Machine learning systems for technical knowledge acquisition
A machine learning system processes SME data to create domain models for training novice users, addressing inefficiencies in traditional training methods by capturing tacit knowledge and enhancing learning through interactive multimedia, thus reducing costs and improving training efficiency.
Patent Information
- Application Number
- JP2023019137
- Authority / Receiving Office
- JP · JP
- Patent Type
- Patents
- Current Assignee / Owner
- Priority Date
- 2019-12-23
- Filing Date
- 2023-02-10
- Publication Date
- 2024-11-19
- Estimated Expiration
- 2040-12-23
AI Technical Summary
Existing training methods for complex tasks in workplaces, such as those requiring tacit knowledge, are inefficient and costly, often relying on subject matter experts (SMEs) who are scarce and prone to omitting important information, and traditional training materials like videos and manuals do not allow for interactive learning or effective transfer of tacit knowledge.
A machine learning system that utilizes video, audio, and sensor data from SMEs performing tasks to create domain models, correlating and processing this data to generate training information for novice users, enabling the capture and transfer of both explicit and tacit knowledge through augmented reality and interactive multimedia manuals.
The system efficiently acquires and transfers complex task knowledge, reducing training costs and time by capturing fine motor movements and subjective information, providing interactive training that enhances learning efficiency and accuracy.
Smart Images

Figure 0007586945000001 
Figure 0007586945000002 
Figure 0007586945000003
Abstract
Description
[Technical field]
[0001] This disclosure relates generally to machine learning systems, and more specifically, to machine learning for knowledge acquisition. [Background technology]
[0002] Workplaces use training programs to train employees to develop knowledge or skills to improve their performance in their roles. However, the more complex the task, the more training is required to successfully teach an employee to perform the task. Training a new employee to perform a highly complex task can be costly and time consuming. As another example, a workplace may use subject matter experts (SMEs) to teach trainees to perform specific tasks. SMEs are individuals, often employees or consultants, who have a deep understanding of a particular job, line of work, department, function, technology, machine, material, or type of equipment. Typically, SMEs develop expertise through a combination of experience and training. It may take an SME years to develop this.
[0003] Some workplaces may train employees by having SMEs create training videos or written instruction guides to teach employees to perform specific tasks. However, videos and instruction guides may be disadvantageous because employees cannot ask questions. Furthermore, SMEs may inadvertently or unconsciously omit important information. Furthermore, global companies spend significant time and budgets creating multimedia technical manuals to train employees. This may be even more expensive when the desired skills involve highly specialized tasks, tacit knowledge, or multiple languages. As another example, a workplace may have employees observe the SME while the SME performs a task. While observing, or "shadowing," the SME may be valuable for an employee, it may be impractical for a large group of employees and may impair the SME's efficiency. In a method similar to training videos, the SME may unintentionally omit verbal explanations of subtle but important details related to the task they are demonstrating. Additionally, in the workplace, employees may attend training sessions in classroom environments that may not provide valuable hands-on experience with technical tasks, and for highly complex tasks or in specialized fields, there may not be many SMEs available to help train new employees to perform the tasks. Summary of the Invention
[0004] This disclosure generally describes machine learning techniques for capturing human knowledge to perform a task and building or refining a domain model that can be used to train others to perform the task. In one example, a video device acquires video data of a first user performing a task. In some examples, the first user is an SME performing the task. The video data may include multiple camera sources, such as video data from a first person perspective of the first user and / or video data from a third person perspective of the first user. An audio input device acquires audio data describing the performance of the task. In some examples, the audio data includes a narrative by the first user of the first user's actions while performing the task. In other examples, the audio data includes a narrative of how to perform the task, narrated by the first user while the first user is not performing the task. In some examples, one or more sensors generate sensor data during the performance of the task by the first user. In some examples, the sensor data includes accelerometer data, pressure data, or force data related to movements or actions taken by the first user, interactions between the first user and one or more objects, such as a tool, a workpiece, etc. In some examples, the document processing unit may obtain textual data describing the performance of a task, such as from an instruction manual, a parts list, or other written guide.
[0005] In some examples, the computation engine applies a machine learning system to the collected data (e.g., video data, audio data, sensor data, and / or text data) to update a domain model that governs the execution of the task. As described herein, a task may be conceptualized as multiple steps performed to achieve a given goal. In some examples, the machine learning system updates the domain model by correlating objects recognized in the video data with references to objects in the audio data and / or text data as well as measurements in the sensor data, thereby identifying portions of the video data, portions of the audio data, portions of the sensor data, and portions of the text data that describe the same steps of multiple steps for performing the task. For example, a pressure sensor on a plate may record the amount of pressure the SME is applying to the plate in addition to a verbal description from the SME describing his or her movements. A training unit applies the updated domain model to generate training information for performing the task. In some examples, the training unit forms a knowledge database that stores the training information. The training information may include, for example, video data, audio data, sensor data, and text data cross-referenced with one another so that the second user may search the knowledge database by concept, task, or sub-task to obtain training information relevant to the second user's query. An output device outputs the training information for use in training the second user to perform the task. In some examples, the second user is a novice. In some examples, the output device outputs the training data in the form of an augmented reality video depicting the performance of the task from a first-person perspective of the second user.
[0006] The techniques of the present disclosure may provide certain technical improvements to the computer-related field of machine learning that have practical applications. For example, the techniques described herein may enable a machine learning system to fuse diverse data sources of a first user performing a complex task to generate training information useful for training a second user to perform the complex task. For example, the techniques of the present disclosure may enable a machine learning system to acquire data from SMEs performing complex tasks, such that training materials useful for others (such as novice users) to perform the complex task may be created. Furthermore, the techniques of the present disclosure may enable a machine learning system to identify important information in performing a task that may be subjective or difficult to communicate by acquiring various aspects of how an SME performs a task that may be unknown, unrecognized, or subjective to the SME when asked to describe the task. Thus, the techniques of the present disclosure may enable the creation of focused, empirical training information that may increase efficiency and reduce the cost of training employees to perform complex tasks.
[0007] In one example, the disclosure describes a system for acquiring knowledge to perform a task, the system comprising: a domain model that defines the performance of the task; a video input device configured to acquire video data of a first user performing the task; an audio input device configured to acquire audio data describing the performance of the task; one or more sensors configured to generate sensor data during the performance of the task; a computational processing engine configured to correlate at least a portion of the video data with at least a portion of the audio data and at least a portion of the sensor data, and process at least a portion of the correlated video data, at least a portion of the audio data, and at least a portion of the sensor data to update the domain model that defines the performance of the task; a training unit configured to apply the updated domain model to generate training information for performing the task; and an output device configured to output the training information for use in training a second user to perform the task.
