Training set optimization method based on abnormal collision detection and related device

Abstract

This application provides a training set optimization method and related apparatus based on abnormal collision detection. The method includes: acquiring a reference dataset and task feature vector corresponding to the robot in a robot operation scenario; concatenating the joint state sequence and task feature vector dimensionally to obtain an input vector; inputting the input vector into a prediction network corresponding to the robot to obtain a predicted current sequence; determining a current deviation sequence based on the actual current sequence and the predicted current sequence; performing dual-threshold anomaly judgment on the current deviation sequence to obtain a reference collision judgment result; evaluating the confidence of the reference collision judgment result and filtering it to obtain a target collision judgment result; and removing data from the reference dataset based on the target collision judgment result to obtain a target dataset. By identifying and removing collision anomaly data through current prediction, dual-threshold judgment, and confidence evaluation, a high-quality training set is constructed.

Description

Technical Field

[0001] This application relates to the field of training set optimization technology, and in particular to a training set optimization method and related apparatus based on abnormal collision detection. Background Technology

[0002] In real-world data acquisition scenarios involving ultra-long-range robot operation, components such as robot arms are prone to accidental collisions with obstacles like tables, chairs, and walls due to remote control delays and blind spots in environmental perception, resulting in abnormal collision events. Existing data acquisition systems lack effective collision detection mechanisms and cannot promptly identify such events, instead continuously collecting data such as joint states and motor currents during the collision process.

[0003] Among them, the data under collision conditions has abnormal fluctuations and low availability and quality. If it is directly used to train the robot control model, it will cause the model to learn the wrong operation mode, which will make it difficult for the robot's actual operation capability to meet the user's expectations, and seriously restrict the reliability and practicality of the robot's ultra-long-range operation.

[0004] Therefore, how to construct a high-quality training set free from abnormal interference, thereby improving the training effect and practical operation capability of robot models, is an urgent problem to be solved. Summary of the Invention

[0005] This application provides a training set optimization method and related apparatus based on abnormal collision detection. By fusing robot joint states and task features to construct an input vector, a predicted current sequence is output based on a prediction network, and collision anomalies are accurately identified by combining dual-threshold anomaly judgment and confidence assessment. Collision anomaly data in the reference dataset are removed, thereby constructing a high-quality training set without abnormal interference, and thus improving the training effect and actual operation capability of the robot model.

[0006] In a first aspect, embodiments of this application provide a training set optimization method based on abnormal collision detection, the method comprising: Obtain the reference dataset and task feature vector corresponding to the robot in the robot operation scenario; the reference dataset includes joint state sequences and actual current sequences. The joint state sequence and the task feature vector are concatenated dimensionally to obtain the input vector; The input vector is input into the prediction network corresponding to the robot to obtain the predicted current sequence; the prediction network is a time-series current prediction model adapted to the robot. Determine the current deviation sequence based on the actual current sequence and the predicted current sequence; The current deviation sequence is subjected to dual threshold anomaly determination to obtain a reference collision determination result; The confidence level of the reference collision determination results is evaluated and filtered to obtain the target collision determination results; The reference dataset is purged based on the target collision determination result to obtain the target dataset; the target dataset is used for iterative training of the prediction network.

[0007] Secondly, embodiments of this application provide a training set optimization device based on anomaly collision detection. The device includes an acquisition module, a dimension splicing module, a prediction module, a determination module, an anomaly judgment module, a filtering module, and a data removal module, wherein: The acquisition module is used to acquire a reference dataset and task feature vector corresponding to the robot in the robot operation scenario; the reference dataset includes joint state sequences and actual current sequences; The dimension concatenation module is used to concatenate the joint state sequence and the task feature vector dimensionally to obtain the input vector. The prediction module is used to input the input vector into the prediction network corresponding to the robot to obtain the predicted current sequence; the prediction network is a time-series current prediction model adapted to the robot. The determining module is used to determine a current deviation sequence based on the actual current sequence and the predicted current sequence; The anomaly determination module is used to perform dual-threshold anomaly determination on the current deviation sequence to obtain a reference collision determination result; The filtering module is used to evaluate the confidence level of the reference collision determination results and filter them to obtain the target collision determination results; The data removal module is used to remove data from the reference dataset based on the target collision determination result to obtain the target dataset; the target dataset is used to iteratively train the prediction network.

[0008] Thirdly, embodiments of this application provide an electronic device, including a processor, a memory, a communication interface, and one or more programs, wherein the one or more programs are stored in the memory and configured to be executed by the processor, and the programs include instructions for performing steps in any method of the first aspect of this application.

[0009] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program for electronic data interchange, wherein the computer program causes a computer to perform some or all of the steps described in any method of the first aspect of this application.

[0010] Fifthly, embodiments of this application provide a computer program product, wherein the computer program product includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps described in any method of the first aspect of this application. The computer program product may be a software installation package.

[0011] By implementing the embodiments of this application, robot joint states and task characteristics can be fused to construct an input vector, a predicted current sequence can be output based on the prediction network, and collision anomalies can be accurately identified by combining dual threshold anomaly judgment and confidence assessment. Collision anomaly data in the reference dataset can be removed, thereby constructing a high-quality training set without anomaly interference, and thus improving the training effect and actual operation capability of the robot model. Attached Figure Description

[0012] To more clearly illustrate the technical solutions of the embodiments of this application, the drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a system architecture diagram of a training set optimization system provided in an embodiment of this application; Figure 2 This is a schematic diagram of the structure of an algorithm layer provided in an embodiment of this application; Figure 3 This is an application scenario diagram of a training set optimization system provided in an embodiment of this application; Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application; Figure 5 This is a flowchart illustrating a training set optimization method based on abnormal collision detection provided in an embodiment of this application; Figure 6 This is a schematic diagram of a process for determining a target collision determination result provided in an embodiment of this application; Figure 7 This is a schematic diagram of a process for performing a data removal operation on a reference dataset, provided in an embodiment of this application; Figure 8 This is a block diagram of the functional modules of a training set optimization device based on abnormal collision detection provided in an embodiment of this application. Detailed Implementation

[0014] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present application, and not all embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present application.

[0015] The terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0016] It should be understood that the term "and / or" in this document is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this document indicates that the preceding and following related objects are in an "or" relationship. In the embodiments of this application, "multiple" refers to two or more.

[0017] In the embodiments of this application, "at least one item" or its similar expression refers to any combination of these items, including any combination of a single item or a plurality of items. "One or more" means one or more, while "multiple" means two or more. For example, "at least one item" of a, b, or c can represent the following seven cases: a, b, c; a and b; a and c; b and c; a, b, and c. Each of a, b, and c can be an element or a set containing one or more elements.

