Method and system for calibrating an automatic charger (ACD)

Abstract

The present invention relates to a method for calibrating one or more automatic charging devices (ACDs) and / or components thereof, wherein the ACD comprises an end effector for supporting and moving a vehicle charging connector, a motion control unit, and a computer vision unit for socket attitude estimation, wherein the computer vision unit comprises at least a baseline neural network, and the motion control unit is configured to control the motion of the effector for plugging the connector into a vehicle socket using attitude estimation from the computer vision unit.

Description

Detailed Description of the Invention

[0001] [Introduction] The present invention relates to a method for calibrating one or more autonomous charging devices (ACD) and / or components thereof, the ACD comprising an end effector for supporting and moving a vehicle charging connector, a motion control unit, and a computer vision unit for socket pose estimation, the computer vision comprising at least a baseline neural network, and the motion control unit being configured to control the motion of the effector for plugging the connector into a vehicle socket using the pose estimation from the computer vision unit.

[0002] [Technical Field of the Invention] The present invention generally relates to an operating autonomous charging device (ACD) that enables automatic charging of electric vehicles, as well as methods and systems related thereto. The methods and systems according to the present invention aim to improve the performance reliability and continuous operation of the ACD.

[0003] [Background of the Invention] Improving performance and reliability in robotic systems, particularly in robotic autonomous charging devices (ACD), remains a highly interesting field.

[0004] Robots are widely used in many different industries, such as assembly lines or production lines for automating manufacturing procedures. In recent years, robots have been utilized in attempts to automate the charging of batteries that power electric vehicles (EV) in the automotive industry. To ensure a good connection between the socket and the connector, it is desirable for the robot to accurately determine the position of the EV and / or the EV socket and align the connector in a continuous and reliable manner. Further improvements are still needed in this field as several components can contribute to poor plug-ins and undesirable operating performance.

[0005] Several solutions have been proposed, including the use of magnetic coupling systems to ensure accurate connections between connectors and sockets. However, many of these techniques require modifications to either the EV or the connector, which can impact the scalability and accessibility of the solutions in the market.

[0006] In recent years, other improvements have also been made, and some ACDs currently known in the art are capable of locating the socket location of an electric vehicle, guiding the charger connector, and plugging it in without operator intervention or modification of the vehicle or the vehicle's charging port. These improvements rely on the use of computer vision-based neural networks trained to accurately identify sockets so that the ACD can reliably complete the plug-in and plug-out process. Apparatuses for this purpose, as well as methods and systems related to such apparatuses, are known in the art, for example, from the international patent applications PCT / NL2020 / 050266, PCT / NL2021 / 050115, PCT / NL2021 / 050410, PCT / NL2021 / 050495, PCT / NL2021 / 05061, PCT / EP2022 / 062233, and PCT / EP2023 / 086489 from the same applicant of the present invention, all of which are incorporated herein by reference. The ACD device described herein may include compliance mechanisms to support several connectors and improve safety, including, but not limited to, CCS-1, CCS-2, MCS, and Tesla connectors.

[0007] An existing challenge in achieving good mating is ensuring precise positioning of the connector in the socket before plugging it in. This must be done with an accuracy of less than a few millimeters and less than a few degrees, which means that the charging device must satisfy this accuracy, as it is not feasible or practical for the vehicle to be positioned in such an alignment position.

[0008] Because the automatic charging device comprises several components, and due to the nature of robotic kinematics, the robot may exhibit some intrinsic inaccuracies. These inaccuracies may be the result of hardware and / or software components. In some cases, these inaccuracies can be resolved by computationally calibrating the ACD's motion model, but in other cases, additional hardware calibration is required, which is generally undesirable and costly, especially when software updates may alter the ACD's inaccuracies. Furthermore, many current calibration techniques are limited to offline processes, requiring the ACD to be shut down if the calibration and / or alignment of components is misaligned, or if the system needs to be recalibrated due to a software update. Therefore, it is desirable that ACD components be calibrated in a way that enables continuous system operation and reliability, and reduces the need for recurring or unexpected inspections and maintenance.

[0009] Neural networks are used to perform complex tasks such as classification tasks in recognizing patterns or objects in images, natural language processing, computer vision, speech recognition, bioinformatics, and other applications. In the context of this disclosure, the ACD comprises a computer vision unit based on a neural network configured to provide or support providing an estimate of the pose of a vehicle socket, where the pose is preferably defined as a combination of position and (angle) orientation. Such a neural network may be a convolutional neural network algorithm or an algorithm based on the "You Only Look Once" (YOLO) model. The ACD may estimate the pose of the socket either directly based on the neural network or by using a 3D pose estimation algorithm. The ACD then moves the end effector supporting the charging connector toward the vehicle socket to complete the mating process (also referred to herein as the connection or plug-in process). Thus, the output of the neural network is important for a good mating process. The quality of the output of the neural network depends, among other factors, on the quality of the training of the neural network. However, even when the data used to train the neural network is considered optimal, the neural network's output may exhibit some inaccuracies that can result in deviations or errors in socket pose estimation. Training and retraining a neural network can help improve the network output and subsequent pose estimation, either under specific conditions or in general. However, it has been observed that updating an ACD neural network with a newly trained or retrained neural network may introduce different or additional inaccuracies that may not have been considered before.

[0010] Furthermore, even when effective object pose estimation is achieved by ACD, unsuitable calibration or use of the robot, or general wear and damage to mechanical components, can adversely affect socket pose estimation.

[0011] The object of the present invention is to eliminate the drawbacks of the prior art, or at least to provide useful alternative forms thereof.

[0012] [A brief explanation of the conventional technology] Some existing systems have various drawbacks for certain applications. Therefore, further contributions in this field are still needed.

[0013] Reference D1 WO2020142496A1 describes a method for training a camera-coupled robot, the process comprising setting robot or camera parameters, capturing training images of a training object with the camera using the robot or camera parameters, changing the settings and capturing another training image, repeating such settings and captures to obtain multiple training images based on different settings, training the system to recognize a training object based on the multiple training images, and evaluating the system using pre-selected test images. This document describes training an industrial robot intended to complete a manufacturing process and makes no mention of initial calibration to compensate for the robot's intrinsic inaccuracies, nor of subsequent calibration relating to residual inaccuracies that may be introduced by a retrained neural network for socket pose detection.

[0014] [Overview of the prefecture] Known techniques devised to calibrate automatic charging devices and any of their components individually or as a whole offer limited benefits. Therefore, there is still a need for improved techniques to provide an effective and automated process for connecting the charging connector to the vehicle's socket.