[0008] In another example, the present disclosure describes a method for acquiring knowledge to perform a task, the method including: acquiring, by a video input device, video data of a first user performing the task; acquiring, by an audio input device, audio data describing the performance of the task; generating, by one or more sensors, sensor data during the performance of the task; correlating, by a processing engine, at least a portion of the video data with at least a portion of the audio data and at least a portion of the sensor data; processing, by the processing engine, at least a portion of the correlated video data, at least a portion of the audio data, and at least a portion of the sensor data to update a domain model that governs the performance of the task; applying, by a training unit, the updated domain model to generate training information for performing the task; and outputting, by an output device, the training information for use in training a second user to perform the task.
[0009] In another example, the disclosure describes a non-transitory computer-readable medium including instructions that, when executed, are configured to cause a processing circuit to acquire video data of a first user performing a task, acquire audio data describing the performance of the task, generate sensor data during the performance of the task, correlate at least a portion of the video data with at least a portion of the audio data and at least a portion of the sensor data, process at least a portion of the correlated video data, at least a portion of the audio data, and at least a portion of the sensor data to update a domain model that governs the performance of the task, apply the updated domain model to generate training information for performing the task, and output the training information for use in training a second user to perform the task.
[0010] The details of one or more examples of the techniques of this disclosure are set forth in the accompanying drawings and the description below. Other features, objects, and advantages of the techniques will become apparent from the description and drawings, and from the claims. [Brief description of the drawings]
[0011] [Figure 1] FIG. 1 is a block diagram illustrating an example system for generating training information in accordance with the techniques of this disclosure. [Diagram 2] FIG. 1 is a block diagram illustrating an example computing system for generating training information in accordance with the techniques of this disclosure. [Diagram 3] 1 is a flowchart illustrating example operations for generating training information in accordance with the techniques of this disclosure. [Figure 4] FIG. 1 illustrates an example system for generating training information in accordance with the techniques of this disclosure. [Diagram 5] FIG. 1 illustrates an example system for generating training information in accordance with the techniques of this disclosure. [Figure 6] FIG. 1 is a diagram of labeled video data used in generating training information in accordance with the techniques of this disclosure. [Figure 7] FIG. 1 is a diagram of labeled video data used in generating training information in accordance with the techniques of this disclosure. [Figure 8] 11A-11C are diagrams of example user interfaces depicting training information generated in accordance with the techniques of this disclosure. [Figure 9] 1 is a flowchart illustrating example operations for generating training information in accordance with the techniques of this disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0012] Like reference characters refer to like elements throughout the figures and description.
[0013] Knowledge acquisition is the process by which knowledge is transformed from tacit to explicit knowledge. "Tacit" or "implicit" knowledge is a type of knowledge that is contained or embodied in an SME who possesses institutional knowledge of a particular domain or field, or institutional knowledge directed to a particular organization or domain of organizations, through the SME's sheer talent or accumulated experience. For example, examples of tacit knowledge include knowledge that is difficult to describe, such as knowledge of hearing, vision, touch, smell, taste, or other senses that are highly discriminatory (e.g., small differences between the performance of the same task by two different SMEs), or knowledge that involves changes that may occur simultaneously in several different sensory domains and are interrelated. An SME may not be able to identify the totality of tacit knowledge that he or she possesses. In contrast to tacit knowledge, "explicit" or explicit knowledge is knowledge that a person may easily articulate, codify, store, or access. Whereas an SME may easily transfer explicit knowledge to others, tacit knowledge is not necessarily easily transferable. It would be very valuable to be able to quickly capture tacit knowledge for later use. As described herein, technologies are described that capture both explicit and tacit knowledge to generate training information, thereby accelerating the knowledge acquisition process and enabling the ability to create interactive multimedia instruction manuals, as well as providing augmented reality content in different languages by applying artificial intelligence (AI) and / or machine learning (ML).
[0014] Traditionally, SMEs create documents or videos that may be used to train others, e.g., novices, as a way to transfer knowledge. Creating content in this manner is very time-consuming. Moreover, trainees may have difficulty finding relevant information within a large amount of training content. The technology described herein enables a system to capture explicit knowledge as well as tacit knowledge from SMEs. In an example of the technology described herein, the system includes one or more 3D video cameras for capturing the SME's activities (e.g., a sequence of maintenance tasks, machining activities, or fabrication activities, etc.) in both first-person and third-person perspective views. Such a system further includes one or more microphones configured to capture the SME's dictation as he or she performs the activities. A computational engine, e.g., using AI and / or ML, then performs a series of steps to convert the captured data into training information. In some examples, the computational engine uses AI to apply design patterns to the data to generalize the activities (also referred to herein as "tasks") into one or more steps for performing the activities. Additionally, the computing engine may use AI and / or ML to guide the SME through a well-defined process, such that the AI and / or ML support the active learning process. In some examples, the system generates a series of questions that may be posed to the SME to elicit training-oriented information. For example, the system may query the SME to state the purpose of the presentation, provide a brief description of the activity, and / or explain the process and / or outline the end goal of the process before demonstrating it. Additionally, the system may query the SME for an explanation of any deviations from normal procedures and the reasons for such deviations.
[0015] In some examples, the computing engine of the systems described herein may apply AI to understand information in video from multiple cameras at different angles and build a 3D model from such information. In some examples, the system may perform activity recognition on the manual tasks of the SME. While conventional systems may focus on global human motions such as sitting, standing, walking, etc., the systems described herein may capture finer hand motions such as finger movements, wrist rotation, hand / finger pressure applied to the tool, and / or fine tool motions. To this end, such systems described herein improve accuracy by using multiple 3D cameras to capture a close-up view of the SME while performing the activity or task. In some examples, the system further comprises one or more sensors that capture motion data from the SME, e.g., sensors worn or embedded in wearable clothing. The system uses the sensors in conjunction with captured audio and video data to help capture gestures and movements of the SME that may be shown to a trainee. For example, if the task involves maintaining a particular machine, the system may depict very fine hand manipulations such as cleaning an object, loosening a bolt, removing a disk, checking flatness using a dial gauge, etc. In some examples, the system captures the performance of such tasks from the SME's perspective and replays the trainee's performance of the tasks in the form of augmented reality content, hi other examples, the computational engine uses ML with or without AI to decipher data and capture information from video and audio sources.