[0018] In this application, the term "connection" refers to various connection methods, such as direct connection or indirect connection, to achieve communication between devices. This application does not impose any limitations on this.

[0019] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0020] In real-world data acquisition scenarios involving ultra-long-range robot operation, components such as robot arms are prone to accidental collisions with obstacles like tables, chairs, and walls due to remote control delays and blind spots in environmental perception, resulting in abnormal collision events. Existing data acquisition systems lack effective collision detection mechanisms and cannot promptly identify such events, instead continuously collecting data such as joint states and motor currents during the collision process.

[0021] Among them, the data under collision conditions has abnormal fluctuations and low availability and quality. If it is directly used to train the robot control model, it will cause the model to learn the wrong operation mode, which will make it difficult for the robot's actual operation capability to meet the user's expectations, and seriously restrict the reliability and practicality of the robot's ultra-long-range operation.

[0022] Therefore, how to construct a high-quality training set free from abnormal interference, thereby improving the training effect and practical operation capability of robot models, is an urgent problem to be solved.

[0023] To address the aforementioned issues, this application provides a training set optimization method and related apparatus based on abnormal collision detection. The method involves acquiring a reference dataset and task feature vector corresponding to the robot in a robot operation scenario. The reference dataset includes joint state sequences and actual current sequences. The joint state sequences and task feature vectors are concatenated dimensionally to obtain an input vector. This input vector is then input into a prediction network corresponding to the robot to obtain a predicted current sequence. The prediction network is a time-series current prediction model adapted to the robot. A current deviation sequence is determined based on the actual current sequence and the predicted current sequence. A dual-threshold anomaly determination is performed on the current deviation sequence to obtain a reference collision determination result. The confidence level of the reference collision determination result is evaluated and filtered to obtain a target collision determination result. Data is removed from the reference dataset based on the target collision determination result to obtain a target dataset. The target dataset is used for iterative training of the prediction network.

[0024] It is evident that by fusing robot joint states and task features to construct input vectors, outputting predicted current sequences based on prediction networks, and combining dual-threshold anomaly detection and confidence assessment to accurately identify collision anomalies, collision anomaly data in the reference dataset are eliminated, thereby constructing a high-quality training set free from anomaly interference, which in turn improves the training effect and actual operation capability of the robot model.

[0025] For easier understanding, please refer to Figure 1 , Figure 1 This is a system architecture diagram of a training set optimization system provided in an embodiment of this application. The training set optimization system includes a perception layer, an algorithm layer, a control layer, and a management layer. The modules at each level work together to realize abnormal collision detection of the robot and optimization of training set data.

[0026] The perception layer collects raw joint position data, raw joint velocity data, and raw joint current data from the robot's operation scenario, while also acquiring the corresponding task target parameters. It uses hardware timestamps to align frames of the joint encoder, force / torque sensor, and RGBD camera (which collects environmental point cloud data), ensuring temporal consistency across all data types. Then, Kalman filtering is performed on the raw joint position / velocity data to obtain reference joint position / velocity data, which is then integrated into a joint state sequence. Median filtering is performed on the raw joint current data to obtain the actual current sequence. The RGBD point cloud data is downsampled and obstacle segmented to generate an environmental particle set, while task target location and task type labels are extracted to construct a task feature vector.

[0027] For easier understanding, please refer to Figure 2 , Figure 2 This is a schematic diagram of the structure of an algorithm layer provided in an embodiment of this application. The algorithm layer includes a task feature embedding module, a neural network prediction module, an anomaly detection module, a confidence evaluation module, and a collision quantization module.

[0028] The task feature embedding module concatenates the joint state sequence with the task feature vector to generate the input vector for the prediction network.

[0029] The neural network prediction module loads a time-series current prediction model adapted to the robot, takes the input vector as input, and outputs a predicted current sequence. This time-series current prediction model incorporates a Multi-Layer Perceptron (MLP) and a Long Short-Term Memory (LSTM) network. The MLP (with hidden layer dimensions of 256 / 128 / 64) can extract static features (such as task type and target location), while the LSTM (with a hidden layer dimension of 128) can capture the temporal dependencies of joint states. The output layer of the time-series current prediction model uses linear activation, which can directly predict the current of each joint, thus obtaining the predicted current sequence.

[0030] The anomaly detection module calculates the current deviation sequence corresponding to the actual current sequence and the predicted current sequence; constructs a sliding window based on the preset sliding step size and window size; calculates the mean and standard deviation of the current deviation within the sliding window and determines the first threshold; combines the motor torque constant and the maximum load torque of the joint to determine the second threshold; and determines the collision time through double threshold verification to obtain the reference collision judgment result (i.e., a collision times).

[0031] The confidence assessment module retrieves the force / torque sensor data and the robot Jacobian matrix corresponding to the collision moment, evaluates the confidence of the collision moment based on the confidence calculation formula, filters out the collision moments that are in the confidence interval, and obtains the target collision judgment result (i.e., b collision moments).

[0032] The collision quantization module calculates the collision point location based on the principle of distance minimization, combining environmental particle sets, sensor data, Jacobian matrix, current deviation, and motor torque constant. It also calculates the magnitude and direction of the collision force using the pseudo-inverse of the Jacobian matrix and the torque constant, providing parameter support for obstacle avoidance strategy generation.

[0033] As can be seen, by embedding task features, predicting time-series currents, determining collisions with dual thresholds, and quantifying collisions with confidence assessment and collision fusion of multi-source data, accurate identification of robot collision events and calculation of core parameters are achieved, providing a reliable decision-making basis for training set optimization and obstacle avoidance strategy generation.

[0034] The control layer can determine the target joint associated with the collision based on the magnitude and direction of the collision force, retrieve the first stiffness coefficient of the target joint, and adjust it according to a preset second ratio to obtain the second stiffness coefficient. Using the collision point position as the obstacle avoidance constraint, the model predictive control algorithm is used to replan the joint trajectory within a preset time after the collision moment to generate the target joint trajectory. Combining the second stiffness coefficient and the target joint trajectory, an obstacle avoidance operation strategy including stiffness adjustment and trajectory execution is generated to drive the robot to perform obstacle avoidance actions.

[0035] Specifically, the management team can define an invalid data filtering interval based on b collision times in the target collision determination results and the sliding window size, and mark invalid data in the reference dataset; remove data within the invalid data filtering interval in the reference dataset to obtain the valid dataset; classify and organize the valid dataset according to the task type in the task feature vector to generate the target dataset for iterative training of the prediction network, and complete the optimization and update of the training set.