[0015] In one embodiment, the present invention relates to a method for calibrating an automatic charging device (ACD), wherein the ACD comprises a camera, an end effector for supporting and moving a vehicle charging connector, a motion control unit, and a computer vision unit for socket attitude estimation, wherein the computer vision unit comprises at least a baseline neural network, and the motion control unit is configured to control the motion of the effector for plugging the connector into the socket using attitude estimation from the computer vision unit using an image of the vehicle socket recorded by the camera, and the method is as follows: a) Obtaining baseline inaccuracy between at least one target orientation of the end effector and at least one actual orientation of the end effector, - The target posture is the posture that the motion control unit controls to move the end effector toward, and the actual posture is the final posture toward which the end effector has effectively moved. - The target pose is determined using the output of the baseline neural network. Obtaining baseline inaccuracies and b) Adjusting the motion control unit by setting a baseline correction function using the acquired baseline inaccuracy to compensate for the acquired inaccuracy, c) Receiving an updated neural network for the computer vision unit, d) Determining the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, e) Calibrating the motion control unit by updating the baseline correction function based on the determined variation in attitude estimation inaccuracy. This includes methods.

[0016] The method according to this embodiment enables the calibration of the ACD, or at least some components of the ACD, such as motion control, thereby taking into account some inaccuracies of the ACD as well as inaccuracies related to the computer vision components of the ACD. The method enables remote calibration of the ACD, in particular remote calibration of the ACD related to the updating of a neural network by a trained or retrained neural network.

[0017] The method according to this embodiment facilitates the deployment of a retrained neural network to improve vehicle socket attitude detection, thereby avoiding the need for overall ACD calibration when deploying a new or retrained network.

[0018] In one embodiment, the present invention is a method for updating an automatic charging device (ACD), wherein the ACD is - End effector for supporting and moving the vehicle charging connector, - Motion control unit, - A computer vision unit for socket attitude estimation, wherein the computer vision comprises at least a baseline neural network, and the motion control unit is configured to use attitude estimation from the computer vision unit to control the motion of an effector for plugging a connector into a vehicle socket. The motion control unit includes a baseline correction function that compensates for the inaccuracy between the actual posture and the target posture, where the target posture is the posture towards which the motion control unit controls the end effector to move, and the actual posture is the final posture to which the end effector has effectively moved, and the computer vision unit and Equipped with, The method comprises a) receiving an updated neural network for a computer vision unit; b) determining a variation in pose estimation inaccuracy between a baseline neural network and the updated neural network; c) calibrating a motion control unit by updating a baseline correction function based on the determined variation in pose estimation inaccuracy and relates to a method.

[0019] In one embodiment, the invention is an automatic charging device (ACD) comprising - an end effector for supporting and moving a vehicle charging connector; - a camera; - a motion control unit; - a computer vision unit for pose estimation comprising a baseline neural network, wherein the motion control unit comprises a baseline correction function for correcting inaccuracies between an actual pose and a target pose; - a computer vision unit configured such that the motion control unit uses pose estimation from the computer vision unit to control the movement of the end effector to plug the connector into a vehicle socket; - a control device comprising - receiving an updated neural network for the computer vision unit; - determining a variation in pose estimation inaccuracy between the baseline neural network and the updated neural network; - calibrating the motion control unit by updating the baseline correction function based on the determined variation in pose estimation inaccuracy and relates to an automatic charging device (ACD).

[0020] ​In one embodiment, the present invention relates to a method for calibrating a motion control unit of a plurality of automatic charging devices (ACDs), wherein the ACDs are arranged in an area where the operating parameters of the ACDs are substantially equivalent.

Brief Description of the Drawings

[0021] [Figure 1] It is a reference diagram showing an ACD (1) that supports the connection of a charger connector (2) to a socket (4) of an electric vehicle (5). [Figure 2] It is a schematic diagram of an automatic charging device. An end effector (10), a motion control unit (20), a computer vision unit (30), a control device (40), and a computer system (50) are shown. [Figure 3] It is a flowchart of the process steps of the method according to the first embodiment of the present invention. [Figure 4] It is a chart showing the determination of the variation in pose estimation inaccuracy between a baseline neural network and an updated neural network, and generating an updated calibration function based on the variation in pose estimation inaccuracy.

Modes for Carrying Out the Invention

[0022] [Detailed Description of the Invention] In the following description, for purposes of explanation and not limitation, specific details such as specific methods, steps, devices, components, etc. are set forth in order to provide a thorough understanding of the present invention. However, it will be apparent to those skilled in the art that the present invention may be practiced in other embodiments that depart from these specific details.

[0023] In one embodiment, the present invention relates to a method for calibrating an automatic charging device (ACD) configured to support and connect a charging connector to a vehicle charging socket based on a computer vision neural network and a motion control unit.

[0024] automatic charging device The automatic charging devices (or simply robots) referred to herein may comprise several hardware and software components. In particular, an ACD may comprise an end effector for supporting and moving a vehicle charging connector, a motion control unit, and a computer vision unit for attitude estimation, comprising at least a baseline neural network, the motion control unit being configured to control the motion of the end effector for plugging the connector into a vehicle socket using attitude estimation from the computer vision unit. The ACD may include, or communicate with, a camera for acquiring an image of the vehicle socket as input for the computer vision unit. The end effector may be a controllable actuation mechanism comprising means for supporting the vehicle charging connector in a releasable or non-releasable manner, enabling such an automatic charging process. A robot may further comprise one or more linkages having one or more joints and operable actuators (e.g., electric motors, stepper motors, and solenoids) coupled to move the linkages in response to control signals or drive signals.

[0025] The ACD connection process generally includes, at a minimum, the steps of: determining the orientation of the vehicle charging port (including a socket cover, socket features, socket pins, reference features, markers, or a combination thereof); moving the vehicle charging connector toward the vehicle charging port; connecting the charger to the vehicle socket; enabling the vehicle to be charged; optionally releasing the connector from the ACD end effector; and optionally disconnecting the charging connector. The orientation of the vehicle socket is preferably determined by the ACD's computer vision unit (or a computer vision unit communicating with such an ACD) based on a neural network that utilizes image data from the socket to enable orientation estimation by the computer vision unit, and the orientation includes, for example, the position and orientation of the socket.

[0026] Due to the nature of the kinematic model, the robot may be subject to several intrinsic and / or extrinsic factors, which may result in inaccurate pose estimation of sockets and / or poor mating of connectors to sockets. Such inaccuracies may be the result of hardware and / or software components of the ACD. In some cases, the inaccuracies can be resolved by computationally calibrating the ACD's motion model to characterize its actual behavior. However, in some cases, additional hardware calibration may be required. Other ACD components that may give rise to inaccuracies include, in particular, camera hardware inaccuracies, camera calibration, ACD kinematics, hardware component play, compliance components, motion control, computer vision, hardware wear, and damage.

[0027] The ACD according to the present invention is intended to be deployed in the field, such as in an EV charging station, where one or more charging devices are preferably positioned to facilitate the automatic charging of such EVs. The ACD may require one or more calibration processes to achieve suitable performance in the field, and the calibration processes may be performed in different circumstances. Such calibrations may include i) an internal camera calibration (to correct lens distortion) which may be performed before assembly of the ACD or after installation in the field; ii) an external camera calibration, i.e., hand-to-eye calibration, which may be performed during assembly or after installation in the field, to find a precise reference frame for the rest of the system's kinematic chain; and iii) a motion control unit calibration aimed at calibrating motion behavior (such as control accuracy and / or model accuracy) which may be performed during or after assembly, during or at a later stage of robot commissioning.