[0016] In some examples, the computational engine applies AI and / or ML components to integrate visual entities, actions, and concepts captured via video data, audio data, or sensor data, as well as narration by the SME. Systems as described herein may use minimal training examples to extract information from audio data, e.g., dictations by the SME related to the task. Systems as described herein may also use text data obtained from domain documentation related to the task, if available, or apply rule-based techniques in the absence of such documentation of the task. In some examples, systems as described herein fuse audio data, such as narrative descriptions or dictations by the SME, with analysis of video data of the SME performing the task in various sequences. Additionally, the SME may provide such narration of action-related information at any time (e.g., before, during, or after performing the task). When the SME incorporates dictations into the video recording, the systems described herein may correlate objects depicted in the video data to references to such objects described via the audio recording. For example, the system may identify from the audio data a statement by the SME that for a particular procedure, the SME ensures that the flatness measurement of the object is within 4 microns. The audio data may include this statement before the SME performs the inspection operation, while the SME performs the inspection operation, or after the SME performs the inspection operation. A system as described herein may identify correlations between the statements by the SME and objects recognized in the video data and reconcile multiple sources of information, regardless of the period during which the relevant statements occur in the audio data or the period during which the relevant objects are identified in the video data.
[0017] In another example, the system may identify a correlation between an action taken by the first user with respect to one or more objects depicted in the video data and one or more measurements obtained via one or more sensors. For example, the system may correlate actions depicted in the video data with accelerometer, pressure, or force measurements sensed from a tool used by the first user and generated contemporaneously with a period during which one or more objects are identified in the video data.
[0018] In some examples, the computing engine may apply AI and / or ML to perform other valuable tasks. For example, a system as described herein may use AI to convert a task learning step into a generalized step or to identify and flag missing or unclear segments in the task description. For example, a system as described herein may identify missing information based on a given domain model for the task. As an example, the system may start with a task aimed at calibrating the use of the device and inspecting the range of one or more readouts of the device. The system classifies the dictation from the SME into various categories such as the goal of the task, safety, and additional explanation of the SME's current actions.
[0019] Systems as described herein may perform object recognition to recognize objects in video data with minimal training examples. Systems as described herein may accomplish this by identifying objects in a first video and leveraging this knowledge across multiple subsequent videos. In some examples, systems as described herein may use recognition of multiple operations performed near one or more objects to refine the identification of one or more objects. Additionally, systems as described herein may update a domain model of the task and apply the domain model to generate training information for use by a second user to perform the task. In some examples, the system may use audio or text data obtained from the SME in a first language (e.g., Japanese) to generate training information in a second language (e.g., English) for use in training the second user to perform the task. Thus, systems as described herein may leverage existing documentation created by the company or available from external sources to arrive at useful terminology for training users who do not share a language with the SME. In some examples, the techniques disclosed herein may speed up knowledge acquisition by more than three times over conventional knowledge acquisition techniques.
[0020] 1 is a block diagram illustrating an example system 100 for generating training information in accordance with the techniques of this disclosure. System 100 includes one or more video devices 106, one or more audio devices 108, one or more sensors 120, a machine learning system 112, a domain model 114, and a knowledge database 116.
[0021] According to the techniques of the present disclosure, a first user 102 performs a task (e.g., an activity). In some examples, the first user 102 is an SME performing a task. The task may be, for example, performing maintenance or cleaning of an industrial machine, using an industrial machine to machine or manufacture a machine part, a consumer good, or an art piece, a task involving playing a musical instrument, or any other task requiring practice, training, or expertise not explicitly described herein. The video device 106 generates video data 107 of the first user 102 performing the task. The audio device 108 generates audio data 109 describing the performance of the task. Additionally, the sensor 120 generates sensor data 121 of the performance of the task. The machine learning system 112 of the computing system 130 receives the video data 107, the audio data 109, and the sensor data 121 as well as the domain documents 104 and processes the data obtained from the data 104, 107, 109, and 121 to update the domain model 114, which defines the performance of the task. A computational engine 130 applies the domain model 116 to generate training information 117 for performing the task and stores the training information 117 in the knowledge database 116. A second user 118 may access the training information 117 stored in the knowledge database 116 to train the second user 118 in performing the task.
[0022] As depicted in FIG. 1 , a video device 106 generates video data 107 of a first user 102 performing a task and provides such video data 107 to a machine learning system 112. The video data 107 may include multiple camera sources. For example, the video device 106 includes a first video device and a second video device. The first video device is configured to capture video data of the first user 102 performing a task from a first-person perspective of the first user 102. The second video device is configured to capture video data of the first user 102 performing a task from a third-person perspective of the first user 102. In other examples, the video device 106 may include multiple video devices positioned at a first-person perspective of the first user 102 as well as multiple video devices positioned at various third-person perspectives of the first user 102 (e.g., at various different positions or poses in a room or environment in which the first user 102 performs the task). In some examples, the video data 107 includes close-up 3D variations in the surface, machine vision of the color of the workpiece. The use of multiple cameras from multiple viewpoints allows for the creation of video data 107 that is more comprehensive and more useful to the machine learning system 112 in understanding the first user 102's interaction with the environment, tool, or workpiece, as described in more detail below.
[0023] Each video device 106 is an example of an imaging device that creates multiple two-dimensional (2D) frames from poses of the video device 106. In some examples, the video device 106 may be another type of imaging device that generates 2D or 3D images of a scene, such as a video camera that generates a stream of image data, a laser scanner or other optical device, a depth sensor that generates image data indicating a range of features in the environment, a stereo vision system having multiple cameras to generate 3D information, a Doppler radar, or other imaging device. In some examples, the video device 106 comprises a three-dimensional (3D) camera. Such a 3D camera may record 3D video using two or more imaging devices positioned at different angles to obtain video data in multiple dimensions from multiple poses. The frames generated by the video device 106 may represent 2D images generated periodically on demand, such as frames of a video stream. Such 2D frames may be of various resolutions and may be generated in various formats that can be processed by various units of the system 100.
[0024] The audio device 108 generates audio data 109 describing the performance of the task and provides such audio data 109 to the machine learning system 112. In some examples, the audio data 109 includes speech by the first user 102 describing the first user's actions while performing the task. In other examples, the audio data 109 includes speech by the first user 102 describing how the first user performs the task while not performing the task, such as during an initial interview with the first user 102 before performing the task, or after an interview with the first user 102 after performing the task. An example of an audio device 108 includes a microphone, such as a dynamic microphone, a condenser microphone, or a contact microphone. However, the techniques of this disclosure may use other devices to capture or record audio during the performance of a task not explicitly described herein.