[0036] It is evident that by using a collision detection and confidence screening mechanism based on multi-source sensor fusion to accurately eliminate invalid data and constructing a high-quality target dataset by combining task feature classification, it is possible to achieve iterative optimization of the prediction network and synergistic improvement of robot collision and obstacle avoidance.

[0037] For easier understanding, please refer to Figure 3 , Figure 3This is an application scenario diagram of a training set optimization system provided in an embodiment of this application. The robot, as the data source, provides the training set optimization system with a reference dataset containing joint state sequences, actual current sequences, and task feature vectors containing information such as task type and task objectives. The training set optimization system, through collaborative computation of the perception layer, algorithm layer, control layer, and management layer, completes processes such as current prediction, dual-threshold collision detection, confidence assessment, invalid data removal, and valid data classification and organization. Finally, it outputs an optimized target dataset, which contains only valid data from the robot's normal operation and can be directly used for iterative training of the robot's prediction network, achieving continuous improvement in model performance and ensuring the robot's safe operation.

[0038] The following is combined Figure 4 The electronic devices in the embodiments of this application will be described. Figure 4 This is a schematic diagram of the structure of an electronic device provided in an embodiment of this application, such as... Figure 4 As shown, the electronic device includes one or more processors, a memory, a communication interface, and one or more programs. The processor is connected to the memory and the communication interface via an internal communication bus.

[0039] The processor can be used for: Obtain the reference dataset and task feature vector corresponding to the robot in the robot operation scenario; the reference dataset includes joint state sequences and actual current sequences. The joint state sequence and the task feature vector are concatenated dimensionally to obtain the input vector; The input vector is input into the prediction network corresponding to the robot to obtain the predicted current sequence; the prediction network is a time-series current prediction model adapted to the robot. Determine the current deviation sequence based on the actual current sequence and the predicted current sequence; The current deviation sequence is subjected to dual threshold anomaly determination to obtain a reference collision determination result; The confidence level of the reference collision determination results is evaluated and filtered to obtain the target collision determination results; The reference dataset is purged based on the target collision determination result to obtain the target dataset; the target dataset is used for iterative training of the prediction network.

[0040] The one or more programs are stored in the aforementioned memory and configured to be executed by the aforementioned processor, and the one or more programs include instructions for performing any step in the above method embodiments.

[0041] The processor can be a central processing unit (CPU), a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), a field-programmable gate array (FPGA), or other programmable logic devices, transistor logic devices, hardware components, or any combination thereof. It can implement or execute the various exemplary logic blocks, cells, and circuits described in conjunction with the disclosure of this application. The processor can also be a combination that implements computational functions, such as a combination of one or more microprocessors, a combination of a DSP and a microprocessor, etc. The communication unit can be a communication interface, transceiver, transceiver circuit, etc., and the storage unit can be a memory.

[0042] The memory can be volatile or non-volatile, or a combination of both. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), used as an external cache. By way of example, but not limitation, many forms of random access memory (RAM) are available, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate synchronous DRAM (DDR SDRAM), enhanced synchronous DRAM (ESDRAM), synchronous linked DRAM (SLDRAM), and direct rambus RAM (DR RAM).

[0043] It is understood that the electronic device may include more or fewer structural elements than those shown in the block diagram above, such as a power module, physical buttons, a Wi-Fi module, a speaker, a Bluetooth module, sensors, a display module, etc., without limitation. It is understood that the electronic device may incorporate elements such as... Figure 1 The system architecture described above.

[0044] After understanding the software and hardware architecture of this application, the following will be combined with... Figure 5 This application describes a training set optimization method based on abnormal collision detection in its embodiments. Figure 5 This is a flowchart illustrating a training set optimization method based on anomaly collision detection provided in an embodiment of this application, specifically including the following steps: Step S501: Obtain the reference dataset and task feature vector corresponding to the robot in the robot operation scenario.

[0045] The reference dataset includes joint state sequences and actual current sequences. The specific steps for obtaining the reference dataset and task feature vector corresponding to the robot in the robot operation scenario include: A1. Obtain the original joint position data, original joint velocity data, original joint current data, and task target parameters corresponding to the robot in the robot operation scenario; the task target parameters include: task target position and task type label; A2. Perform Kalman filtering on the original joint position data and the original joint velocity data respectively to obtain reference joint position data and reference joint velocity data; A3. Determine the joint state sequence based on the reference joint position data and the reference joint velocity data; A4. Perform median filtering on the original joint current data to obtain the actual current sequence; A5. Determine the task feature vector based on the task target location and the task type label.

[0046] In a specific embodiment, firstly, through the robot's multi-source sensor data and task scheduling system, the original joint position data (collected by the joint encoder, reflecting the real-time angle of each joint), the original joint velocity data (which can be directly collected by the position data differential or velocity sensor, characterizing the joint rotation rate), the original joint current data (collected by the current sensor, used to reflect the load state of the joint motor), and the task target parameters (including the task target position and task type label, issued by the task scheduling system, used to distinguish different operation scenarios) are acquired synchronously.

[0047] Then, Kalman filtering is applied to the original joint position data and original joint velocity data to obtain reference joint position data and reference joint velocity data, effectively filtering out random noise introduced by factors such as long-range operation communication delays and mechanical vibrations. The reference joint position data and reference joint velocity data are then dimensionally integrated according to the time sampling order to obtain a joint state sequence. For example, for a 6-axis robot, the joint state at each moment can be represented as a vector containing three position components and three velocity components. Arranging the joint state vectors at all moments along the time axis forms a complete joint state sequence. Finally, median filtering is used to denoise the pulse noise (such as motor start-stop spikes) and random noise in the original joint current data, and the data is then integrated to obtain the actual current sequence.

[0048] Next, the task type labels are encoded (e.g., 'grabbing' is labeled as '1', 'transporting' as '2', which can be done using one-hot encoding or numerical mapping). The task target position is normalized, mapping the position component to the [0, 1] interval to eliminate the dimensional influence caused by differences in the range of motion of different joints. Then, the encoded task type features and the normalized target position features are integrated to form a basic feature vector. Finally, the basic feature vector is expanded to a preset dimension (e.g., 128 dimensions) through fully connected layer mapping or zero-padding to obtain the final task feature vector.

[0049] It is evident that by preprocessing and standardizing the features of multi-source sensor data, a highly reliable and consistent input data foundation is provided for subsequent current prediction, collision detection, and training set optimization.