[0028] It is desirable to tune the motion control unit by setting a baseline correction function using acquired baseline inaccuracies to compensate for acquired baseline inaccuracies during or prior to ACD deployment, thereby compensating for at least some, and preferably most, of the possible inaccuracies in the robot's software or hardware components, also known as bias or offset, which may contribute to undesirable execution of the mating process. Such setting of the baseline correction function may be done when the ACD is assembled or when the ACD is installed in the field.

[0029] The ACD end effector is part of a controllable actuation mechanism, also referred to herein as a manipulator or robotic arm, which supports a vehicle charging connector, whether in a retractable form or not, and is configured to guide the connector toward the vehicle charging port to complete the plug-in and / or plug-out process.

[0030] According to the present invention, sensors may be mounted on or around the ACD and used to capture data from the vehicle, vehicle socket, or its surroundings. Multiple sensors may be mounted on the same ACD and may be of various types for collecting information on several types of parts / objects. Sensors include data-collecting sensors such as 2D or 3D cameras configured to collect images, video, and / or audio. However, such data may also be obtained from sensors positioned outside the ACD system. Sensors according to this disclosure further include force, light, wind, geographic location and orientation, temperature and pressure sensors, and any combination thereof. Sensors according to the present invention may also provide the date and time the data was captured.

[0031] Socket posture estimation An object's orientation describes how the object is positioned in the three-dimensional space it uses. An object's orientation can be determined relative to a viewpoint, such as a camera viewpoint or camera coordinate system. It may include three-dimensional information characterizing the object's rotation relative to the camera viewpoint or camera coordinate system. Alternatively or additionally, an object's orientation may include three-dimensional information characterizing the object's translation relative to the camera viewpoint or camera coordinate system.

[0032] In the context of this disclosure, attitude determination is performed on an electric vehicle socket, but is not limited thereto. Preferably, the attitude of an electric vehicle charging port (also called a socket) may be understood as the position and orientation of the charging port on the vehicle, and in some embodiments, is represented by the 3D Cartesian position and yaw (x,y,θ) of the socket. However, in some embodiments, the attitude is a 6D attitude, where the position is defined by the 3D Cartesian position and the orientation is defined by the roll, pitch, and yaw of the socket.

[0033] Determining the electric vehicle socket attitude is useful for representing the attitude and orientation of the charging port relative to the ACD camera position in order to guide the EV connector toward the charging port and complete the plug-in process. By obtaining the determination of the charging port relative to the camera attitude and by performing eye-to-hand calibration, the motion control unit can guide the end effector to move toward the estimated socket attitude.

[0034] An object's pose can be obtained from a single image, possibly from multiple images from the same pose, possibly from multiple images from different poses, or by utilizing an earlier determined pose. Alternatively or additionally, an object's pose can be obtained from a multiview image or video.

[0035] Computer Vision Unit Neural Network In some embodiments of this disclosure, the orientation of the socket may be determined by a computer vision unit based on a neural network. Preferably, the neural network may be trained to determine the orientation through the analysis of one or more images. The estimated socket orientation may include estimates of the socket's dominant axis, roll, elevation, angular position, attitude, and azimuth. Alternatively, the neural network may be trained to detect geometric features of the socket, which may serve as reference markers, and based on these reference markers, an orientation estimation algorithm may determine the orientation of the socket relative to the camera.

[0036] A computer vision unit may be based on or comprise a neural network. In some embodiments, the computer vision unit is configured to host one or more computer vision neural networks, also referred to herein simply as neural networks. The computer vision unit may communicate with or host computer vision training modules and / or computer vision test modules. The training modules and / or test modules may be hosted independently of the computer vision unit. In some embodiments, the computer system may further comprise an edge computer that hosts the training modules and / or test modules, the edge computer which can run machine learning algorithms on collected data to form knowledge of neural networks for general or case-specific applications. A computer vision neural network may be any type of neural network that can be used in image processing. For example, a neural network may be a feedforward neural network, a regulatory feedback neural network, a convolutional neural network, or a recurrent neural network.

[0037] Improving the robustness of computer vision components, which are capable of determining the relative position and orientation of charging ports, is a critical task for increasing process performance. Since the performance of computer vision components generally depends on the neural network on which they are built, it is desirable to provide the computer vision unit with a robust neural network suitable for the ACD system.

[0038] In some embodiments of this disclosure, the orientation of a vehicle or vehicle charging port may be determined directly by a neural network. In some embodiments, the neural network may be trained to determine the location of a geometric feature of a socket to be used as a reference marker, and based on that location, an algorithm such as a PnP (perspective-n-point) algorithm like solvePnP or RANSAC (Random Sample Consensus) can estimate the orientation of the socket. This may utilize a single image or multiple images.

[0039] The training process of a neural network typically determines the quality of the network's output. Training and testing a neural network can involve several steps. First, a large dataset of images is collected that capture various scenarios relevant to ACD operation, such as different lighting conditions, socket orientation, and environmental variations. These images are then annotated with ground truth pose information, which serves as a baseline for training the neural network.

[0040] The training process generally involves feeding these annotated images into a neural network and iteratively adjusting the network's internal parameters to minimize the difference between the predicted pose and the ground truth pose. This adjustment is achieved through optimization algorithms such as stochastic gradient descent, allowing the network to learn to extract relevant features from those images and make accurate pose estimates.

[0041] To evaluate the performance of a trained neural network, a separate dataset, distinct from the training data, may be used for testing. This test dataset helps assess how well the network generalizes to unseen data. Various metrics, such as precision, recall, and mean squared error, can be calculated to quantify the network's performance.

[0042] Furthermore, techniques such as cross-validation may be employed to ensure that the network's performance is consistent across different subsets of the data. Cross-validation involves dividing the dataset into multiple subsets, training the network on one subset, testing the network on the remaining subsets, and alternating this process to cover all possible combinations. This helps detect overfitting, where the network performs well on the training data but fails to generalize to new data.

[0043] Furthermore, fine-tuning and optimization strategies can be applied to improve network performance. This may involve adjusting hyperparameters such as the learning rate, network architecture, or regularization method to achieve better pose estimation results.

[0044] Generally, large amounts of training data are collected and annotated manually or automatically. To improve the performance of ACDs, it is desirable to retrain the neural network based on newly collected data or data deemed more suitable for a particular ACD or group of ACDs.

[0045] Training and retraining a neural network can improve the network output and subsequently help improve pose estimation. In some cases, it is expected that a neural network may exhibit some inaccuracies in pose detection when it is trained for pose detection of a new socket, for example, when the amount or quality of training data is insufficient. In some cases, deploying a new neural network to ACD (also referred herein as neural network update or updating) may introduce additional inaccuracies, also called offsets, which may not have been considered previously. Similarly, deploying or updating a new neural training may reduce or alter the magnitude of inaccuracies in an existing baseline neural network. In the context of the present invention, it may be assumed that computer vision units and / or their neural networks are always prone to inaccuracies, meaning that they do not always determine ground truth. In such cases, a proper comparison between the baseline neural network and the updated neural network will determine the variation in their inaccuracies.

