Reinforcement learning-based object location estimation and catching method and system

Abstract

A reinforcement learning object location estimation method according to one embodiment of the present invention comprises the steps of: acquiring a plurality of image data captured from a camera or sensor; detecting an object in the plurality of image data using a first neural network based on a convolutional neural network (CNN) model; when an object is detected, identifying the location of the object within each of the image data using a second neural network based on a convolutional neural network (CNN) model and vectorizing the location information and the image data to generate an embedding vector; and using a reinforcement learning model to take the embedding vector as input data and output object trajectory prediction information. The step of detecting an object includes the step of training a first neural network using a set of training image data so that the first neural network outputs a binary classification of object-presence or object-non-presence for each input data; and the step of applying the trained first neural network to the acquired image data to automatically classify the image data into object-presence or object-non-presence classifications to detect the object, thereby providing an object location estimation method exhibiting excellent retention performance.

Description

Technology Field

[0001] The present invention relates to a method for estimating the location of an object and catching it using reinforcement learning, and more specifically, to a method for detecting objects of various sizes using a convolutional neural network (CNN), identifying the movement trajectory of an object, predicting the movement trajectory of an object using a transformer model, and grasping it using a robot. Background Technology

[0002] Recently, a vast amount of image data of various types is being generated, and diverse technologies are being developed to automate its processing. In particular, with the advancement of artificial neural network technology, processing that automatically learns and classifies image data using this technology is being applied in the industry.

[0004] Among these artificial neural network technologies, the relatively widely used Convolutional Neural Network (CNN) technology has a structure consisting of one or more convolutional layers, pooling layers, and fully connected layers, and is known to be mainly used for the analysis of two-dimensional input data such as images and voice.

[0006] Conventionally, deep learning-based data analysis methods have been studied to identify the movement of objects and predict their trajectories, and in particular, many studies have been conducted to identify the movement of objects and predict their trajectories using LSTM models specialized for time series analysis.

[0008] However, existing artificial neural network-based object orbit prediction methods had a limitation in that the accuracy of orbit prediction was low.

[0009] delete Prior art literature

[65535] BALL TRAJECTORY TRACKING (US 2022-0044423 A1, 2022.02.10.) The problem to be solved

[0010] The present invention is designed to solve the problems described above. An object location estimation system according to one embodiment of the present invention estimates the presence or absence of an object in an image through a neural network, identifies the location of the object in the image, and then uses a reinforcement learning model to predict the movement trajectory of the object, including the time of arrival and arrival trajectory of the object in a pre-set area, and provides a system for grasping an object in a pre-set area using a grasping unit. means of solving the problem

[0012] An object location estimation method according to an embodiment of the present invention comprises: a step of acquiring a plurality of image data captured from a camera or sensor; a step of detecting an object in the plurality of image data using a first neural network based on a convolutional neural network (CNN) model; a step of generating an embedding vector in which, when an object is detected, the location of the object within each of the image data is identified using a second neural network based on a convolutional neural network (CNN) model and vectorizes the location information and the image data; and a step of predicting an object's orbit using a reinforcement learning model, wherein the embedding vector is used as input data and the object's orbit prediction information is output. The step of detecting an object is characterized by comprising: a step of training a first neural network using a set of training image data so that the first neural network outputs a binary classification of object-presence or object-non-presence for each input data; and a step of applying the trained first neural network to the acquired image data to automatically classify the image data into object-presence or object-non-presence classifications to detect an object. Effects of the invention

[0014] An object position estimation and retention method using reinforcement learning according to one embodiment of the present invention learns the patterns of continuous image data using a decision transformer model in which reinforcement learning is implemented as sequence modeling, thereby providing an object position estimation method and device that exhibit excellent orbit prediction performance. Brief explanation of the drawing

[0016] FIG. 1 is a block diagram of an AI device according to one embodiment of the present invention. FIG. 2 is a diagram illustrating a catching system using a reinforcement learning-based robot according to one embodiment of the present invention. FIG. 3 is a diagram illustrating a framework of a catching method using a reinforcement learning-based robot according to an embodiment of the present invention. FIG. 4 is a diagram illustrating a first neural network and a second neural network according to an embodiment of the present invention. FIG. 5 is an algorithm diagram of a catching method using a reinforcement learning-based robot according to an embodiment of the present invention. FIG. 6 is a flowchart of a catching method using a reinforcement learning-based robot according to an embodiment of the present invention. FIG. 7 is a diagram comparing the performance of a catching method using a reinforcement learning-based robot according to one embodiment of the present invention. Specific details for implementing the invention