[0050] Step S502: Concatenate the joint state sequence and the task feature vector dimensionally to obtain the input vector.

[0051] Specifically, the joint state sequence can be vectorized to obtain a joint state vector (which can be 6-dimensional), and then fused with the task feature vector (which can be 128-dimensional) using a tail concatenation method to obtain the input vector.

[0052] Step S503: Input the input vector into the prediction network corresponding to the robot to obtain the predicted current sequence.

[0053] The prediction network is a time-series current prediction model adapted to the robot. Input vectors are sequentially fed into the pre-trained prediction network in time-step order. This prediction network is a time-series current prediction model optimized for the robot's joint structure, motor characteristics, and typical operational tasks. It incorporates a time-series feature extraction module (such as LSTM, GRU, or a Transformer encoder) to synchronously learn the robot's motion state features and task scenario features from the input vectors. Through forward inference, the prediction network outputs the predicted current values ​​of the joint motors corresponding one-to-one with the input time steps. By integrating the predicted current values ​​from all time steps in time order, a complete predicted current sequence can be obtained.

[0054] Step S504: Determine the current deviation sequence based on the actual current sequence and the predicted current sequence.

[0055] Specifically, the absolute difference between the actual current sequence and the predicted current sequence is calculated point by point at the same time step to obtain the current deviation value at each time step. Then, the current deviation values ​​of all time steps are integrated in chronological order to obtain the current deviation sequence.

[0056] Step S505: Perform dual-threshold anomaly determination on the current deviation sequence to obtain the reference collision determination result.

[0057] The current deviation sequence includes m current deviations, where m is an integer greater than 1; the reference collision determination result includes a collision times, where a is a positive integer less than or equal to m; the specific steps for performing dual-threshold anomaly determination on the current deviation sequence to obtain the reference collision determination result include: B1. Determine the first sliding window corresponding to the first current deviation according to the preset sliding step size and window size; the first current deviation is any one of the m current deviations; B2. Calculate the first mean and first standard deviation of all current deviations in the first sliding window; B3. Determine the first threshold based on the first mean and the first standard deviation; B4. If the first current deviation is greater than the first threshold, the first predicted torque and the first actual torque corresponding to the first current deviation are calculated according to the preset motor torque constant. B5. Obtain the maximum load torque of the joints corresponding to the robot; B6. Determine the second threshold based on the preset first ratio and the maximum load torque of the joint; B7. If the deviation between the first predicted torque and the first actual torque is greater than the second threshold, then obtain the first moment corresponding to the first current deviation; B8. Determine the first moment as the first collision moment; the first collision moment is the collision moment corresponding to the first current deviation among the a collision moments.

[0058] In a specific embodiment, firstly, the current deviation sequence is traversed using a preset sliding step size (e.g., 1 time step) and window size (e.g., 50 consecutive time steps) as parameters. For the first current deviation (i.e., any current deviation in the current deviation sequence), k consecutive current deviations (k being the window size) centered on (or ending at) the first current deviation are selected to form the first sliding window.

[0059] Then, statistical calculations are performed on all current deviations within the first sliding window to obtain a first mean and a first standard deviation. The first mean reflects the average level of current deviations within the first sliding window, and the first standard deviation reflects the dispersion of the deviations. The combination of these two values ​​characterizes the normal fluctuation range of the current deviations within a local time period. Then, a first threshold is determined based on the first mean and the first standard deviation; for example, the first threshold is set to... ,in, This represents the first mean; This represents the first standard deviation. When the first current deviation... If the first current deviation is initially deemed abnormal, further torque verification is needed to confirm whether it was caused by a collision. Then, based on the preset motor torque constant, the first predicted torque and the first actual torque corresponding to the first current deviation are calculated respectively. The calculation formulas for the first predicted torque and the first actual torque are as follows:

[0060] in, This represents the first predicted torque, i.e. Predicted joint torque at any given moment; It represents the motor torque constant and characterizes the conversion coefficient between current and torque; This represents the first predicted current corresponding to the first current deviation in the predicted current sequence, i.e. Predicted current of the joint motor at any given moment; This represents the first actual torque, i.e. The actual torque of the joint at any given moment; This represents the first actual current corresponding to the first current deviation in the actual current sequence, i.e. The actual current of the joint motor at any given moment.

[0061] It should be noted that if the first current deviation is less than or equal to the first threshold, the first moment corresponding to the first current deviation is determined to be the normal operating moment, and no subsequent operation is performed.

[0062] Next, the maximum load torque of the robot's joints is obtained. This maximum load torque is determined by the joint's motor rated power, reducer transmission ratio, etc., and can be retrieved from the robot's hardware parameter manual. It represents the maximum load torque that the joint can withstand within its safe operating range; exceeding this value is considered an overload anomaly. Then, a second threshold is determined based on a preset first ratio and the maximum load torque of the joint. For example, the second threshold can be set to... The first ratio is preset to 0.1; This represents the maximum load torque of the joint. If the deviation between the first predicted torque and the first actual torque is greater than the second threshold, i.e. If the first current deviation is determined to be caused by a sudden change in joint load, which conforms to the mechanical characteristics of a collision, the timestamp corresponding to the first current deviation is retrieved and recorded as the first moment. Then, the first moment is marked as the first collision moment, and all current deviations in the current deviation sequence are traversed, and steps B1-B8 are repeated. Finally, all collision moments are integrated to form a reference collision determination result.

[0063] It is evident that the dual-threshold determination mechanism of sliding window statistical threshold and dynamic torque verification enables rapid and accurate identification of robot collision events, effectively filters out instantaneous noise interference, and improves the robustness and reliability of collision detection.

[0064] Step S506: Confidence assessment and screening of the reference collision determination results are performed to obtain the target collision determination results.

[0065] For easier understanding, please refer to Figure 6 , Figure 6 This is a flowchart illustrating a process for determining a target collision determination result according to an embodiment of this application. The specific steps for evaluating and filtering the reference collision determination results to obtain the target collision determination result include: C1. Obtain the first sensor data corresponding to the first collision moment and the Jacobian matrix corresponding to the robot; the first sensor data includes three-dimensional force data and three-dimensional torque data; C2. Based on the preset confidence calculation formula, determine the first confidence level according to the first sensor data, the Jacobian matrix, the first current deviation and the motor torque constant; C3. If the first confidence level is within a preset confidence level range, then the first collision time is retained in the target collision determination result.