[0046] As used herein, the term "baseline neural network" is used when referring to the initial neural network model deployed within the ACD's computer vision unit. This baseline neural network may serve as a baseline for socket pose estimation and provide an initial framework for determining the relative position and orientation of the charging port. As used herein, the term "trained neural network" is adopted when describing a neural network that has undergone one or more training and possibly subsequent retraining processes to improve the performance and adaptability of the neural network within the ACD system.

[0047] Based on the foregoing, in some embodiments of the present disclosure, the present invention may include determining the variation in pose estimation inaccuracy between a baseline neural network and an updated neural network, and calibrating a motion control unit by updating baseline correction parameters based on the determined variation in pose estimation inaccuracy, the motion control unit being calibrated when the determined variation exceeds a predetermined threshold.

[0048] The ACD may preferably include, but not limited to, a computer system suitable for controlling any of its components, including, or networking with, one or more end effectors, one or more cameras, one or more computer vision units, and one or more motion control units. The computer system may be any type of suitable computer system, such as an edge computer, which is capable of controlling one or more components of the ACD, either directly or via a controller.

[0049] The ACD workspace represents the three-dimensional space in which the robot can operate and move, although in some implementations, the workspace may represent a two-dimensional space. Preferably, the workspace refers to the full range of poses in which the ACD's end effector can move. The workspace may vary or be adapted depending on the type of ACD or the specific configuration of the ACD.

[0050] The robot may be equipped with or communicating with a motion control unit, which may further be equipped with a motion planner, configured to dynamically create a motion plan for the end effector based on a determined socket posture and a determined posture of the end effector. The motion control may preferably include an algorithm for controlling the motion of the end effector supporting the charging connector.

[0051] ACD calibration According to this disclosure, the performance, reliability, and / or accuracy of the plug-in process may be affected by the individual behavior of various components of the ACD, particularly the performance of computer vision components, mechatronics, system and component calibration (such as internal and external calibration of the camera), hand-eye calibration of the relative position of the camera and end effectors and / or connectors, wear and damage. Advantageously, ACD reliability and automatic operation can be improved by a suitable calibration of the motion control unit, which takes into account some, preferably all, of the components of the ACD that may be contributing to undesirable performance or process inaccuracies. By performing calibration of the motion unit, it is possible to compensate for existing and newly introduced accuracy in the system. Preferably, the calibration of the motion control unit takes into account previous calibrations performed in the ACD. Within the scope of the present invention, the motion control unit can be calibrated by setting baseline correction parameters, and preferably by updating the baseline correction parameters when an updated neural network for the computer vision unit is received and intended to be deployed to the ACD computer vision unit. Such adjustments and / or calibrations may be performed by a computer system configured to adjust algorithms for controlling the motion of the end effector.

[0052] In a first embodiment, a method for calibrating an automatic charging device (ACD) is described, wherein the ACD comprises an end effector for supporting and moving a vehicle charging connector, a motion control unit, and a computer vision unit for socket attitude estimation, wherein the computer vision unit comprises at least a baseline neural network, and the motion control unit is configured to control the motion of the effector for plugging the connector into a vehicle socket using attitude estimation from the computer vision unit.

[0053] Figure 3 shows a schematic flowchart of an exemplary embodiment of the method according to the present invention.

[0054] The method begins with step 100, which obtains a baseline inaccuracy between at least one target pose of the end effector and at least one actual pose of the end effector, where the target pose is the pose in which the motion control unit controls the end effector to move toward it, and the actual pose is the final pose in which the end effector has effectively moved toward it, and the target pose is determined using the output of a baseline neural network.

[0055] The method continues with step 110, which adjusts the motion control unit by setting a baseline correction function using the acquired baseline inaccuracy to compensate for the acquired inaccuracy.

[0056] This method continues with step 120, which involves receiving an updated neural network for the computer vision unit.

[0057] This method continues with step 130, which determines the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network.

[0058] The method continues with step 140, which involves calibrating the motion control unit by updating the baseline correction function based on the determined variation in attitude estimation inaccuracy.

[0059] More specifically, the method according to the first embodiment is: a) Obtaining baseline inaccuracy between at least one target orientation of the end effector and at least one actual orientation of the end effector, - The target posture is the posture that the motion control unit controls to move the end effector toward, and the actual posture is the final posture toward which the end effector has effectively moved. - The target pose is determined using the output of the baseline neural network. Obtaining baseline inaccuracies and b) Adjust the motion control unit by setting a baseline correction function using the acquired baseline inaccuracy to compensate for the acquired inaccuracy. It may include.

[0060] Baseline inaccuracy Baseline inaccuracy between the target and actual orientations may result from the sum of some or all inaccuracies introduced by one or more components in bringing the connector to the intended orientation. Such inaccuracies may be representations of specific values, such as specific vectors and magnitudes, and may also be values ​​or sets of values ​​that vary over a given time frame, i.e., varying vectors and magnitudes. Such inaccuracies may be defined as a set of coordinates, including translational and rotational coordinates, representing the difference between the target and actual orientations. In some cases, baseline inaccuracy may be defined as the size of the offset between the target and actual orientations of the end effector.

[0061] Baseline inaccuracies include at least one or more individual inaccuracies. Baseline inaccuracies include at least one of the following: baseline neural network inaccuracies, pose estimation inaccuracies, inaccuracies due to ACD hardware components, inaccuracies due to ACD software components, and inaccuracies in the motion control unit. Baseline inaccuracies related to hardware components may include at least one of the following: camera hardware inaccuracies, inaccuracies in the manufacture and assembly of structural and / or movable components, inaccuracies in the control of movable components, play in movable components, compliant deformation, wear and damage to ACD structural components, these compared to their respective representations in the system's control algorithms. Baseline inaccuracies may determine the overall system inaccuracy because it takes into account a set of multiple components of individual inaccuracies. Preferably, the magnitude and source of each of the individual inaccuracies can be identified.

[0062] The target attitude is preferably determined by estimating the attitude of the vehicle socket using the ACD's computer vision unit, and by instructing the ACD to move the end effector toward such estimated attitude. Any inaccuracies in the attitude estimation by the computer vision unit shall be reflected in the socket attitude estimation. Both the target attitude and the actual attitude may be defined as a set of coordinates including translational and rotational coordinates.

[0063] The target orientation may be the orientation towards which the motion control unit controls the end effector to move, and the actual orientation may be the final orientation towards which the end effector has effectively moved. Hand-to-eye calibration allows the ACD and motion control unit to determine the relative orientation of the end effector and / or charging connector with respect to the camera and move the end effector accordingly using the hand-to-eye transformation matrix. In some cases, the actual orientation is equal to the target orientation. In such cases, a good mating process can be expected. In some cases, the target orientation is slightly different from the actual orientation. In such cases, good mating can also be expected, given the ACD's standard-compliant features or the presence of guide surfaces in the vehicle socket. In some cases, the target orientation is considerably different from the actual orientation, which may be referred to herein as baseline inaccuracy. In such cases, mating that ends in failure can be expected.