[0017] Hereinafter, embodiments disclosed in this specification will be described in detail with reference to the attached drawings. Identical or similar components regardless of drawing symbols will be assigned the same reference number, and redundant descriptions thereof will be omitted. The suffixes "module" and "part" used for components in the following description are assigned or used interchangeably solely for the ease of drafting the specification and do not inherently possess distinct meanings or roles. Furthermore, in describing the embodiments disclosed in this specification, if it is determined that a detailed description of related prior art could obscure the essence of the embodiments disclosed in this specification, such detailed description will be omitted. Additionally, the attached drawings are intended only to facilitate understanding of the embodiments disclosed in this specification; the technical concept disclosed in this specification is not limited by the attached drawings, and it should be understood that they include all modifications, equivalents, and substitutions that fall within the concept and technical scope of this specification.

[0018] Terms including ordinal numbers, such as first, second, etc., may be used to describe various components, but said components are not limited by said terms. These terms are used solely for the purpose of distinguishing one component from another.

[0019] When it is stated that one component is "connected" or "connected" to another component, it should be understood that while it may be directly connected or connected to that other component, there may also be other components in between. On the other hand, when it is stated that one component is "directly connected" or "directly connected" to another component, it should be understood that there are no other components in between.

[0020] A singular expression includes a plural expression unless the context clearly indicates otherwise.

[0021] In this application, terms such as “comprising” or “having” are intended to specify the existence of the features, numbers, steps, actions, components, parts, or combinations thereof described in the specification, and should be understood as not precluding the existence or addition of one or more other features, numbers, steps, actions, components, parts, or combinations thereof.

[0023] reinforcement learning

[0024] The Reinforcement Learning class is It is represented as a standard MDP (Markov Decision Process) with tuples. Here is the state space, is an action space, is the reward space, is the probability of transition It is a space of, and is the discounting factor. A reinforcement learning agent (or reinforcement learning algorithm) is a state If received, stochastic policy Activities according to Decide to perform, and then the reward Obtains. The main goal of reinforcement learning is cumulative reward It is to maximize.

[0026] FIG. 1 is a block diagram of an AI device according to one embodiment of the present specification.

[0027] The AI ​​device (20) may include an electronic device including an AI module capable of performing AI processing, or a server including said AI module. The device of the present invention refers to an AI device. Additionally, the AI ​​device (20) may be configured to be included as a component of at least part of an electronic device to perform at least part of the AI ​​processing together.

[0028] The AI ​​device (20) may include an AI processor (21), memory (25) and / or a communication unit (27).

[0029] The AI ​​device (20) is a computing device capable of learning neural networks and can be implemented as various electronic devices such as a server, desktop PC, laptop PC, tablet PC, etc.

[0030] The AI ​​processor (21) can train a neural network using a program stored in memory (25). In particular, the AI ​​processor (21) can generate an artificial intelligence model for estimating action prediction information, which is the allocation ratio of stock items, and can reinforce this artificial intelligence model using collected stock price data and technical indicator data.

[0031] Meanwhile, the AI ​​processor (21) that performs the functions described above may be a general-purpose processor (e.g., CPU), but may be an AI-dedicated processor for artificial intelligence learning (e.g., GPU, graphics processing unit).

[0032] The memory (25) can store various programs and data required for the operation of the AI ​​device (20). The memory (25) may be implemented as non-volatile memory, volatile memory, flash memory, hard disk drive (HDD), or solid-state drive (SDD), but is not limited thereto. For example, the memory may include other external storage devices that can be accessed by the device.

[0033] The memory (25) is accessed by the AI ​​processor (21), and the AI ​​processor (21) can perform reading / writing / modifying / deleting / updating of data.

[0034] Additionally, the memory (25) can store a neural network model (e.g., a deep learning model) generated through a learning algorithm for data classification / recognition according to one embodiment of the present specification.

[0035] Meanwhile, the AI ​​processor (21) may include a data learning unit that learns a neural network for data classification / recognition. For example, the data learning unit may learn a deep learning model by acquiring training data to be used for learning and applying the acquired training data to a deep learning model.

[0036] The communication unit (27) can transmit the AI ​​processing results by the AI ​​processor (21) to an external electronic device.