[0066] In a specific embodiment, firstly, the three-dimensional force data and three-dimensional torque data of the robot's end effector at the moment of the first collision are retrieved. This data is collected in real time by the robot's built-in force / torque sensors and can directly reflect the contact mechanics characteristics between the robot and the environment. Simultaneously, the Jacobian matrix of the robot at the moment of the first collision is retrieved. This Jacobian matrix is ​​a mapping matrix between the robot's joint space and Cartesian space, and its elements are determined by the joint angles at the moment of the first collision. It is used to realize the mutual conversion between joint torque and end effector Cartesian force / torque.

[0067] Then, based on the preset confidence level calculation formula, the first confidence level is determined according to the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. The confidence level calculation formula is as follows:

[0068] in, Indicates the first confidence level; This represents the data from the first sensor; Represent the pseudo-inverse of the Jacobian matrix to achieve the mapping from joint space to Cartesian space; Indicates the first current deviation; This represents the motor torque constant.

[0069] Next, a pre-set confidence interval (e.g., [0.7, 1.0]) is obtained. If the first confidence level falls within this interval, the first collision moment is determined to be a valid collision moment and retained in the target collision determination result. If the first confidence level does not fall within this interval, the first collision moment is determined to be a misjudgment moment (e.g., caused by sensor noise) and is not stored in the target collision determination result. Then, all collision moments in the reference collision determination result are iterated over, and steps C1-C3 are repeated to finally obtain the target collision determination result after confidence level filtering.

[0070] As can be seen, by integrating end force / torque sensor data, Jacobian matrix and current deviation confidence assessment mechanism, the initial collision time is re-verified, effectively eliminating false judgments and significantly improving the accuracy and reliability of collision detection results, providing accurate invalid data labeling basis for subsequent training set optimization.

[0071] Step S507: Based on the target collision determination result, remove data from the reference dataset to obtain the target dataset.

[0072] For easier understanding, please refer to Figure 7 , Figure 7This is a flowchart illustrating a data removal operation on a reference dataset provided in an embodiment of this application. The target dataset is used for iterative training of the prediction network. The target collision determination result includes b collision times, where b is a positive integer less than or equal to a. The specific steps for removing data from the reference dataset based on the target collision determination result to obtain the target dataset include: D1. Determine the b invalid data filtering intervals corresponding to the b collision times based on the window size; D2. Remove the data corresponding to the b invalid data filtering intervals in the reference dataset to obtain the valid dataset; D3. Classify and organize the data in the effective dataset according to the task feature vector to obtain the target dataset.

[0073] In a specific embodiment, for each collision moment in the target collision determination result Based on the window size of the sliding window corresponding to the moment of collision. Define the corresponding invalid data filtering interval. For example, the invalid data filtering interval at the moment of collision could be... No specific limitations are made here. It should be noted that, because the current data at the moment of the collision and for a period of time before and after it will be severely distorted by the sudden change in load caused by the collision, it cannot reflect the current characteristics of the robot's normal operation. Therefore, all data in this period (i.e., the invalid data screening interval) should be marked as invalid data.

[0074] Then, iterate through all the data in the reference dataset, batch remove all data within the b invalid data filtering intervals, retain the remaining data that were not included in the invalid data filtering intervals, and reassemble them in chronological order to obtain the valid dataset.

[0075] Next, based on the task type labels (such as grasping, moving, and assembling) in the task feature vector, the effective dataset is divided into multiple sub-datasets, each corresponding to a specific task scenario. Data consistency is verified for each sub-dataset (such as checking whether the data dimensions and time steps are consistent). Then, all the classified sub-datasets are integrated and encapsulated to obtain the target dataset for iterative training of the prediction network.

[0076] It is evident that by removing invalid data and classifying and organizing the data, the reference dataset was accurately cleaned and structurally optimized, resulting in a high-quality target dataset that effectively ensured the accuracy and iterative efficiency of the prediction network training.

[0077] In one possible embodiment, the method further includes the following steps: E1. Obtain the set of environmental particles corresponding to the first collision moment; E2. Based on the preset distance minimization principle, the location of the first collision point is determined according to the environmental particle set, the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. E3. Determine the first collision force corresponding to the first collision point position based on the Jacobian matrix, the motor torque constant, and the first current deviation; E4. Determine the magnitude and direction of the first collision force; E5. Determine the obstacle avoidance strategy of the robot based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force.

[0078] In a specific embodiment, firstly, the environmental particle set collected by the robot at the moment of the first collision is retrieved. ,in, This represents the Nth particle in the environmental particle set, where N is the number of particles, typically between 2000 and 4000. Each particle is a three-dimensional spatial coordinate point, representing a candidate location of an environmental obstacle within the robot's workspace. This environmental particle set is generated from environmental point cloud data collected by a depth camera or LiDAR, after downsampling and noise reduction processing. It covers the potential collision area around the robot's end effector, providing environmental coordinate references for collision point localization.

[0079] Then, based on the preset distance minimization principle, the location of the first collision point is determined according to the environmental particle set, the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. The relevant calculation formula for the location of the first collision point is shown below:

[0080] in, Indicates the location of the first collision point; Indicates the moment of the first collision. The corresponding set of environmental particles contains the coordinates of all candidate obstacles within the workspace; Indicates the moment of the first collision. The transpose of the corresponding Jacobian matrix; Indicates the moment of the first collision. The corresponding current deviation, i.e., the first current deviation; Indicates the moment of the first collision. The corresponding sensor data, i.e., the first sensor data.

[0081] Next, based on the Jacobian matrix, the motor torque constant, and the first current deviation, the first collision force corresponding to the first collision point is determined. The relevant calculation formula for the first collision force is shown below:

[0082] in, This represents the first collision force, which is a three-dimensional vector. Indicates the moment of the first collision. The pseudo-inverse of the corresponding Jacobian matrix.

[0083] Then, calculate the first collision force. The corresponding magnitude of the first collision force And based on the magnitude of the first collision force For the first collision force After normalization, the direction of the first collision force is obtained. The direction of the first collision force is the first collision force. The unit vector.

[0084] Finally, the obstacle avoidance strategy of the robot is determined based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force.

[0085] It is evident that by integrating environmental particle sets, multi-source sensor data, and robot dynamic parameters, the collision point can be accurately located, the collision force characteristics can be quantified, and a targeted obstacle avoidance strategy can be generated. This enables refined perception of robot collision events and safe obstacle avoidance, ensuring operational safety and task continuity.