[0025] The sensor 120 generates sensor data 121 of the performance of the task. The sensor 120 may include, for example, one or more motion sensors, pressure sensors, force sensors, or acceleration sensors. In some examples, the sensor 120 generates sensor data obtained from the first user 102, the workspace of the first user 102, or one or more objects with which the first user 102 interacts, such as one or more tools or workpieces, during the performance of the task. In some examples, the sensor data 121 includes data related to at least one of micro-movements or actions of the first user 102. In some examples, the sensor data 121 includes data related to one or more objects with which the first user 102 interacts during the performance of the task. In some examples, the sensor data 121 includes data related to one or more finger or hand movements of the first user, rotation of the first user's wrist, or hand pressure or finger pressure of the first user applied to one or more objects. In some examples, the sensor data 121 includes data related to one or more of an angle between one or more tools and one or more objects, a pressure applied to one or more objects, surface characteristics of one or more objects, or an acceleration of one or more objects.
[0026] For example, the sensor 120 may be incorporated into an article worn by the first user 102 or worn by the first user 102, such as a motion tracking glove that detects the motion and / or force of the user's fingers, hand and / or arm. In some examples, the sensor 120 is incorporated into one or more tools used by the first user 102, such as a smart tool that incorporates one or more pressure sensors to detect the motion and force of the tool being used by the user. In some examples, the sensor 102 is an external sensor that senses data related to the first user 102, the workspace of the first user 102, or an object with which the first user 102 interacts, such as a force pad that detects the force the user applies to a surface, or an inertial measurement unit (IMU) that detects the acceleration of a work surface, workpiece, tool, or the first user 102. In some examples, the sensor 102 includes a wearable glove equipped with an accelerometer that allows for the creation of a 3D model of the body or hand movements of the first user 102.
[0027] The domain documents 104 include text data that describes the performance of a task. Examples of domain documents 104 include instruction manuals for performing a task, a parts list of parts needed to perform a task, a tools list of tools needed to perform a task, a fault report, a machine guide, or other written guides. The first user 102 may provide the domain documents 104 to the computational engine 130. The computational engine 130 performs text recognition to extract the text data that describes the performance of the task and provides such text data to the machine learning system 112.
[0028] The domain model 114 provides a model of the task to be performed by the first user 102. Typically, the domain model 114 integrates knowledge from the first user 102, which may be one or more SMEs. In some examples, the domain model 114 is initially generated during an interview process with the first user 102 configured to rapidly form a baseline model of expert decision-making during the performance of the task and / or procedural skills in the machine, technology, and craft domains related to the performance of the task. In some examples, the interview process may take the form of a first-person narrative by the first user 102 describing a step-by-step process for performing the task, problems the first user 102 may encounter, solutions to such problems, techniques the first user 102 pursues during the performance of the task, or details that the first user 102 believes are important to pay attention to. Typically, the interview is configured to elicit both explicit and tacit knowledge. For example, the interview may take the form of a narrative describing the performance of a given task by the SME to elicit explicit knowledge followed by a series of follow-up questions to elicit tacit knowledge. Tacit knowledge questions are typically exploratory. For example, exploratory questions designed to elicit tacit knowledge might ask, "What will you test before starting a step?" or "How will you measure progress?" Additional follow-up questions may be conducted based on answers provided by the first user 102.
[0029] In some examples, the computation engine 130 may initially generate the domain model 114 from documents related to the task. In some examples, the computation engine 130 may apply a rule-based approach to generate or update the domain model 114. For example, the domain model 114 may identify a rule that a calibration of a tool must be performed before a measurement can be performed using the tool. The computation engine 130 may use such rules to identify or resolve steps in performing a task or to identify missing steps in performing a task modeled by the domain model 114. In some examples, a first user 102 (e.g., one or more SMEs) provides such rules that are codified in the domain model 114.
[0030] Machine learning systems may be used to process images to generate various data about the images. For example, a machine learning system may process an image to identify one or more objects in the image. Some machine learning systems may apply a model generated by a neural network, such as a convolutional neural network, to process the image. Machine learning systems may require large amounts of "training data" to build accurate models. However, once trained, machine learning systems may be capable of performing a variety of image recognition tasks previously thought to be only capable of being performed by humans. For example, machine learning systems may be used in a variety of applications, such as security, commercial applications, scientific and zoological research, and industrial applications such as inventory management and quality control.
[0031] The computation engine 130 applies the machine learning system 112 to the collected data (e.g., the domain documents 104, the video data 107, the audio data 109, and / or the sensor data 121) to update or refine the domain model 114 that governs the execution of the task. In some examples, the domain documents 104, the video data 107, the audio data 109, and / or the sensor data 121 are converted into vectors and tensors (e.g., multi-dimensional arrays). The machine learning system 112 may apply mathematical operations to the vectors and tensors, such as linear algebraic operations, non-linear operations, or other computational operations. In some examples, the machine learning system 102 applies techniques from the field of deep learning. In some examples, the machine learning system 102 is an example of a supervised learning system, an unsupervised learning system, a semi-supervised learning system, or a reinforcement learning system.
[0032] The machine learning system 112 may be initialized by training the machine learning system 112 with training sample data (not depicted in FIG. 1 ) including video data, text data, audio data, and / or sensor data. In some examples, the machine learning system 112 uses such training sample data to teach a machine learning model to identify elements depicted in the video data, text data, audio data, and / or sensor data and train the machine learning system 112 to determine whether such elements are more or less likely to be related to one another, such as by assigning different weights to various elements, applying different coefficients to such elements, etc.
[0033] In some examples, the machine learning system 112 updates the domain model 114 by correlating objects recognized in the video data 107 with references to objects in the audio data 108, text data obtained from the domain documents 104, and / or sensor data 121 to identify portions of the video data 107, audio data 109, text data, and / or sensor data 121 that describe the same step of multiple steps for performing a task. In some examples, the machine learning system 112 performs task learning to generalize the task performed by the first user 102 into one or more steps. In some examples, the machine learning system 112 may apply one or more templates to generalize the task by the first user 102 into one or more steps.