[0064] Baseline inaccuracy can be determined by instructing the ACD to move the charging connector to a target orientation and measuring the offset between the target orientation and the actual orientation, i.e., the size of the offset. Preferably, baseline inaccuracy can be obtained by instructing the ACD to move the charging connector to at least one target orientation and measuring the translational and rotational differences between at least one target orientation and each of at least one actual orientation. In some cases, baseline inaccuracy is obtained by measuring the translational and rotational differences between multiple target orientations and their respective actual orientations. In some cases, baseline inaccuracy can be obtained by a predictive function from one or more previous offset determinations. In some cases, baseline inaccuracy can be obtained by measuring the rotational and / or translational compliance motion caused by the charging connector due to the self-aligning components of the charging connector and / or the guide surfaces of the socket when connected to the socket, the compliance motion referring to the movement or displacement of the self-aligning components due to an external force or load. Compliance motion may include at least one of deflection and deformation of a compliant component, where deflection refers to bending or displacement of the loaded component without any change in shape or size, and deformation involves a change in shape or size of the loaded component, such as elongation, compression, or twisting. In some cases, baseline inaccuracy is obtained by receiving sensor data from sensors including force sensors, torque sensors, force and torque sensors, one or more cameras, one or more distance sensors, motion capture systems, or combinations thereof.

[0065] Baseline inaccuracies can be obtained by iteratively determining the baseline inaccuracies for different estimated poses until the iterative step no longer produces a significant difference in the obtained baseline inaccuracies. The different estimated poses cover the majority of poses in the ACD's working space.

[0066] Baseline correction function Once baseline inaccuracy is acquired, the method may include adjusting the motion control unit by setting a baseline correction function to compensate for the acquired inaccuracy, which utilizes the acquired baseline inaccuracy. Adjusting the motion control unit may allow the actual attitude to approach or approach the target attitude, which may be accurate enough to facilitate good insertion and removal of the charging connector.

[0067] Tuning a motion control unit by setting a baseline correction function may include at least one offset assigned to at least one of the three-dimensional spatial coordinates, and the motion control unit is tuned by setting such a baseline correction function to compensate for acquired inaccuracies. An example of such at least one offset includes a rotational offset of some magnitude and a translational offset of some magnitude, and such offsets enable compensation for acquired inaccuracies. Setting a baseline correction function may include assigning at least one offset to the motion control unit to compensate for acquired inaccuracies, and said offset includes a rotational offset, a translational offset, or a combination thereof. In some cases, setting a baseline correction function may include assigning the motion control unit offsets for at least one, at least two, at least three, at least four, at least five, or at least six degrees of freedom to compensate for acquired inaccuracies. Setting a baseline correction function may include assigning the motion control unit offsets for at least one, two, or three rotational degrees of freedom and at least one, two, or three translational degrees of freedom to compensate for acquired inaccuracies. In this manner, when the motion control unit is operated, it takes into account the assigned offset and thereby compensates for any acquired inaccuracies. For example, if the unit is intended to move in a straight line along a certain axis but exhibits a slight deviation due to inherent inaccuracies, the baseline correction function may adjust the motion accordingly by applying the assigned offset. Similarly, if there is a rotational misalignment during motion, the correction function may correct it by applying a specified rotational offset. As used herein, the term offset refers to a predetermined adjustment applied to the motion control unit of the ACD to compensate for acquired inaccuracies. Such an offset represents a corrective measure implemented within the motion control unit to ensure that the actual attitude is precisely aligned with the target attitude, facilitating good insertion and removal of the charging connector into and out of the socket.

[0068] The baseline correction function may include at least one of qualitative and quantitative correction instructions that enable the ACD to achieve good plugging of the connector into the vehicle socket. In some cases, the baseline correction function may be generated by using recorded position values ​​of the charging connector. In some cases, the baseline correction function may be a function of the attitude or range of attitudes of the end effector in the ACD workspace, which determines the baseline inaccuracy. The baseline correction function may be selected from one or more of the following: a black-box model, a gray-box model, or a white-box model, and a parameterized model or a non-parameterized model.

[0069] The baseline correction function may include a combination of calibrations for the individual components of the ACD, and may, in addition or alternatively, include an overall calibration function to compensate for remaining inaccuracies so that the ACD can successfully insert and remove the connectors from the EV, respectively. In some cases, the baseline correction function may include at least one subfunction, each of which compensates for acquired baseline inaccuracies. In some cases, the baseline correction function may include at least one baseline correction subfunction, each compensating for at least one of the following: baseline neural network inaccuracies, attitude estimation inaccuracies, inaccuracies due to ACD hardware components, inaccuracies due to ACD software components, inaccuracies in the motion control unit, or a combination thereof. In some cases, the baseline correction function may include at least one subfunction that corrects at least one of the following: inaccuracies in the computer vision output, inaccuracies in the baseline neural network, pose estimation inaccuracies, inaccuracies due to ACD hardware components, inaccuracies due to ACD software components, and inaccuracies in the motion control unit. In some cases, the baseline correction function includes a subfunction isolated to either the pose estimation output or the computer vision output, i.e., the baseline correction function includes at least one subfunction that corrects only the pose estimation output or the computer vision output. In such cases, updating the baseline correction function based on a determined variation in pose estimation inaccuracies can be advantageously done by considering only such pose estimation output or computer vision output, without necessarily having to consider other subfunctions.

[0070] The baseline correction function can be set as an overarching system calibration function or as a function for calibrating an ACD system or its subsystems. The baseline correction function may be selected from one or more linear, nonlinear, or polynomial functions to represent behavior that is time-dependent, space-dependent, or dependent on other parameters, such as a function for modifying signal values ​​using a constant offset, and these functions may have multiple inputs and outputs. Another concrete example is a lookup table, with or without interpolation or extrapolation, for finding intermediate values ​​between entries.

[0071] In some cases, when the baseline inaccuracy exceeds a predetermined threshold, such as when the target posture differs significantly from the actual posture, the method may include adjusting the motion control unit by setting a baseline correction function using the acquired baseline inaccuracy to compensate for the acquired inaccuracy. In some cases, the motion control unit is adjusted when the baseline inaccuracy exceeds a threshold of 0.1 mm or 0.1 degree offset in Cartesian space.

[0072] The actions described herein, namely obtaining or determining baseline inaccuracy, setting baseline correction functions, and adjusting the motion control unit, may be performed by the ACD controller, by the ACD operator, and / or by the controller in a remote manner. In the context of the present invention, the term “obtain” may generally mean to retrieve or receive information. In the context of the present invention, the term “determine” may generally mean to measure either qualitatively or quantitatively.