[0086] The specific steps of determining the robot's obstacle avoidance strategy based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force include: F1. Determine the target joint of the robot that collided with the first collision force based on the magnitude and direction of the first collision force. F2. Obtain the first stiffness coefficient of the target joint; F3. Adjust the first stiffness coefficient according to the preset second ratio to obtain the second stiffness coefficient; F4. Using the position of the first collision point as the obstacle avoidance constraint, a preset model prediction control algorithm is used to replan the joint trajectory within a preset time after the first collision moment to obtain the target joint trajectory. F5. Determine the obstacle avoidance operation strategy based on the second stiffness coefficient and the target joint trajectory; the obstacle avoidance operation strategy is used to control the robot to perform obstacle avoidance operations.

[0087] In a specific embodiment, firstly, a collision force-joint torque transmission model is constructed based on the magnitude and direction of the first collision force. The first collision force is decomposed in reverse along the robot linkage, and the additional torque of each joint is calculated. Then, the joint with the largest additional torque exceeding a preset torque threshold is identified as the target joint where the collision occurs. Next, the inherent mechanical parameters of the target joint, namely the first stiffness coefficient, are retrieved. The first stiffness coefficient is jointly determined by the stiffness of the joint servo motor, the stiffness characteristics of the reducer, and the material properties of the connecting rod. The first stiffness coefficient is then adjusted according to a preset second ratio to obtain the second stiffness coefficient. .in, , This indicates a preset second ratio, which can take the value 0.5, and 0 < 0. <1. It should be noted that the damping coefficient B of the target joint can remain unchanged to reduce rebound after the collision.

[0088] Next, using the location of the first collision point as an obstacle avoidance constraint, a minimum safe distance between the robot's end effector and the collision point is set, and a cost function is constructed with the objectives of maximizing trajectory smoothness and minimizing task path offset. A pre-defined model predictive control algorithm is then used to replan the joint trajectories within a pre-defined timeframe (e.g., 50ms) after the first collision, resulting in the target joint trajectory. .in, Indicates the first The desired angle vectors of each joint of the robot at any given time.

[0089] Finally, based on the second stiffness coefficient and the target joint trajectory, an obstacle avoidance operation strategy is generated, which includes stiffness adjustment instructions, trajectory execution instructions, and real-time monitoring instructions. This obstacle avoidance operation strategy is used to control the robot to perform obstacle avoidance operations.

[0090] It is evident that by locating collision-related joints, dynamically adjusting joint stiffness, and planning obstacle avoidance trajectories, the robot achieves flexible buffering and safe obstacle avoidance after collisions, taking into account operational safety, smooth motion, and task continuity.

[0091] The above primarily describes the solutions of the embodiments of this application from the perspective of the method execution process. It is understood that, in order to achieve the above functions, the electronic device includes corresponding hardware structures and / or software modules for executing each function. Those skilled in the art should readily recognize that, in conjunction with the units and algorithm steps of the various examples described in the embodiments provided herein, this application can be implemented in hardware or a combination of hardware and computer software. Whether a function is executed by hardware or by computer software driving hardware depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0092] This application embodiment can divide the electronic device into functional units according to the above method example. For example, each function can be divided into a separate functional unit, or two or more functions can be integrated into one processing unit. The integrated unit can be implemented in hardware or as a software functional unit. It should be noted that the unit division in this application embodiment is illustrative and only represents one logical functional division. In actual implementation, there may be other division methods.

[0093] When dividing each function into modules according to its corresponding function. Figure 8 This is a functional block diagram of a training set optimization device based on anomaly collision detection provided in an embodiment of this application. The training set optimization device 800 based on anomaly collision detection includes an acquisition module 810, a dimension splicing module 820, a prediction module 830, a determination module 840, an anomaly judgment module 850, a filtering module 860, and a data removal module 870, wherein: The acquisition module 810 is used to acquire a reference dataset and task feature vector corresponding to the robot in the robot operation scenario; the reference dataset includes a joint state sequence and an actual current sequence. The dimension concatenation module 820 is used to concatenate the joint state sequence and the task feature vector dimensionally to obtain the input vector. The prediction module 830 is used to input the input vector into the prediction network corresponding to the robot to obtain the predicted current sequence; the prediction network is a time-series current prediction model adapted to the robot. The determining module 840 is used to determine a current deviation sequence based on the actual current sequence and the predicted current sequence; The anomaly determination module 850 is used to perform dual-threshold anomaly determination on the current deviation sequence to obtain a reference collision determination result. The filtering module 860 is used to evaluate the confidence level of the reference collision determination result and filter it to obtain the target collision determination result. The data removal module 870 is used to remove data from the reference dataset according to the target collision determination result to obtain the target dataset; the target dataset is used to iteratively train the prediction network.

[0094] Optionally, regarding the acquisition of the reference dataset and task feature vector corresponding to the robot in the robot operation scenario, the acquisition module 810 is specifically used for: Obtain the original joint position data, original joint velocity data, original joint current data, and task target parameters corresponding to the robot in the robot operation scenario; the task target parameters include: task target position and task type label; Kalman filtering is performed on the original joint position data and the original joint velocity data respectively to obtain reference joint position data and reference joint velocity data. The joint state sequence is determined based on the reference joint position data and the reference joint velocity data; The original joint current data is subjected to median filtering to obtain the actual current sequence; The task feature vector is determined based on the task target location and the task type label.

[0095] Optionally, the current deviation sequence includes m current deviations, where m is an integer greater than 1; the reference collision determination result includes a collision times, where a is a positive integer less than or equal to m; in terms of performing dual-threshold anomaly determination on the current deviation sequence to obtain the reference collision determination result, the anomaly determination module 850 is specifically used for: The first sliding window corresponding to the first current deviation is determined according to the preset sliding step size and window size; the first current deviation is any one of the m current deviations. Calculate the first mean and first standard deviation of all current deviations in the first sliding window; A first threshold is determined based on the first mean and the first standard deviation; If the first current deviation is greater than the first threshold, then the first predicted torque and the first actual torque corresponding to the first current deviation are calculated according to the preset motor torque constant. Obtain the maximum load torque of the joints corresponding to the robot; The second threshold is determined based on a preset first ratio and the maximum load torque of the joint; If the deviation between the first predicted torque and the first actual torque is greater than the second threshold, then the first moment corresponding to the first current deviation is obtained; The first moment is determined as the first collision moment; the first collision moment is the collision moment corresponding to the first current deviation among the a collision moments.