[0034] The computation engine 130 applies the domain model 116 to generate training information 117 for performing the task and stores the training information 117 in the knowledge database 116. The second user 118 may access the training information 117 stored in the knowledge database 116 to train the second user 118 in performing the task. The training information 117 may include, for example, video data, audio data, text data, and portions of sensor data of the performance of the task cross-referenced such that the second user may search the knowledge database 116 with a concept, task, or sub-task to obtain training information related to the second user's query. For example, the training information 117 may include objects recognized in the video data 107, portions of the audio data 109 describing the objects recognized in the video data 107, portions of the domain documents 104 related to the objects recognized in the video data 107, and / or sensor data acquired from the sensors 120 during the performance of a step of a task related to the objects recognized in the video data 107. In some examples, the training data 117 is personalized for the second user 118. In some examples, the knowledge database 116 outputs the training information 117 to the second user 118 for use in training the second user 118 to perform a task. In some examples, the knowledge database 116 outputs the training information 117 in the form of an augmented reality video depicting the performance of a task from a first-person perspective of the second user 118. In other examples, the knowledge database 116 outputs the training information 117 in the form of an interactive technical manual for training the second user 118 to perform a task.
[0035] In some examples, the knowledge database 116 receives a query from the second user 118 for instructions to perform a task or one of a plurality of steps to perform a task. In response to the query, the knowledge database 116 applies the domain model 114 to generate training information 117 for performing the task or one of a plurality of steps to perform the task and outputs such training information 117 to the second user 118. For example, given a procedure modeled for the task and a current state of the domain model 114 derived from data observed from the environment of the system 100, the computational engine 130 may use the domain model 114 to predict or identify a next step to be performed by the second user 118 and suggest performing such a next step. Thus, the computational engine 130 may use the domain model 114 to identify a next step that the SME would perform in performing the task and suggest such a step, for example, to train the second user 118 to perform the task. Additionally, the computational engine 130 may use the domain model 114 to evaluate the performance of a novice user, for example, by comparing steps performed by the novice user with steps performed by an SME.
[0036] 2 is a block diagram illustrating an example computing system 200 for generating training information according to the techniques of this disclosure. In the example of FIG. 2, computing system 200 includes a computing engine 130, one or more input devices 252, and one or more output devices 254. In some examples, computing system 200 includes one or more computing devices interconnected with each other, such as one or more mobile phones, tablet computers, laptop computers, desktop computers, servers, Internet of Things (IoT) devices, etc. In some examples, computing system 200 is a single computing device. In some examples, computing system 200 is distributed across multiple computing devices and interconnected by a computer network (e.g., implemented as a cloud-based application).
[0037] In the example of FIG. 2, the computing system 200 may provide user input to the computing engine 130 via one or more input devices 252. A user of the computing system 200 may provide input to the computing system 200 via one or more input devices 252. The input devices may include a keyboard, a mouse, a microphone, a touch screen, a touch pad, or another input device coupled to the computing system 120 via one or more hardware user interfaces. Additionally, the computing system 200 may receive data via the input devices 252 from various other sources, such as domain documents 104 of FIG. 1, video data 107 via one or more video devices 106 of FIG. 1, audio data 109 via one or more audio devices 108 of FIG. 1, or sensor data 121 via one or more sensors 120 of FIG. 1, which are processed by various components of the computing engine 130, as described in more detail below.
[0038] The input device 252 may include hardware and / or software for establishing a connection with the processing engine 130. In some examples, the input device 252 may communicate with the processing engine 130 via a direct wired connection, a network such as the Internet, or any public or private communication network, e.g., broadband, cellular, Wi-Fi, and / or other type of communication network capable of transmitting data between computing systems, servers, and computing devices. The input device 252 may be configured to transmit and receive data, control signals, commands, and / or other information via such a connection using any suitable communication technology for receiving sensor data. In some examples, the input device 252 and the processing engine 130 may each be operably coupled to the same network using one or more network links. The link coupling the input device 252 and the processing engine 130 may be a wireless wide area network link, a wireless local area network link, Ethernet, asynchronous transfer mode (ATM), or other type of network connection, and such connection may be wireless and / or wired.
[0039] Output device(s) 254 may include a display, an audio source card, a video graphics adapter card, a speaker, a presence screen, one or more USB interfaces, a video and / or audio output interface, or any other type of device capable of generating tactile, audio, video, or other output. Output device(s) 254 may include a display device that may function as an output device using technologies including liquid crystal display (LCD), quantum dot display, dot matrix display, light emitting diode (LED) display, organic light emitting diode (OLED) display, cathode ray tube (CRT) display, electronic ink, or monochrome, color, or other type of display capable of generating tactile, audio, and / or visual output. In other examples, output device(s) 254 may generate output to a user in another manner, such as via an audio source card, a video graphics adapter card, a speaker, a presence screen, one or more BS interfaces, a video and / or audio output interface, or any other type of device capable of generating tactile, audio, video, or other output. In some examples, the output device(s) 254 may include a presence-aware display that may function as a user interface device operating both as one or more input devices and as one or more output devices. In some examples, the output device(s) include one or more interfaces for transmitting data to another computing device via a wired or wireless connection.
[0040] The computational engine 130 comprises a machine learning system 112, a domain model 114, a text recognition unit 202, a speech recognition unit 212, a video recognition unit 214, and a training unit 210. Each of the components 112, 114, 202, 210, 212, and 214 may operate in a manner substantially similar to the analogous components of FIG. 1. The computational engine 130 may represent software executable by a processing circuit 256 and stored in a storage device 258, or a combination of hardware and software. Such processing circuit 256 may comprise any one or more of a microprocessor, a controller, a digital signal processor (DSP), an application specific integrated circuit (ASIC), a field programmable gate array (FPGA), or equivalent discrete or integrated logic circuitry. The storage devices 258 may include memory such as random access memory (RAM), read only memory (ROM), programmable read only memory (PROM), erasable programmable read only memory (EPROM), electronically erasable programmable read only memory (EEPROM), flash memory, etc. that contain executable instructions for causing one or more processors to perform actions resulting from the instructions.
[0041] The text recognition unit 202 receives the domain documents 104 and performs text recognition to obtain text data suitable for use by the machine learning system 112. Examples of domain documents 104 include instruction manuals for performing a task, a parts list of parts needed to perform the task, a tools list of tools needed to perform the task, instruction manuals for each of the tools used in performing the task, schematics or specifications for a finished product, written narratives or other written guides provided by an SME. In some examples, the text recognition unit 202 may obtain the domain documents 104 from a first user 102 (e.g., an SME), a repository of technical documentation (e.g., maintained by a company or an external source such as the Internet), other users, etc. In some examples, the text recognition unit 202 implements a machine learning system for performing text recognition on the domain documents 104. In some examples, the text recognition unit 202 performs text recognition on the domain documents 104 using off-the-shelf text recognition software.