[0037] Here, external electronic devices may include other terminals and servers.

[0038] Meanwhile, although the AI ​​device (20) illustrated in FIG. 1 is described by functionally separating it into an AI processor (21), memory (25), and communication unit (27), the aforementioned components may be integrated into a single module and referred to as an AI module or an artificial intelligence (AI) model.

[0040] Hereinafter, with reference to FIGS. 2 to 7, a stock trading system and a stock trading method using reinforcement learning according to the present invention will be described.

[0042] [Object Location Estimation System Using Reinforcement Learning Models]

[0043] An object location estimation system according to one embodiment of the present invention means a system for estimating the presence or absence of an object in an image through a neural network, identifying the location of an object in an image through a neural network, and predicting the movement trajectory of an object including the time of arrival and arrival trajectory of the object in a preset area using a reinforcement learning model, and for grasping an object in a preset area using a grasping unit.

[0044] FIG. 2 is a diagram illustrating a catching system using a reinforcement learning-based robot according to one embodiment of the present invention, and FIG. 3 is a diagram illustrating a framework of a catching method using a reinforcement learning-based robot according to one embodiment of the present invention.

[0045] Referring to FIGS. 2 and FIGS. 3, an object position estimation system according to one embodiment of the present invention may include a shooting unit, a position estimation device, and a gripping unit.

[0046] The capturing unit may include a camera, an image sensor, an infrared sensor, etc., and refers to various devices and sensors capable of acquiring an image or sensing data of an object. The capturing unit may acquire image data or sensing information and transmit it to a position estimation device.

[0048] [Location Estimation Device - AI Device]

[0049] The object location estimation device according to the present invention means a device that estimates the presence of an object within image data based on data acquired from a shooting unit, and if an object exists, estimates the movement trajectory of the object by identifying the location of the object within the image, thereby estimating the location of the object after time t, and transmits the corresponding information to a gripping unit.

[0050] A location estimation device according to the present invention may include a memory in which at least one program is stored and a processor that estimates the location of an object by executing at least one program.

[0051] In the present invention, only the components of the device related to the present embodiment are described. Therefore, those skilled in the art related to the present embodiment will understand that other general-purpose components may be further included in the device in addition to the described components. For example, the device may further include a display (not shown), a communication device (not shown), and an output device (not shown). The display may be used to display a plurality of image data or images, along with final positions determined for each of the plurality of image data.

[0053] Memory is hardware that stores various data processed within a device; for example, memory can store data processed by the device and data to be processed. Additionally, memory can store applications, drivers, etc., to be executed by the device.

[0055] The processor can play an overall role in controlling the device. The processor can be implemented by one or more processors. For example, the processor can be implemented as an array of multiple logic gates, or as a combination of a general-purpose microprocessor and memory storing a program that can be executed on the microprocessor.

[0057] FIG. 4 is a diagram illustrating a first neural network (detection network) and a second neural network (embedding network) according to an embodiment of the present invention.

[0058] Referring to FIG. 4, the processor can estimate the orbit and position of an object by executing at least one program. Specifically, the processor can acquire multiple image data captured by a camera, determine the presence or absence of an object within the multiple image data using a first neural network, and if it is estimated that an object exists, determine the position of the object within the acquired image using a second neural network, and generate an embedding vector by embedding the position information of the object and the image.

[0060] More specifically, the processor acquires multiple image data captured through the camera unit and can estimate the presence or absence of an object within the acquired image using a first neural network. That is, the first neural network may be a model trained to output a binary classification of the presence or absence of an object for the input images. Data sets used for training may be generated using a simulator. In this case, to improve the object detection performance of the first neural network, data sets regarding objects of various sizes and colors may be generated using a simulator.

[0062] The first neural network may be a Deep Neural Network (DNN) or an n-layer neural network including two or more hidden layers. For example, the first neural network may be a Convolutional Neural Network (CNN). However, it is not necessarily limited thereto, and the first neural network may include Recurrent Neural Networks (RNN), Deep Belief Networks, Restricted Boltzmann Machines, etc. According to one embodiment of the present invention, the first neural network may be a CNN model trained to classify the presence or absence of an object.

[0064] The first neural network can be trained through supervised learning or reinforcement learning. For example, the first neural network can be trained to output the presence or absence of an object when image data is input, using pairs of image data training data generated by a simulator. However, it is not necessarily limited to this, and the first neural network can also be trained through unsupervised learning.