[0096] Optionally, in the process of evaluating the confidence level of the reference collision determination results and filtering them to obtain the target collision determination results, the filtering module 860 is specifically used for: Acquire the first sensor data corresponding to the first collision moment and the Jacobian matrix corresponding to the robot; the first sensor data includes three-dimensional force data and three-dimensional torque data; Based on a preset confidence calculation formula, the first confidence level is determined according to the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. If the first confidence level is within a preset confidence level range, then the first collision time is retained in the target collision determination result.

[0097] Optionally, the target collision determination result includes b collision times, where b is a positive integer less than or equal to a. In the step of removing data from the reference dataset based on the target collision determination result to obtain the target dataset, the data removal module 870 is specifically used for: Based on the window size, determine the b invalid data filtering intervals corresponding to the b collision times; The data corresponding to the b invalid data filtering intervals in the reference dataset are removed to obtain the valid dataset; The data in the effective dataset are classified and organized according to the task feature vector to obtain the target dataset.

[0098] Optionally, the determining module 840 is specifically used for: Obtain the set of environmental particles corresponding to the first collision moment; Based on the preset distance minimization principle, the location of the first collision point is determined according to the environmental particle set, the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. The first collision force corresponding to the first collision point position is determined based on the Jacobian matrix, the motor torque constant, and the first current deviation. Determine the magnitude and direction of the first collision force; The obstacle avoidance strategy of the robot is determined based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force.

[0099] Optionally, in determining the robot's obstacle avoidance strategy based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force, the determining module 840 is further specifically used for: The target joint of the robot that collides is determined based on the magnitude and direction of the first collision force. Obtain the first stiffness coefficient of the target joint; The first stiffness coefficient is adjusted according to a preset second ratio to obtain a second stiffness coefficient; Using the location of the first collision point as an obstacle avoidance constraint, a preset model prediction control algorithm is used to replan the joint trajectory within a preset time after the first collision moment to obtain the target joint trajectory. The obstacle avoidance strategy is determined based on the second stiffness coefficient and the target joint trajectory; the obstacle avoidance strategy is used to control the robot to perform obstacle avoidance operations.

[0100] It is evident that by fusing robot joint states and task features to construct input vectors, outputting predicted current sequences based on prediction networks, and combining dual-threshold anomaly detection and confidence assessment to accurately identify collision anomalies, collision anomaly data in the reference dataset are eliminated, thereby constructing a high-quality training set free from anomaly interference, which in turn improves the training effect and actual operation capability of the robot model.

[0101] It should be noted that the specific implementation of each operation can be described in the corresponding description of the method embodiments shown above. The training set optimization device 800 based on abnormal collision detection can be used to execute the method embodiments of this application, and will not be described again here.

[0102] This application also provides a computer-readable storage medium storing a computer program for electronic data interchange, which causes a computer to perform some or all of the steps of any of the methods described in the above method embodiments, wherein the computer includes an electronic device.

[0103] This application also provides a computer program product, which includes a non-transitory computer-readable storage medium storing a computer program operable to cause a computer to perform some or all of the steps of any of the methods described in the above method embodiments. The computer program product may be a software installation package, and the computer may include an electronic device.

[0104] It should be noted that, for the sake of simplicity, the above embodiments are all described as a series of actions. Those skilled in the art should understand that this application is not limited to the described order of actions, as some steps in the embodiments of this application can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are preferred embodiments, and the actions, steps, modules, or units involved are not necessarily essential to the embodiments of this application.

[0105] In the above embodiments, the descriptions of each embodiment in this application have different focuses. For parts not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0106] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. This program can be stored in a computer-readable storage medium, and when executed, it can include the processes described in the above method embodiments. The aforementioned storage medium includes various media capable of storing program code, such as ROM or random access memory (RAM), magnetic disks, or optical disks.

[0107] The steps of the methods or algorithms described in the embodiments of this application can be implemented in hardware or by a processor executing software instructions. The software instructions can consist of corresponding software modules, which can be stored in RAM, flash memory, ROM, EPROM, electrically erasable programmable read-only memory (EEPROM), registers, hard disk, portable hard disk, read-only optical disk (CD-ROM), or any other form of storage medium well known in the art. An exemplary storage medium is coupled to a processor, enabling the processor to read information from and write information to the storage medium. Of course, the storage medium can also be a component of the processor. The processor and storage medium can reside in an ASIC. Additionally, the ASIC can reside in a terminal device or management device. Alternatively, the processor and storage medium can exist as discrete components in the terminal device or management device.

[0108] Those skilled in the art will recognize that, in one or more of the examples above, the functions described in the embodiments of this application can be implemented, in whole or in part, by software, hardware, firmware, or any combination thereof. When implemented in software, it can be implemented, in whole or in part, as a computer program product. This computer program product includes one or more computer instructions. When these computer program instructions are loaded and executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another via wired (e.g., coaxial cable, fiber optic, digital subscriber line (DSL)) or wireless (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium accessible to a computer or a data storage device such as a server or data center that integrates one or more available media. The available media can be magnetic media (e.g., floppy disks, hard disks, magnetic tapes), optical media (e.g., digital video discs (DVDs)), or semiconductor media (e.g., solid-state disks (SSDs)).

[0109] The modules / units included in the various devices and products described in the above embodiments can be software modules / units, hardware modules / units, or a combination of both. For example, for devices and products applied to or integrated into a chip, all modules / units can be implemented using hardware methods such as circuits, or at least some modules / units can be implemented using software programs that run on a processor integrated within the chip, while the remaining (if any) modules / units can be implemented using hardware methods such as circuits. For devices and products applied to or integrated into a chip module, all modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components of the chip module, or at least some modules / units can be implemented using hardware methods such as circuits. The implementation is achieved through a software program that runs on the processor integrated within the chip module. The remaining modules / units (if any) can be implemented using hardware methods such as circuits. For various devices and products applied to or integrated into terminal equipment, each of their modules / units can be implemented using hardware methods such as circuits. Different modules / units can be located in the same component (e.g., chip, circuit module, etc.) or different components within the terminal equipment. Alternatively, at least some modules / units can be implemented through a software program that runs on the processor integrated within the terminal equipment, while the remaining modules / units (if any) can be implemented using hardware methods such as circuits.

[0110] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the embodiments of this application. It should be understood that the above descriptions are merely specific embodiments of the embodiments of this application and are not intended to limit the protection scope of the embodiments of this application. Any modifications, equivalent substitutions, improvements, etc., made on the basis of the technical solutions of the embodiments of this application should be included within the protection scope of the embodiments of this application.