[0042] The video recognition unit 214 receives the video data 107 from the video device 106 of FIG. 1 and performs object recognition to identify one or more objects depicted in the video data 107. In some examples, the video recognition unit 214 identifies the first user 102, one or more tools used by the first user 102, one or more workpieces with which the first user 102 interacts, etc. In some examples, the video recognition unit 214 processes the video data 107 to generate video data labeled with human pose, object, or activity sequence annotations. For example, the video recognition unit 214 annotates the video data 107 with human pose data of the first user 102, including skeleton pose data, joint recognition, hand gesture recognition, etc. In some examples, the video recognition unit 214 annotates the video data 107 with object detection data, such as annotations of human skeletons or joints detected in frames of the video data 107, or annotations of objects, such as tools, workpieces, etc., recognized in frames of the video data 107. In some examples, the video recognition unit 214 determines a confidence score or probability that a frame of the video data 107 depicts a particular object. For example, the video recognition unit 214 may annotate each frame in the video data 107 with one or more objects that are recognized as depicted in the frame. Using multiple cameras from multiple viewpoints allows for the creation of video data 107 that is more comprehensive and more useful to the machine learning system 112 in understanding the first user 102's interaction with the environment, tool, or workpiece. For example, video data from multiple angles may aid object recognition by reducing instances of object occlusion and provide a closer view of the movement by the SME, tool, or workpiece from multiple viewpoints. For example, the video recognition unit 214 may use objects recognized in the video data from the first video device 106 to aid in the identification of objects, such as partially occluded objects, present in the second video device 106. Additionally, the video recognition unit 214 may apply AI to facilitate object identification when a recognized sequence of actions is performed in close proximity of two objects.The video recognition unit 214 uses such video data 107 from multiple viewpoints to build a 3D model of the environment including the first user 102, tools and workpieces, as well as to model interactions between the first user 102 and such tools and workpieces. In some examples, the video recognition unit 214 uses a machine learning system to perform human pose detection and / or object recognition on the video data 107. In some examples, the video recognition unit 214 uses off-the-shelf object recognition software to perform human pose detection and / or object recognition on the video data 107.
[0043] The speech recognition unit 212 receives the audio data 109 from the audio device 108 of FIG. 1 and performs speech recognition to identify references to one or more objects or concepts present in the audio data 109. In some examples, the audio data 109 includes a narrative by the first user 102 describing the first user's actions while performing a task. In other examples, the audio data 109 includes a narrative by the first user 102 describing how to perform a task while the first user is not performing the task, such as during an interview with the first user 102 before or after performing the task. In some examples, the speech recognition unit 212 performs speech recognition on the audio data 109 using machine learning. In some examples, the speech recognition unit 212 performs speech recognition on the audio data 109 using off-the-shelf speech recognition software.
[0044] The machine learning system 112 correlates the video data 107, the audio data 109, the sensor data 121, and the text data obtained from the domain documents 104 to identify at least a portion of the text data obtained from the video data 107, the audio data 109, the sensor data 121, and the domain documents 104 that describes the same step of a plurality of steps for performing a task. Further, the machine learning system 112 processes the correlated portions of the text data obtained from the video data 107, the audio data 109, the sensor data 121, and the domain documents 104 to update the domain model 114. As described herein, a task may be conceptualized as a plurality of steps performed to achieve a given goal. Thus, the domain model 114 specifies the execution of the task by specifying a plurality of steps or operations that the first user 102 performs to achieve the goal of the task. In some examples, the domain model 114 specifies the execution of the task by specifying at least one of an ontology, a set of related concepts or objects, entities, actions, events, or rules (e.g., semantic rules) associated with the execution of the task. In some examples, the domain model 114 models the execution of a task by defining at least one of an ontology, a set of related concepts or objects, entities, actions, events, or rules (e.g., semantic rules) related to the execution of a task. An ontology is a semantic relationship between various objects, and may be created by the machine learning system 112 or manually by an SME such as the first user 102. The machine learning system 112 may perform clustering to detect a set of related terms, concepts, or objects by identifying relationships present in the video data 107, the audio data 109, and the text data obtained from the domain documents 104. The machine learning system 112 may identify entities by using object detection performed on the video data 107 and text semantic extraction from the text data obtained from the domain documents 104.The machine learning system 112 may identify events (e.g., actions being performed) by performing action sequence recognition on the video data 107 and performing text semantic extraction from the text data obtained from the domain documents 104. Semantic rules are rules for fusing the video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104 (e.g., processing the video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104 to identify relationships between entities, objects, and actions present in the video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104). In some examples, the semantic rules are handcrafted by an SME, such as the first user 102.
[0045] As an example, the machine learning system 112 may correlate portions of the video data to portions of the audio data by identifying objects depicted in the video data 107, identifying references to the objects from the audio data 109, and correlating the objects depicted in the video data 107 to the references to the objects in the audio data 109. The machine learning system 112 may then define a domain model 114 based on the correlations between the objects identified from the video data 107 and the references to the objects identified from the audio data 109, for example, by using the correlations to define rules of the domain model 114 that govern an ontology, entity, action, event, or task performance.
[0046] In some examples, the machine learning system 112 applies co-clustering to extract domain-related semantic information, such as task names, objects, tools, etc., from each of the video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104 to build or extend an ontology of concepts in the domain model 114. For example, the machine learning system 112 may apply unsupervised machine learning, such as a co-clustering algorithm, to discover a set of related concepts and terms.
[0047] In some examples, the machine learning system 112 implements a first machine learning system that correlates the video data 107, the audio data 109, the sensor data 121, and the text data obtained from the domain document 104 to identify at least a portion of the video data 107, the audio data 109, the sensor data 121, and the text data obtained from the domain document 104 that describe the same step of a plurality of steps for performing a task. For example, the first machine learning system identifies one or more objects used in performing the task from at least a portion of the video data 107. The first machine learning system identifies references to one or more objects used in performing the task from at least a portion of the audio data 109. The first machine learning system identifies one or more physical measurements of one or more objects used in performing the task from at least a portion of the sensor data 121. The first machine learning system correlates the one or more objects identified from the video data 107 with the references to one or more objects identified from the audio data 109 and the physical measurements of the one or more objects identified from the sensor data 121.
[0048] As another example, the machine learning system 112 implements a second machine learning system that processes correlated video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104 to update the domain model 114 that governs the execution of a task. For example, the second machine learning system prescribes ontology, entities, actions, events, or rules that govern the execution of a task described by the domain model 114 based on correlations between the video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104.