[0066] When image data is input, the trained first neural network can automatically classify the image data into object-presence or object-non-presence classifications and output results.

[0068] When the processor classifies that an object exists within an image through the first neural network, it can input multiple classified image data into the second neural network. The second neural network can estimate the location of the object in each of the input multiple images and generate an embedding vector by embedding the estimated location information and image data.

[0070] The second neural network may be a DNN or n-layers neural network containing two or more hidden layers. For example, the second neural network may be a CNN. However, it is not necessarily limited to this, and the second neural network may include an RNN, Deep Belief Networks, Restricted Boltzmann Machines, etc.

[0072] The reinforcement learning model may be a model that takes embedding vectors generated from a second neural network as input data and object trajectory prediction information as output data. The reinforcement learning model may be a transformer that includes multiple decoder blocks. The transformer identifies the correlation between the input embedding vectors. That is, it can estimate the object's movement information and trajectory through the correlation between the object's position at the time the image data was captured, and predict the object's movement speed, position after t seconds, and trajectory.

[0074] In other words, reinforcement learning models acquire information regarding an object's location, direction of movement, and speed through correlations between input data, and can predict information regarding the object's movement trajectory, location, and speed after the image is acquired.

[0076] If necessary, the reinforcement learning model can estimate the time and location information of an object reaching a pre-set area. The pre-set area may refer to an area where a gripping unit capable of grasping an object or a robot can move or reach within a certain time. Accordingly, the reinforcement learning model estimates the time and location information of the object reaching the pre-set area, and based on this, enables the gripping unit or robot to move and grasp the object. In this case, the location information may include the trajectory or direction information of the object's movement. The gripping unit may be in the form of a gripper, and a gripper-type gripping unit can grasp an object in a specific direction. Accordingly, the reinforcement learning model predicts the trajectory and direction information of the object's movement, generates information on the angle (θ) at which the object enters the pre-set area, and transmits this information to the gripping unit, thereby enabling the gripping unit to correctly grasp the object.

[0078] The first neural network, the second neural network, and the reinforcement learning model can each be trained individually.

[0080] Decision Transformer

[0081] If necessary, the present invention may utilize a decision transformer model that implements reinforcement learning as sequence modeling using a Generative Pre-trained Transformer (GPT) model, which is a type of transformer model. The decision transformer model is a model in which modalities such as state, action, and reward are passed to an embedding network and processed by an autoregressive transformer model trained to predict the next action from the previous token using a linear output layer. Autoregressive refers to a model that predicts the current self using the past self, and the decision transformer model is a model capable of sequence modeling. The core layer of the reinforcement learning module is the decoder block, and the decoder block is composed of an attention mechanism so that correlations are learned through the inner product of input elements. If necessary, the reinforcement learning model according to the present invention may include an attention mechanism.

[0083] A gripping unit refers to a device or gripper for grasping an object while moving within a preset area or a specific area. According to one embodiment of the present invention, the gripping unit may be composed of a robot arm and a gripper. The gripping unit obtains information such as object trajectory prediction information, object arrival time information, arrival position information, and incident angle within the preset area from a position estimation device, and can grasp an object by moving the gripping unit.

[0085] According to one embodiment, the gripping unit can grasp an object by moving the gripper to a position where the object will be reached after t seconds, and then performing a gripping operation after t seconds have elapsed.

[0087] According to another embodiment, the gripping unit can control the gripper to grip the object by acquiring information on the angle (θ) at which an object enters a preset area. The angle (θ) can be calculated according to the following formula.

[0088]

[0090] Hereinafter, a method for estimating the location of an object using a reinforcement learning model according to another embodiment of the present invention will be described.

[0091] FIG. 5 is an algorithm diagram of a catching method using a reinforcement learning-based robot according to an embodiment of the present invention, and FIG. 6 is a flowchart of a catching method using a reinforcement learning-based robot according to an embodiment of the present invention.

[0092] Referring to FIGS. 5 and 6, an object location estimation method using a reinforcement learning model according to one embodiment of the present invention may include an image data acquisition step, an object detection step, an embedding vector generation step, and an orbit prediction step.

[0093] The image data acquisition step (S1100) means the step of acquiring image data captured from a camera or sensor.

[0094] The object detection step (S1200) refers to a step of detecting objects in multiple image data using a first neural network based on a convolutional neural network (CNN) model. If necessary, the object detection step may separate and classify only the images in which objects exist among the acquired images.

