the electronic device includes at least one computer-readable storage medium, and the computer program is organized into one or more program modules, which include: a multimodal perception data acquisition module for synchronously reading data from a six-dimensional force sensor, a depth vision sensor, and a robotic arm encoder; a physical semantic graph construction module for constructing a physical semantic graph based on the perception data and quantizing it; a policy model inference module for inputting physical semantic feature vectors into a policy model to generate motion trajectory fragments; a physical constraint correction module for performing physical constraint checks and consistency corrections on the motion trajectory fragments; a force-position hybrid control module for performing force closed-loop corrections on the motion trajectory fragments based on force perception data; and a training module for performing adversarial training between the policy model and the physical intent discriminator and bootstrapping data augmentation.

[0245] In one possible implementation, this application embodiment also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the above-described anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor, or implements the steps of the above-described training method.

[0246] It should be noted that the computer-readable storage medium can be a non-volatile storage medium, including but not limited to: read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), flash memory, hard disk, solid-state drive (SSD), optical disk (CD-ROM / DVD-ROM), or magnetic tape. The computer program can be written using any combination of one or more programming languages, including object-oriented programming languages ​​such as Python and C++, and procedural programming languages ​​such as C.

[0247] It is understood that the computer program stored on the computer-readable storage medium implements all the technical solutions of the foregoing method embodiments. When the computer program is executed by the processor of the electronic device, the electronic device is able to perform anthropomorphic automatic testing of power tools or strategy model training according to the steps of the foregoing method embodiments. The computer program can be divided into code segments corresponding to the program modules in the above-described electronic device embodiments, and each code segment implements the function of the corresponding module.

[0248] In a specific implementation, the computer program includes at least the following instructions: instructions for acquiring multimodal perception data, including force data, depth image data, and pose data; instructions for constructing a physical semantic graph based on the multimodal perception data and converting it into a physical semantic feature vector; instructions for inputting the physical semantic feature vector into a policy model to generate motion trajectory fragments; instructions for performing consistency correction on the motion trajectory fragments based on physical constraints; instructions for sending the motion trajectory fragments to a force-position hybrid controller for force closed-loop correction; and instructions for training the policy model and the physical intent discriminator, including instructions for adversarial training, bootstrap data augmentation, and fine-tuning effectiveness fallback.

[0249] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention shall still fall within the scope of the present invention.

Claims

1. A human-like automatic testing method for power tools based on a six-dimensional force sensor, characterized in that, include: Multimodal perception data is acquired during the testing process of a robotic arm gripping a power tool. The multimodal perception data includes force data, depth image data, and pose data. Based on the multimodal perception data, the terrain features of the target test surface are extracted, and a physical semantic map of the test scene is constructed in combination with preset physical rules. The physical semantic graph is converted into a physical semantic feature vector, and the physical semantic feature vector is input into a pre-trained policy model to generate the motion trajectory fragment of the robotic arm; wherein, the policy model performs consistency correction on the motion trajectory fragment based on physical constraints during the training and generation process; The motion trajectory segment is sent to the force-position hybrid controller, which performs force closed-loop correction based on the force data, and drives the robotic arm to clamp the power tool to perform automated testing actions.

2. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 1, characterized in that, The physical semantic graph includes a set of nodes and a set of edges; the steps of extracting terrain features of the target test surface based on the multimodal perception data and constructing a physical semantic graph of the test scene in combination with preset physical rules specifically include: The depth image data is mapped to the terrain features using a point cloud encoder; Based on the pose data, force data, and terrain features, the node set is constructed, which includes tool chisel head nodes, test surface nodes, obstacle nodes, and robotic arm end effector nodes. The edge set is constructed based on the contact force transmission relationship, kinematic constraint relationship and collision constraint relationship between the nodes in the node set.

3. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 2, characterized in that, The step of converting the physical semantic graph into a physical semantic feature vector specifically includes: The state vectors of each node in the node set are concatenated into a node feature sequence in a preset order, wherein the state vectors include at least one of the corresponding node's position, velocity, force, or curvature. The attribute vectors of each edge in the edge set are concatenated into an edge feature sequence in a preset order, wherein the attribute vectors include at least one of the contact force, fixed offset or distance of the corresponding edge; The node feature sequence and the edge feature sequence are concatenated to form the physical semantic feature vector, and the input of the policy model includes at least the physical semantic feature vector.

4. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 1, characterized in that, The strategy model's step of performing consistency correction on the motion trajectory segments based on physical constraints during training and generation specifically includes: During the generation of the motion trajectory segment by the strategy model, physical constraint checks are performed in real time; the physical constraint checks include force-motion consistency constraints, energy conservation constraints, and contact stability constraints. When the motion trajectory segment is detected to violate any of the physical constraint checks, the motion trajectory segment is physically consistent by projecting it into the direction that satisfies the constraint, scaling it, or filtering and smoothing it.

5. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 4, characterized in that, The method for checking the force-motion consistency constraint is as follows: based on the equivalent mass and equivalent damping coefficient of the power tool, verify the mechanical balance relationship between the generated velocity, acceleration and contact force, wherein the equivalent mass includes the converted value of the end inertia of the robotic arm, and the equivalent damping coefficient is determined based on the dynamic characteristics of the robotic arm and the power tool. The energy conservation constraint is checked by verifying the energy input and output balance relationship based on the input energy of the robotic arm, the impact energy input by the power tool itself, the kinetic energy of the power tool, the damping dissipation energy, and the breaking energy of the test medium, wherein the breaking energy is estimated based on the penetration depth and the hardness of the medium. The contact stability constraint is checked by verifying that the generated normal contact force is positive and that the time change rate of the normal contact force is within a preset range.

6. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 5, characterized in that, The training loss function of the policy model includes a basic generation loss and a physical constraint violation penalty term, wherein the physical constraint violation penalty term is a weighted sum of squares of the violation amount of each physical constraint and the corresponding tolerance threshold. During training, the weights of each loss term are dynamically adjusted. In the early stage of training, the basic generation loss is used as the main factor. As training progresses, the weight of the physical constraint violation penalty term is gradually increased until the weights of the physical constraint violation penalty term and the basic generation loss reach a preset ratio. When the strategy model adopts a diffusion strategy as the backbone network, the physical constraint check and correction are performed after each denoising step. When the strategy model uses an action block transformer as the backbone network, the physical constraint check and correction are performed on the action blocks output by the action block transformer decoder.

7. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 1, characterized in that, The strategy model incorporates a physical intent discriminator for adversarial training during the training phase; the training steps for the physical intent discriminator specifically include: Obtain the sequence of physical semantic feature vectors within a time window; Based on the physical semantic feature vector sequence, the contact force-displacement stiffness curve, impedance parameter variation mode, and energy dissipation trajectory are calculated as physical intent features. Based on the physical intent features, the style of the action trajectory fragment is determined to guide the parameter update of the policy model. The style includes the human expert demonstration style and the policy generation style.

8. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 7, characterized in that, The physical intent discriminator is constructed using a temporal convolutional network with a self-attention mechanism. It takes the physical semantic feature vector sequence as input and outputs the probability that the action trajectory segment belongs to the human expert demonstration style. The physical intent discriminator uses manually generated demonstration data as positive samples and early policy-generated trajectories and violation trajectories constructed by applying physical perturbations to the demonstration data as negative samples. After offline training until convergence, the parameters are frozen and used as a trajectory scorer to give expert style scores to the action trajectory segments output by the policy model. The expert style scores are used as a scoring weighting term in the loss function of the policy model or as a screening condition for high-confidence training samples.

9. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 7, characterized in that, The policy model employs a bootstrap data augmentation mechanism during training, specifically including: In each training round, the policy model generates candidate action trajectory fragments based on the current policy parameters; The physical intent discriminator is used to perform expert style scoring on the candidate action trajectory segments, and candidate trajectories with expert style scores higher than a preset threshold are selected as style qualified samples. The physical constraint deviation of the style qualified samples is calculated based on the physical constraint conditions, and the style qualified samples whose physical constraint deviation meets the preset conditions are selected as high confidence samples. The high-confidence samples are automatically added to the training dataset of the policy model, forming a bootstrapping loop of policy generation, dual screening, and data expansion, so as to continuously expand the training data under the condition of limited manual demonstration data.

10. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 9, characterized in that, When the bootstrapping loop iterates until the physical constraint violation rate on the validation set and the expert style score of the physical intent discriminator reach a preset standard and no longer improve, the training of the policy model is completed. After training, the strategy model is deployed to the robotic arm and tested a preset number of times on a real test medium, while simultaneously collecting real physical feedback data of the robotic arm on the real test medium. The strategy model is fine-tuned using the real physical feedback data at a learning rate lower than that used during the training phase, so that the parameters of the physical constraints in the strategy model are adapted to the actual physical characteristics of the real test medium.

11. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 1, characterized in that, The steps of sending the motion trajectory segment to the force-position hybrid controller, and using the force-position hybrid controller to perform force closed-loop correction based on the force data, and driving the robotic arm to grip the power tool to perform automated testing actions, specifically include: The motion trajectory segment is interpolated into a servo command, the desired contact force is tracked through the force control channel of the force-position hybrid controller, and the lateral motion trajectory is executed through the position control channel. The system monitors forces or torques in all directions collected by the six-dimensional force sensor in real time. When an abnormal force exceeding the threshold of normal operating conditions is detected, a compliant response mechanism is triggered in stages according to the degree of abnormality.

12. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 11, characterized in that, The tiered triggering of the adaptive response mechanism based on the degree of anomaly includes: When a Level 1 anomaly is detected, the stiffness parameter of the force-position hybrid controller is reduced without interrupting the test; the Level 1 anomaly is defined as exceeding the normal operating threshold by less than 20%. When a level 2 anomaly is detected, the feed rate is reduced and admittance control is activated to generate elastic yield; the level 2 anomaly is defined as exceeding the normal operating threshold by 20% to 50%. When a Level 3 anomaly is detected, the chiseling action is paused. After a yielding displacement is generated in the direction of the abnormal force and the tool posture is adjusted, an attempt is made to resume the operation at a reduced feed rate. The Level 3 anomaly is defined as exceeding the normal operating condition threshold by more than 50%. Once the abnormal force is eliminated, the normal testing state will automatically resume.

13. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 11, characterized in that, Before each chisel strike, the normal vector of the test surface is sensed based on the depth image data and the slope angle is calculated. The attack angle of the end of the robotic arm is dynamically planned so that the angle between the tool axis and the normal of the test surface is kept between 0° and 15°. The chisel strike position is planned so that the distance between adjacent chisel strikes is not less than 20mm.

14. The anthropomorphic automatic testing method for power tools based on a six-dimensional force sensor according to claim 1, characterized in that, The tool life distribution and failure mode distribution obtained from automated testing are statistically compared with the results of manual testing of the same batch of tools. The life comparison is deemed qualified when the mean deviation rate of life is not greater than 15% and the standard deviation rate of life is not greater than 20%. The failure mode comparison is deemed qualified when the maximum deviation of the proportion distribution of each failure mode is not greater than 10%. If the representativeness requirement is not met, supplementary manual demonstration data is used to retrain the strategy model.

Images