[0073] This method may also include comparing the determined baseline inaccuracy with a dynamic threshold and calibrating the motion control unit when the baseline inaccuracy exceeds the dynamic threshold. Preferably, the dynamic threshold is a function of the end effector's orientation in the workspace. Within the end effector workspace, the end effector may be required to perform different displacement or rotational motions. Therefore, the inaccuracy is expected to be greater or smaller in some areas of the end effector workspace than in others. The dynamic threshold may be an orientation-dependent threshold of the end effector.

[0074] Attitude estimation variation While training, retraining, updating, and / or deploying a neural network can improve the network output and thus help improve pose estimation, it has been observed that updating a new neural network to a computer vision unit can alter the magnitude of existing inaccuracies and even introduce additional or different inaccuracies that are not accounted for by the baseline correction function. Advantageously, the present invention enables the determination of the variation in pose estimation inaccuracies between a baseline neural network and an updated neural network, and thus the baseline correction function can be updated accordingly, enabling a good connection process during neural network updates and deployments. In other words, the present invention makes it possible to compensate for the impact of the difference in pose estimation inaccuracies of the updated neural network on the motion control unit.

[0075] In particular, the method of the present invention is c) Receiving an updated neural network for the computer vision unit, d) Determining the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, e) Calibrating the motion control unit by updating the baseline correction function based on the determined variation in attitude estimation inaccuracy. This may further include:

[0076] Calibrating the baseline correction function based on the determined variation in attitude estimation inaccuracy may preferably allow the motion control unit to continue achieving good plug-in. Calibrating the motion control is preferably done by updating the baseline correction function, which allows for calibration of the motion control unit or its output.

[0077] Variations in pose estimation inaccuracy can be determined by several means, including measuring the difference in inaccuracy of the pose estimation outputs of the baseline neural network and the updated neural network on a benchmark dataset containing multiple images on which ACD is likely to plug in or plug in well.

[0078] Alternatively, variability in pose estimation inaccuracy can be determined by subtracting the output of the updated neural network from the output of the baseline neural network across a benchmark dataset of images.

[0079] Alternatively, variability in pose estimation inaccuracy could be the average difference between the output of the baseline neural network and the output of the retrained neural network across a benchmark dataset of images.

[0080] Alternatively, variability in pose estimation inaccuracy may include the average of values ​​resulting from subtracting the output of the updated neural network from the output of the baseline neural network.

[0081] Alternatively, variability in pose estimation inaccuracy may be determined by comparing the updated absolute socket pose estimation error function of the neural network to ground truth with the absolute socket pose estimation error function of the baseline neural network to ground truth, the socket pose estimation error function may be determined by comparing the output of each neural network with ground truth measurements. According to the present invention, ground truth may refer to manually labeled data representing the correct classification or annotation of the socket or its features present in the image of the socket.

[0082] Additionally or alternatively, images in the benchmark dataset may include means for determining ground truth measurements of the socket orientation relative to the camera, and means for determining ground truth may include at least one reference marker placed at a known position relative to the socket.

[0083] Variation in attitude estimation inaccuracy can be defined as a set of coordinates that represent variation in attitude estimation inaccuracy, including at least one of translational and rotational coordinates.

[0084] The baseline correction function can be updated by adding or subtracting a determined variation in pose estimation inaccuracy to or from the baseline correction function. This determined variation can be added to or subtracted from the computer vision module's pose estimate, either as an addition or substitution, thereby adjusting the output of the motion control unit. Similarly, the baseline correction function can be updated by adding a variation in pose estimation inaccuracy to the computer vision module's pose estimate, such as the average of the values ​​obtained by subtracting the updated neural network output from the baseline neural network output, thereby adjusting the output of the motion control unit.

[0085] In some cases, the baseline correction function may be updated by adding a 6-degree-of-freedom constant offset to the output of the computer vision module, the offset being determined by subtracting the 6-degree-of-freedom output of the computer vision module when running the benchmark dataset with the updated neural network from the 6-degree-of-freedom output of the computer vision module when running the benchmark dataset with the baseline neural network. In some cases, the motion control unit is calibrated by updating the baseline correction function, in which at least one subfunction of the baseline correction function is updated using variations in pose estimation inaccuracy.

[0086] In some cases, the baseline correction function may be updated by determining a 6-degree-of-freedom offset function and adjusting the output of the computer vision module using the offset function, the offset function being obtained by determining the updated absolute estimation error of the neural network based on ground truth measurements. In such cases, the updated absolute estimation error of the neural network may be determined by placing a reference marker at a known position relative to the EV socket, estimating the pose of both the socket and the reference marker in one image from several recording positions, using the estimate of the reference marker as ground truth, estimating the error of the socket pose by ground truth, averaging the estimation errors across several recording positions, and determining a 6-degree-of-freedom offset function based on the absolute estimation error. Preferably, in such cases, the average of the absolute estimation errors is obtained based on the averaged estimation errors across several recording positions.

[0087] Advantageously, the motion control unit can be calibrated from a remote device via a communication device, which compares the variation in attitude estimation with an optimal variation range from a database of historical decisions and adjusts the threshold based on the optimal variation. In some cases, the calibration of the motion control unit is a recalibration.

[0088] Preferably, the method of the present invention may include deploying the updated neural network to a computer vision unit. In some cases, the neural network is deployed on the condition that an improvement in pose estimation is observed between the baseline neural network and the updated neural network.

[0089] The actions described herein, namely receiving an updated neural network, determining variations in pose estimation inaccuracy, and calibrating the motion control unit, may be performed by an ACD controller or control device, by an ACD operator, and / or by the controller or control device in a remote manner.

[0090] In another embodiment, the present invention relates to a method for updating an automatic charger (ACD), wherein the ACD is - End effector for supporting and moving the vehicle charging connector, - Camera and, - Motion control unit, - Computer vision unit for socket pose estimation and Equipped with, The computer vision system is configured to include at least a baseline neural network, and the motion control unit is configured to control the motion of the effector for plugging the connector into the socket using attitude estimation from the computer vision unit using images of the vehicle socket recorded by a camera. The motion control unit includes a baseline correction function that compensates for the inaccuracy between the actual posture and the target posture. The target posture is the posture that the motion control unit controls to move the end effector toward. The actual position is the final position where the end effector is effectively moved toward it. Book 2, a) Receiving an updated neural network for the computer vision unit, b) To determine the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, c) Calibrating the motion control unit by updating the baseline correction function based on the determined variation in attitude estimation inaccuracy. This includes methods.

[0091] In another embodiment, the present invention is an automatic charging device, - End effector for supporting and moving the vehicle charging connector, - Camera and, - Motion control unit, - A computer vision unit for pose estimation, equipped with a baseline neural network, - The motion control unit is configured to control the motion of the end effector for plugging the connector into the socket using attitude estimation from a computer vision unit using images of the vehicle socket recorded by a camera, and the motion control unit has a baseline correction function that compensates for the inaccuracy between the actual attitude and the target attitude, - Control device and Equipped with, The control device - Receiving updated neural networks for the computer vision unit, - To determine the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, - Calibrating the motion control unit by updating baseline correction parameters based on the determined variation in posture estimation inaccuracy. The present invention relates to an automatic charging device, characterized by being configured to perform the following actions.