[0095] The object detection step (S1200) may include a step of training a first neural network and a step of automatically classifying and detecting objects.

[0096] The first neural network may be a model trained to output a binary classification of the presence or absence of an object for input images. That is, the device can input acquired image data into the trained first neural network to estimate the presence or absence of an object.

[0098] The step of generating an embedding vector (S1300) is a step of identifying the location of an object within each image data using a second neural network based on a convolutional neural network (CNN) model when an object is detected to exist through a first neural network or classified to exist through a first neural network, and generating an embedding vector by vectorizing the object's location information and the image data.

[0100] The orbit prediction stage refers to the step of identifying correlations between input embedding vectors using a reinforcement learning model and outputting predicted orbit information of the object.

[0101] Correlation may refer to the correlation between the positions of objects within each image data, and orbit prediction information may include movement information, including the object's movement speed and direction of movement, and orbit information.

[0102] If necessary, the orbit prediction step may include a step of identifying correlations between the embedding vectors using a transformer model including a plurality of decoder blocks, an orbit prediction step of predicting the orbit of an object, and a step of predicting the arrival time and arrival location information of the object within a preset region.

[0103] The step of predicting arrival location information may include the step of predicting the angle (θ) at which the object enters a pre-set area.

[0104] According to one embodiment, the angle (θ) at which an object enters can be calculated based on trajectory prediction information. This information is used to facilitate the robot or gripping unit from grasping the object, and can be calculated according to the following formula.

[0105]

[0106] If necessary, the trajectory prediction step may include a step of predicting the trajectory of an object, including information on the position, movement trajectory, and movement direction of the object after a preset time, through the correlation between input embedding vectors using a reinforcement learning model.

[0108] In this case, the reinforcement learning model is a decision transformer-based reinforcement learning model, meaning a model that identifies correlations between input data. Furthermore, the reinforcement learning model of the present invention can estimate the trajectory of an object by identifying correlations between the continuously acquired image data using an attention mechanism-based model.

[0110] If necessary, the method for estimating an object's position using a reinforcement learning model according to one embodiment of the present invention may further include the step of generating a control signal and transmitting it to a robot so that the robot can grasp an object based on the result of trajectory prediction. The control signal may include information regarding the time taken to reach a preset region of the object, the angle of reaching the preset region, and the position of reaching the preset region, and the robot may move based on the control signal to grasp the object. Additionally, the robot may grasp the object by moving to the corresponding position before the object is reached based on the acquired control signal, and then executing a grasping action in accordance with the time taken to reach the object.

[0112] FIG. 7 is a diagram comparing the performance of a reinforcement learning-based catching method according to one embodiment of the present invention.

[0113] In other words, this experiment predicted the trajectory using reinforcement learning models such as CNN, LSTM, and Transformer, and then evaluated the performance of the models based on two criteria. Specifically, the performance of the models was evaluated in terms of accessibility of the gripper (whether the tip of the gripper touched the ball) and catching (whether the object was successfully caught).

[0114] Each model was tested 30 times, with 10 trials each for golf balls, tennis balls, and baseballs, and the success rates are shown in Fig. 7. The Transformer model according to the present invention showed the best results. The performance of the Transformer was 93.3% for accessibility and 73.3% for catching, which means that the Transformer model analyzes sequential image data excellently, with only slight differences in accuracy. On the other hand, other models showed poor catching performance and even poor accessibility. The success rate for all criteria was less than 30%. The LSTM model showed the lowest performance and appears unable to predict the trajectory of the ball at all. In the case of the LSTM model, it was expected to show superior performance compared to CNN because it is specialized for learning sequential data; however, it was confirmed that the CNN algorithm showed superior performance. This was assumed to be due to the loss of positional information in the image embeddings when predicting trajectories using LSTM and CNN, whereas, conversely, it appears to be because the Transformer learned the trajectory of the ball well.

[0116] That is, the object position estimation and retention method using reinforcement learning according to the present invention utilizes a decision transformer model in which reinforcement learning is implemented as sequence modeling to learn the patterns of continuous image data, thereby providing an object position estimation method and device that exhibit excellent orbit prediction performance.