[0049] The training unit 210 applies the domain model 114 to generate training information 117 or for use in training another user (e.g., second user 118 of FIG. 1 ) to perform a task. The training information 117 may include, for example, video data, audio data, sensor data, and / or textual data related to, for example, the task, one or more of the steps that make up the task, or an object (e.g., a tool or workpiece) related to the task, where each type of data is cross-referenced with each of the other types of data, and where the second user searches the knowledge database by concept, task, or sub-task to obtain audio, video, sensor, or textual information related to the second user's query.
[0050] In some examples, the training unit 210 outputs the training information 117 via output device 254 to a knowledge database 116 of Figure 1 to create a repository of training information for use by one or more users. In some examples, the training unit 210 outputs the training information 117 via output device 254 to, for example, a second user 118 of Figure 1 to train the second user 118 to perform a task.
[0051] The personalization unit 204 converts the training information acquired from the first user 102 of FIG. 1 using the above techniques into personalized training information for the second user 117. As an example, the training unit 210 acquires the training information in the first language of the first user 102 of FIG. 1. The personalization unit 204 generates the training information 117 in the second language of the second user 118 from the training information in the first language of the first user 102. This may be more useful for training the second user 118 to perform a task. In some examples, the personalization unit 204 may leverage the domain documents 104 created by the company in different languages to determine the terms that may be applied to the elements of the domain model 114 in defining the task. In some examples, the personalization unit 204 may use an off-the-shelf translation tool to help translate the training information 117 from the first language to the second language.
[0052] The training manual generation unit 206 converts the training information obtained from the first user 102 of FIG. 1 using the techniques described above into a training manual for use in training the second user 118 to perform a task. In some examples, such a training manual may be a text document. In some examples, such a training manual may be an interactive multimedia manual in the form of text, audio and / or video information. The second user 118 may interact with such a manual to experience a more comprehensive or effective training experience.
[0053] The augmented reality content unit 208 converts the training information obtained from the first user 102 of FIG. 1 using the techniques described above into augmented reality content for use in training the second user 118 to perform the task. The training unit 210 may output such augmented reality content, for example, to a head mounted display (HMD) worn by the second user 118 to provide an experiential first person perspective of the SME's performance of the task. In some examples, the augmented reality content may include relevant portions of the audio data 109, such as a narration by the first user 102, and relevant portions of the video data 107, such as the perspective of the first user 102 as it performs the task. In some examples, the second user 118 may interact with a controller that provides force feedback such that the augmented reality content may provide force feedback based on sensor data 121 that replicates the performance of the task by the first user 102.
[0054] 3 is a flow chart illustrating example operations for generating training information in accordance with the techniques of this disclosure. For convenience, FIG. 3 is depicted with respect to FIG.
[0055] As depicted in the operations of FIG. 3 , the system 100 performs 304 active knowledge acquisition of the first user 102 performing the task. The active knowledge acquisition of the first user 102 allows the system 100 to acquire the explicit knowledge of the SME related to the performance of the task, and may also allow the system 100 to acquire some tacit knowledge of the SME related to the performance of the task. In some examples, the system 100 performs active knowledge acquisition in the form of a well-structured interview. In some examples, during active knowledge acquisition, the computational engine 130 generates a series of questions to the first user 102 to elicit training-oriented information. In some examples, the series of questions includes the purpose of the task, a brief description of the task, the end goal of the sequence of work, and a demonstration of the execution of the task. In some examples, if the first user 102 deviates from the normal procedure, the computational engine 130 prompts the first user 102 to provide an explanation, such as the deviation and the purpose of the deviation.
[0056] For example, the video device 106 captures video data 107 of the first user 102 performing the task. The video data 107 may include multiple camera sources. For example, the video device 106 includes a first video device and a second video device. The first video device is configured to capture the video data of the first user 102 performing the task from a first person perspective of the first user 102. The second video device is configured to capture the video data of the first user 102 performing the task from a third person perspective of the first user 102.
[0057] Additionally, the audio device 108 captures audio data 109 of the first user 102 performing the task. During active knowledge acquisition, the audio data 109 includes narration by the first user 102 describing the first user's actions while actively performing the task.
[0058] Additionally, the sensor device 120 acquires sensor data 121 of the first user 102 performing the task. The sensor 120 may include, for example, one or more motion sensors, pressure sensors, force sensors, or acceleration sensors. In some examples, the sensor 120 is incorporated into an article worn by or worn by the first user 102, such as a motion tracking glove that detects the motion and / or force of the user's fingers, hand, and / or arm. In some examples, the sensor 120 is incorporated into one or more tools used by the first user 102, such as a smart tool that incorporates one or more pressure sensors to detect the motion and force of the tool being used by the user. In some examples, the sensor 102 is an external sensor that senses data related to the first user 102, the workspace of the first user 102, or an object with which the first user 102 interacts, such as a force pad that detects the force applied by the user to the surface, such as a work surface, workpiece, tool, or an IMD that detects the acceleration of the first user 102.
[0059] The system 100 further performs 304 passive knowledge acquisition of the first user 102 performing the task. Passive knowledge acquisition of the first user 102 enables the system 100 to acquire both explicit and tacit knowledge of the SME related to performing the task. For example, during passive knowledge acquisition, the audio device 108 acquires audio data 109 including a narrative by the first user 102 describing how to perform the task while the first user is not performing the task, such as during an interview with the first user 102 before or after performing the task.
[0060] Additionally, the text recognition unit 202 of the computational engine 130 receives text data in the form of domain documents 104 related to the performance of the task. Examples of domain documents 104 include instruction manuals for performing the task, parts lists of parts needed to perform the task, tools lists of tools needed to perform the task, or other written guides. The text recognition unit 202 performs text recognition to extract company-specific terms from the domain documents 104 (302). In some examples, the text recognition unit 202 performs text recognition to obtain text data suitable for use by the machine learning system 112. The computational engine 130 may train the machine learning system 112 with such text data, thereby training the machine learning system 112 to accurately identify tasks and reducing the number of training examples required to refine a useful domain model and / or training information for the second user 118. Thus, the machine learning system 112 using the techniques of the present disclosure may require only a minimal number of training sample data to provide useful output (e.g., training information 117 for training the second user 118 to perform the task). In some examples, the machine learning system 112 can acquire knowledge from three examples of SMEs performing the task and compare the knowledge acquired from the SMEs with examples of novices performing the task, and then generate training information 117 for training a second user 118 to perform the task.
[0061] In instances where domain documents 104 are not available for a particular task, the machine learning system 112 may instead apply rule-based techniques to process the multimodal data and update the domain model 114 so that the domain model 114 more accurately describes the performance of the task. Alternatively, or in addition, the machine learning system 112 may apply one or more templates to process the multimodal data and update or refine the domain model 114.