[0092] The ACD may further include a camera configured to record an image of at least one vehicle charging socket, thereby the image being used as input for a computer vision unit.

[0093] The ACD may further comprise a computer system communicating with a motion control unit, the computer system comprising a storage unit configured to store one or more of the baseline inaccuracy determination, baseline correction parameters, and / or adjusted correction parameters.

[0094] The control device may be further configured to acquire a baseline inaccuracy between the target pose of the end effector and the actual pose of the end effector, and to generate a baseline correction function based on the baseline inaccuracy, wherein the target pose is determined using the output of a computer vision baseline neural network, and to adjust a motion control unit or its output by setting baseline correction parameters using the baseline correction function to compensate for the acquired inaccuracy.

[0095] Exemplary Implementations In an exemplary implementation, the variation in pose estimation offset between a baseline neural network and a retrained neural network is achieved by training two networks using a general dataset containing 3409 images from a CCS2 socket. These two networks are individually trained using a dataset consisting of 50 / 50 splits of the general dataset, and then, respectively, 80 / 10 / 10 for each split for training / testing / validation. In parallel, a benchmark dataset of 141 recorded images is generated by the validated poses, and the trained networks are used to generate pose estimates for these 141 images in the benchmark dataset. The resulting pairs of pose estimates are compared, and their differences are visualized. Figure 4 shows histograms representing the observed offsets between the two networks for both translational and rotational differences. The majority of the translational and rotational offsets, i.e., the variation in pose estimation offsets between the baseline neural network and the retrained neural network, fall within a discernible range, and based on this, the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network is determined. The motion control unit can be calibrated by updating the baseline correction function based on the determined variation in pose estimation inaccuracy. In this exemplary embodiment, the variation in pose estimation inaccuracy is the average of the differences between the output of the baseline neural network and the output of the retrained neural network across a benchmark dataset of images.

[0096] The experimental results demonstrate that applying the calibration method of the present invention allows the ACD to perform socket pose estimation and take into account extrinsic and intrinsic accuracy, including that related to the updating of the neural network for the computer vision unit.

[0097] Advantageously, a calibrated ACD can estimate the socket's orientation and complete the mating process across a large ACD workspace. Furthermore, calibrated kinematic parameters achieve adequate accuracy within a defined area, and the robot's accuracy in nearby areas does not change dramatically. Therefore, the ACD can move across a larger workspace and still make relatively accurate predictions.

[0098] To facilitate understanding of the principles of the present invention, references to embodiments shown in the drawings are made, and specific terminology is used to describe the embodiments. However, it should be understood that this is not intended to limit the scope of the present invention. Any changes and further modifications to the embodiments described, and any further applications of the principles of the present invention described herein, are contemplated, as would normally be conceivable to those skilled in the art.

[0099] In some instances, one or more components may be referred to herein as “configured to,” “configured by,” “configurable to,” “operable to,” “adapted to,” “capable of,” or “conformed to.” Those skilled in the art will recognize that such terms (e.g., “configured to”) generally encompass active state components and / or inactive state components and / or standby state components, unless otherwise required by context.

[0100] In particular, conditional statements used herein, such as “can,” “could,” “might,” “may,” and “for example,” generally convey that some features, elements, and / or steps are included in some embodiments but not in others, unless otherwise specified or understood in the context in which they are used. Therefore, such conditional statements generally do not imply that features, elements, and / or steps are required in any way for one or more embodiments, nor do they imply that one or more embodiments necessarily include logic for determining, with or without author input or prompts, whether these features, elements, and / or steps are included or should be implemented in any particular embodiment.

[0101] Terms such as “comprising,” “including,” and “having” are synonymous, used in an open-ended manner, comprehensively, and without excluding additional elements, features, actions, or behaviors. Furthermore, the term “or” is used in its inclusive sense (and not in its exclusive sense), for example, when used to connect a list of elements, so that the term “or” means one, some, or all of the elements in the list. Additionally, the articles “a,” “an,” and “the” used in this application and the attached claims should be interpreted as meaning “one or more” or “at least one” unless otherwise specified. As used herein, phrases referring to “at least one of” or “and / or” a list of items refer to any combination of those items, including a single member.

[0102] Similarly, while method steps may be shown in a diagram in a specific order, it should be recognized that such method steps do not need to be performed in the specific or sequential order shown, or that not all shown actions need to be performed, in order to achieve the desired result. Furthermore, a diagram may schematically illustrate another exemplary process in the form of a flowchart. However, other actions not shown may be incorporated into the schematicly shown exemplary methods and processes. For example, one or more additional actions may be performed before, after, simultaneously with, or in between any of the actions described. Furthermore, those actions may be rearranged or reordered in other implementations. In some situations, multitasking and parallel processing may be advantageous.

[0103] The detailed descriptions provided above are essentially illustrative, and it should be understood that variations that do not deviate from the essence and / or spirit of the claimed subject matter are intended to fall within the scope of the claims. Such variations should not be considered deviations from the spirit and scope of the claimed subject matter.

Claims

1. A method for calibrating an automatic charging device (ACD), wherein the ACD comprises a camera, an end effector for supporting and moving a vehicle charging connector, a motion control unit, and a computer vision unit for socket attitude estimation, wherein the computer vision unit comprises at least a baseline neural network, and the motion control unit is configured to control the motion of the effector for plugging the connector into the socket using attitude estimation from the computer vision unit using an image of the vehicle socket recorded by the camera. The aforementioned method, a) Obtaining baseline inaccuracy between at least one target orientation of the end effector and at least one actual orientation of the end effector, - The target posture is the posture to which the motion control unit controls the end effector to move toward it, and the actual posture is the final posture to which the end effector has effectively moved toward it. - The target attitude is determined using the output of the baseline neural network. Obtaining baseline inaccuracies and b) Adjusting the motion control unit by setting a baseline correction function to compensate for the acquired baseline inaccuracy, c) Receiving the updated neural network for the computer vision unit, d) Determining the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, e) Calibrating the motion control unit by updating the baseline correction function based on the determined variation in the attitude estimation inaccuracy. A method characterized by including

2. The method according to claim 1, wherein setting a baseline correction function includes assigning at least one offset to the motion control unit to correct the acquired inaccuracy, the offset including a rotational offset, a translational offset, or a combination thereof.

3. The method according to claim 1, wherein the baseline inaccuracy includes at least one of the following: the baseline neural network inaccuracy, the pose estimation inaccuracy, the inaccuracy due to the ACD hardware components, the inaccuracy due to the ACD software components, and the inaccuracy in the motion control unit.

4. The method according to any one of the preceding claims, wherein the baseline correction function includes at least one baseline correction subfunction, each correcting at least one of the following: inaccuracies in the baseline neural network, inaccuracies in pose estimation, inaccuracies due to ACD hardware components, inaccuracies due to ACD software components, inaccuracies in the motion control unit, or combinations thereof.