Claims

Claim 1 A method for estimating the location of an object using an object location estimation system comprises: a step of acquiring a plurality of image data captured from a camera or sensor; a step of detecting an object in the plurality of image data using a first neural network based on a convolutional neural network (CNN) model; a step of generating an embedding vector in which, when an object is detected, the location of the object within the image data is identified using a second neural network based on a convolutional neural network (CNN) model and the location and the image data are vectorized; and a trajectory prediction step in which the embedding vector is used as input data and trajectory prediction information of the object is output using a reinforcement learning model. The step of detecting the object includes a step of training the first neural network using a set of training image data so that the first neural network outputs a binary classification of object-presence or object-non-presence for each input data. A reinforcement learning-based object location estimation method comprising: a step of detecting an object by applying the above-mentioned trained first neural network to the above-mentioned image data to automatically classify the above-mentioned image data into object-presence or object-non-presence classifications; wherein the orbit prediction step comprises: a step of identifying correlations between the above-mentioned input embedding vectors using a transformer model including a plurality of decoder blocks; an orbit prediction step of predicting the orbit of the object; and a step of predicting the arrival time and arrival location information of the object within a pre-set region. Claim 2 delete Claim 3 A reinforcement learning-based object position estimation method according to claim 1, wherein the step of predicting the arrival position information includes the step of predicting the angle (θ) at which the object enters the preset region; and further includes the step of generating a control signal and transmitting it to the robot so that the robot can grasp the object based on the result of the trajectory prediction. Claim 4 In claim 1, the reinforcement learning model estimates the orbit of the object by identifying the correlation between the image data acquired continuously using an attention mechanism-based model. Claim 5 A reinforcement learning-based object position estimation method according to claim 1, wherein the trajectory prediction step comprises: a step of predicting the trajectory including information on the position, movement trajectory, and movement direction of the object after a preset time through the correlation between the input embedding vectors. Claim 6 A device for estimating the location of an object comprises: a memory in which at least one program is stored; and a processor that estimates the location of the object by executing the at least one program; wherein the processor acquires a plurality of images through a camera, estimates the presence of an object in the plurality of images using a first neural network, and when the object is estimated to be present, identifies the location of the object within the image using a second neural network, generates an embedding vector by embedding the location information of the object and the images, inputs the embedding vector into a reinforcement learning model to output trajectory prediction information of the object, wherein the first neural network is a Convolutional Neural Network (CNN)-based network trained to output a binary classification of object presence or object non-presence for input data, and the reinforcement learning model is a Transformer, and the processor uses the Transformer, which includes a plurality of decoder blocks, to identify the correlation between the input embedding vectors, predict the trajectory of the object, and predict the arrival time and arrival location information of the object within a preset region. Claim 7 In claim 6, the processor predicts the arrival time and arrival location information of the object within a preset region and the angle (θ) at which the object is incident within the preset region, and generates a control signal to enable the robot to grasp the object based on the result of the trajectory prediction and transmits it to the robot. Claim 8 In claim 6, the reinforcement learning model is a device that estimates the trajectory of the object by identifying the correlation between the images acquired consecutively using an attention mechanism-based model. Claim 9 An object position estimation system comprises: a capturing unit for acquiring image data; an object position estimation device for detecting an object from the acquired image data and predicting the trajectory of the object; and a grasping unit that moves in a preset area, acquires information on the time and location of the object's arrival within the preset area, and grasps the object using the acquired information; wherein the object position estimation device includes a memory in which at least one program is stored; An object location estimation system comprising: a processor that estimates the location of an object by executing at least one program; wherein the processor acquires a plurality of image data through the capturing unit, estimates the presence of an object within the plurality of image data using a first neural network, and when it is estimated that the object exists within the image, identifies the location of the object within the image data using a second neural network, generates an embedding vector by embedding the location information of the object and the image data, inputs the embedding vector into a reinforcement learning model to output orbit prediction information of the object, wherein the first neural network is a network based on a CNN model trained to output a binary classification of object presence or object non-presence for input data, and the reinforcement learning model is a transformer model based on an attention mechanism, and the processor identifies the correlation between the input embedding vectors using the transformer model including a plurality of decoder blocks, predicts the orbit of the object, and predicts the arrival time and arrival location information of the object within a preset region. Claim 10 In claim 9, the processor predicts the arrival time and arrival location information of the object within a preset region and the angle (θ) at which the object enters the preset region, and generates a control signal based on the result of the trajectory prediction to enable the gripping unit to grip the object and transmits the control signal to the gripping unit.

Images