Claims

1. A training set optimization method based on anomaly collision detection, characterized in that, The method includes: Obtain the reference dataset and task feature vector corresponding to the robot in the robot operation scenario; the reference dataset includes joint state sequences and actual current sequences. The joint state sequence and the task feature vector are concatenated dimensionally to obtain the input vector; The input vector is input into the prediction network corresponding to the robot to obtain the predicted current sequence; the prediction network is a time-series current prediction model adapted to the robot. Determine the current deviation sequence based on the actual current sequence and the predicted current sequence; The current deviation sequence is subjected to dual threshold anomaly determination to obtain a reference collision determination result; The confidence level of the reference collision determination results is evaluated and filtered to obtain the target collision determination results; Based on the target collision determination result, data is removed from the reference dataset to obtain the target dataset; the target dataset is used for iterative training of the prediction network. Wherein, the first current deviation is any current deviation in the current deviation sequence; the step of evaluating and filtering the confidence of the reference collision determination result to obtain the target collision determination result includes: The first sensor data corresponding to the first collision moment and the Jacobian matrix corresponding to the robot are obtained; the first sensor data includes three-dimensional force data and three-dimensional torque data; the first collision moment is jointly determined by the first current deviation and the torque deviation corresponding to the first current deviation calculated in combination with the preset motor torque constant; Based on a preset confidence calculation formula, the first confidence level is determined according to the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. If the first confidence level is within a preset confidence level range, then the first collision time is retained in the target collision determination result; The confidence level is calculated using the following formula: in, This represents the first confidence level; This represents the data from the first sensor; The pseudo-inverse of the Jacobian matrix is ​​used to implement the mapping from joint space to Cartesian space; This indicates the first current deviation; This represents the torque constant of the motor.

2. The method as described in claim 1, characterized in that, The acquisition of the reference dataset and task feature vector corresponding to the robot in the robot operation scenario includes: Obtain the original joint position data, original joint velocity data, original joint current data, and task target parameters corresponding to the robot in the robot operation scenario; the task target parameters include: task target position and task type label; Kalman filtering is performed on the original joint position data and the original joint velocity data respectively to obtain reference joint position data and reference joint velocity data. The joint state sequence is determined based on the reference joint position data and the reference joint velocity data; The original joint current data is subjected to median filtering to obtain the actual current sequence; The task feature vector is determined based on the task target location and the task type label.

3. The method as described in claim 1, characterized in that, The current deviation sequence includes m current deviations, where m is an integer greater than 1; the reference collision determination result includes a collision times, where a is a positive integer less than or equal to m. The step of performing dual-threshold anomaly determination on the current deviation sequence to obtain a reference collision determination result includes: The first sliding window corresponding to the first current deviation is determined according to the preset sliding step size and window size; the first current deviation is any one of the m current deviations. Calculate the first mean and first standard deviation of all current deviations in the first sliding window; A first threshold is determined based on the first mean and the first standard deviation; If the first current deviation is greater than the first threshold, then the first predicted torque and the first actual torque corresponding to the first current deviation are calculated based on the motor torque constant. Obtain the maximum load torque of the joints corresponding to the robot; The second threshold is determined based on a preset first ratio and the maximum load torque of the joint; If the deviation between the first predicted torque and the first actual torque is greater than the second threshold, then the first moment corresponding to the first current deviation is obtained; The first moment is determined as the first collision moment; the first collision moment is the collision moment corresponding to the first current deviation among the a collision moments.

4. The method as described in claim 3, characterized in that, The target collision determination result includes b collision times, where b is a positive integer less than or equal to a. The step of removing data from the reference dataset based on the target collision determination result to obtain the target dataset includes: Based on the window size, determine the b invalid data filtering intervals corresponding to the b collision times; The data corresponding to the b invalid data filtering intervals in the reference dataset are removed to obtain the valid dataset; The data in the effective dataset are classified and organized according to the task feature vector to obtain the target dataset.

5. The method as described in claim 1 or 4, characterized in that, The method further includes: Obtain the set of environmental particles corresponding to the first collision moment; Based on the preset distance minimization principle, the location of the first collision point is determined according to the environmental particle set, the first sensor data, the Jacobian matrix, the first current deviation, and the motor torque constant. The first collision force corresponding to the first collision point position is determined based on the Jacobian matrix, the motor torque constant, and the first current deviation. Determine the magnitude and direction of the first collision force; The obstacle avoidance strategy of the robot is determined based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force.

6. The method as described in claim 5, characterized in that, The step of determining the robot's obstacle avoidance strategy based on the location of the first collision point, the magnitude of the first collision force, and the direction of the first collision force includes: The target joint of the robot that collides is determined based on the magnitude and direction of the first collision force. Obtain the first stiffness coefficient of the target joint; The first stiffness coefficient is adjusted according to a preset second ratio to obtain a second stiffness coefficient; Using the location of the first collision point as an obstacle avoidance constraint, a preset model prediction control algorithm is used to replan the joint trajectory within a preset time after the first collision moment to obtain the target joint trajectory. The obstacle avoidance strategy is determined based on the second stiffness coefficient and the target joint trajectory; the obstacle avoidance strategy is used to control the robot to perform obstacle avoidance operations.

7. A training set optimization apparatus based on anomaly collision detection, used to perform the method as described in any one of claims 1-6, characterized in that, The device includes an acquisition module, a dimension splicing module, a prediction module, a determination module, an anomaly detection module, a filtering module, and a data removal module, wherein: The acquisition module is used to acquire a reference dataset and task feature vector corresponding to the robot in the robot operation scenario; the reference dataset includes joint state sequences and actual current sequences; The dimension concatenation module is used to concatenate the joint state sequence and the task feature vector dimensionally to obtain the input vector. The prediction module is used to input the input vector into the prediction network corresponding to the robot to obtain the predicted current sequence; the prediction network is a time-series current prediction model adapted to the robot. The determining module is used to determine a current deviation sequence based on the actual current sequence and the predicted current sequence; The anomaly determination module is used to perform dual-threshold anomaly determination on the current deviation sequence to obtain a reference collision determination result; The filtering module is used to evaluate the confidence level of the reference collision determination results and filter them to obtain the target collision determination results; The data removal module is used to remove data from the reference dataset based on the target collision determination result to obtain the target dataset; the target dataset is used to iteratively train the prediction network.

8. An electronic device, characterized in that, include: Processor, memory, communication interface, and one or more programs; The one or more programs are stored in the memory and configured to be executed by the processor, the programs including instructions for performing the steps of the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, the computer program including program instructions that, when executed by a processor, cause the processor to perform the method as described in any one of claims 1-6.

Images