5. The method according to claim 4, wherein the at least one subfunction corrects only the inaccuracy in pose estimation.

6. The method according to claim 4, wherein the at least one subfunction includes assigning at least one of a rotational offset and a translational offset to the motion control unit in order to compensate for the acquired inaccuracy.

7. The method according to any one of the preceding claims, wherein the baseline correction function is selected from at least one or more of a black-box model, a gray-box model, or a white-box model, the model is a parameterized model or a non-parameterized model, and the baseline correction function is set as an inclusive system calibration function, or as a function for calibrating the ACD or a subsystem of the ACD, or a combination thereof.

8. The method according to any one of the preceding claims, wherein the baseline inaccuracy is obtained by measuring the rotational and / or translational compliance motion generated by the charging connector by the self-aligning components of the charging connector and / or the guide surface of the socket when connected to the socket, the compliance motion comprising at least one of deflection and deformation of the compliant components.

9. The method according to any one of the preceding claims, wherein the baseline inaccuracy is obtained by instructing the ACD to move the charging connector to the at least one target posture and by measuring the translational and rotational differences between the at least one target posture and each of the at least one actual postures, and the baseline inaccuracy is obtained by measuring the translational and rotational differences between a plurality of target postures and each of their actual postures.

10. Here, the baseline inaccuracy is obtained by iteratively determining the baseline inaccuracy for different target postures and their respective actual postures, until the iterative step no longer produces a significant difference in the obtained baseline inaccuracies, and the different estimated postures cover the majority of postures in the ACD's working space.

11. The method according to any one of the preceding claims, wherein the baseline correction function includes at least one of a qualitative correction instruction and a quantitative correction instruction that enables the ACD to achieve good plugging of the connector into the vehicle socket.

12. The method according to any one of the preceding claims, wherein the baseline correction function is set to obtain an acceptable difference between the target attitude and the actual attitude and / or adjust the target attitude so that the ACD can successfully insert and remove the vehicle charging connector into and out of the vehicle socket.

13. The method according to any one of the preceding claims, wherein the baseline correction function includes at least one offset assigned to at least one of the three-dimensional spatial coordinates, and the motion control unit is adjusted by assigning the correction function to the motion control unit.

14. The method according to any one of the preceding claims, wherein the baseline correction function is a function of the orientation of the end effector in the workspace when the baseline inaccuracy is determined.

15. The method according to any one of the preceding claims, wherein the variation in pose estimation inaccuracy is determined by measuring the difference in inaccuracy of the pose estimation output of the baseline neural network and the updated neural network to a benchmark dataset comprising multiple images on which the ACD is likely to plug in or plug in well.

16. The method according to claim 15, wherein the images in the benchmark dataset include means for determining ground truth measurements of the socket posture relative to the camera.

17. The method according to claim 15, wherein the benchmark dataset includes a plurality of images annotated with at least one ground truth label.

18. The method according to any one of the preceding claims, wherein the variation in the pose estimation inaccuracy is determined by subtracting the output of the updated neural network from the output of the baseline neural network.

19. The method according to any one of the preceding claims, wherein the variation in the pose estimation inaccuracy is the average of the differences between the output of the baseline neural network and the output of the retrained neural network across a benchmark dataset of images.

20. The method according to any one of the preceding claims, wherein the variation in the pose estimation inaccuracy is determined by comparing the absolute socket pose estimation error function of the updated neural network with respect to the ground truth with the absolute socket pose estimation error function of the baseline neural network with respect to the ground truth, the socket pose estimation error function is determined by comparing the output of each of the neural networks with the ground truth measurement.

21. The method according to claim 20, wherein the means for determining the ground truth includes at least one reference marker placed at a known position relative to the socket.

22. The method according to any one of the preceding claims, wherein the baseline correction function is updated by adding the variation in the attitude estimation inaccuracy to the attitude estimate of the computer vision module.

23. The method according to any one of the preceding claims, wherein the baseline correction function is updated by using the updated absolute pose estimation error function of the neural network instead of adjusting the baseline correction function.

24. The method according to any one of the preceding claims, wherein the baseline correction function is updated by adding a six-degree-of-freedom constant offset to the output of the computer vision module, the offset being determined by subtracting the six-degree-of-freedom output of the computer vision module when the updated neural network runs the benchmark dataset from the six-degree-of-freedom output of the computer vision module when the benchmark dataset is run by the baseline neural network.

25. The method according to any one of the preceding claims, wherein the baseline correction function is updated from a remote device via a communication device, and the remote device compares the variation in the attitude estimation inaccuracy with an optimal variation range from a database of past decisions, and adjusts a threshold based on the optimal variation.

26. The method according to any one of the prior claims, further comprising deploying the updated neural network to the computer vision unit.

27. The method according to claim 26, wherein the neural network is deployed to the computer vision unit when an improvement in pose estimation is observed between the baseline neural network and the updated neural network.

28. A method for updating an automatic charging device (ACD), wherein the ACD is - End effector for supporting and moving the vehicle charging connector, - Camera and, - Motion control unit, - Computer vision unit for socket pose estimation and Equipped with, The computer vision system comprises at least a baseline neural network, and the motion control unit is configured to control the motion of the effector for plugging the connector into the socket using attitude estimation from the computer vision unit using images of the vehicle socket recorded by the camera. The motion control unit includes a baseline correction function that compensates for the inaccuracy between the actual posture and the target posture. The aforementioned target posture is the posture to which the motion control unit controls the end effector to move toward it. The aforementioned actual position is the final position in which the end effector is effectively moved toward it. The method described above is d) Receiving the updated neural network for the computer vision unit, e) Determining the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, f) Calibrating the motion control unit by updating the baseline correction function based on the determined variation in the attitude estimation inaccuracy. Methods that include...

29. It is an automatic charging device, - End effector for supporting and moving the vehicle charging connector, - Camera and, - Motion control unit, - A computer vision unit for pose estimation, equipped with a baseline neural network, - The motion control unit is configured to control the motion of the end effector for plugging the connector into the socket using attitude estimation from the computer vision unit using an image of the vehicle socket recorded by the camera, and the motion control unit includes a baseline correction function that corrects for inaccuracies between the actual attitude and the target attitude, - Control device and Equipped with, The control device, - Receiving the updated neural network for the aforementioned computer vision unit, - To determine the variation in pose estimation inaccuracy between the baseline neural network and the updated neural network, - Calibrating the motion control unit by updating the baseline correction parameters based on the determined variation in the attitude estimation inaccuracy. An automatic charging device characterized by being configured to perform the following actions.

30. The control device is - To obtain the baseline inaccuracy between the target pose of the end effector and the actual pose of the end effector, and to generate a baseline correction function based on the baseline inaccuracy, wherein the target pose is determined using the output of a computer vision baseline neural network, - Adjusting the motion control unit by setting baseline correction parameters using the baseline correction function to compensate for the acquired inaccuracies. The automatic charging device according to claim 29, further configured to perform the following:

Images