[0062] The speech recognition unit 212 receives the audio data 109 and extracts 308 first semantic information relevant to the performance of the task from the dictation of the first user 102. For example, the speech recognition unit 212 performs speech recognition to identify references to one or more objects or concepts present in the audio data 109. In some examples, the dictation of the first user 102 includes a description of a machine maintenance activity, such as performing an inspection operation to ensure that a workpiece has a flatness within 4 microns. The first user 102 may provide this narration before, during, or after performing the inspection operation. As described below, regardless of the chronological occurrence of this description of the inspection operation by the first user 102, the machine learning system 112 may correlate this description to the occurrence of this inspection operation extracted from the video data 107. In some examples, the speech recognition unit 212 may perform complex speech recognition to extract explicit and tacit knowledge scattered across multiple sentences.
[0063] The video recognition unit 214 receives the video data 107 and extracts 310 second semantic information from the video data 107 that is related to the performance of the task. For example, the video recognition unit 214 performs object recognition to identify one or more objects depicted in the video data 107. In some examples, the video recognition unit 214 identifies the first user 102, one or more tools used by the first user 102, one or more workpieces with which the first user 102 interacts, etc. In some examples, the video recognition unit 214 processes the video data 107 to generate video data labeled with annotations of human postures, objects, or activity sequences. In some examples, the video recognition unit 214 performs activity recognition on the video data 107. Activity recognition may typically be performed to identify human movements such as sitting, standing, walking, etc. 3, the video recognition unit 214 performs activity recognition to identify machine maintenance actions such as cleaning, loosening bolts, removing disks, very fine manual manipulations such as inspecting flatness using a dial gauge, etc. In some examples, the video recognition unit 214 may perform complex human pose and / or object recognition to extract explicit and tacit knowledge scattered across multiple different video sources and cross-reference such knowledge with knowledge extracted from audio data 109.
[0064] The machine learning system 112 processes the sensor data 121 to extract (309) third semantic information relevant to the performance of the task from the sensor data 121. For example, the machine learning system 112 may identify portions of the sensor data 121 that correspond to micro-movements or movements by the first user 102 or interactions by the first user 102 with one or more objects, such as a tool or a workpiece.
[0065] The machine learning system 112 fuses (312) information from multiple modalities. For example, the machine learning system 112 processes the video data 107, the audio data 109, the sensor data 121, and text data obtained from the domain documents 104 to update the domain model 114. The domain model 114 provides a model of a task performed by the first user 102. In some examples, the domain model 114 models the task as multiple steps performed to achieve a particular goal. In some examples, the domain model 114 models the task by specifying at least one of an ontology, a cluster, an entity, an action, an event, or a rule (e.g., a semantic rule) associated with performing the task. For example, the machine learning system 112 may correlate a portion of the video data 107 to a portion of the audio data 109 by identifying an object depicted in the video data 107, identifying a reference to the object from the audio data 109, and correlating the object depicted in the video data 107 to the reference to the object in the audio data 109. The machine learning system 112 may then define a domain model 114 based on correlations between objects identified from the video data 107 and references to objects identified from the audio data 109, for example by using the correlations to define rules in the domain model 114 that define an ontology, an entity, an action, an event, or the performance of a task.
[0066] As another example, the machine learning system 112 processes the sensor data 121 to synchronize data from each of a plurality of sensors 120 of various types with the video data 107 and / or the audio data 109. For example, the machine learning system 112 may correlate a portion of the video data 107 with a portion of the sensor data 121 by identifying an action of the first user 102 depicted in the video data 107, identifying a portion of the sensor data 121 generated contemporaneously with the action depicted in the video data 107, and correlating the action depicted in the video data 107 with the portion of the sensor data 121 generated contemporaneously with the action depicted in the video data 107. In one example, the machine learning system 112 receives the sensor data 121 from the sensor 120 that describes changes in the surface of the workpiece. Additionally, the machine learning system 112 receives the video data 107 annotated with human pose data of the first user 102. The machine learning system processes the sensor data 121 and the video data 107 to correlate changes in the surface of the workpiece with the human pose data in the video data 107 as well as the body movements of the first user 102 depicted in other machine logs to build a model of the physical movements performed by the first user 102 to perform one or more of a plurality of steps to perform a task.
[0067] As another example, the machine learning system 112 processes the video data 107 and the audio data 109 to correlate the video data 107 with the audio data 109 to identify at least a portion of the video data 107 depicting a step of a plurality of steps for performing a task and at least a portion of the audio data 109 describing the same step of the plurality of steps for performing a task. For example, the machine learning system 112 identifies at least a portion of the video data 107 depicting a step of a plurality of steps for performing a task and at least a portion of the audio data 109 describing the same step of the plurality of steps for performing a task from the first semantic information obtained from the audio data 109 and the second semantic information obtained from the video data 107. As described herein, "semantic information" refers to meaningful information regarding the performance of a task that the system 100 obtains from the environment, such as the audio data 109, the video data 107, the domain documents 104, and the sensor data 121. For example, the machine learning system 112 may identify semantic information such as a sentence type (e.g., action, warning, list of tools, preconditions) as well as an action, object or tool derived from the action sentence from such first and second semantic information obtained from the audio data 109 and the video data 107. The machine learning system 112 may fuse such semantic information from, for example, the video data 107, the audio data 109, the sensor data 121 and the domain documents 104 to create a coherent sequence of semantic steps to perform a task.
[0068] In some examples, the computation engine 130 identifies ambiguities in one or more steps to perform the task and / or one or more semantic differences or inconsistencies in the performance of the task (314). For example, the computation engine 130 may identify a discrepancy between, for example, the video data 107 of the performance of the task by the first user 102 and the audio data 109 including a description of the performance of the task provided by the first user 102. The computation engine 130 may query the first user 102 for an explanation of such a discrepancy or difference. In some examples, the computation engine 130 uses the goal of the task to generate a query to the first user 102 for an explanation. For example, the computation engine 130 may recognize that the use of the device requires calibration initially and that the reading of the device always has a range to check. Thus, if the first user 102 relies on the reading of the device without calibrating the device, the computation engine 130 may generate a query to the first user 102 to explain why. The computational processing engine 130 may use the description of the identified semantic differences between the video data 107 and audio data 109 received from the first user 102 to update the domain model 114, thereby increasing the accuracy of the domain model 114 during task execution and reducing any ambiguity in the actions performed by the first user 102.
[0069] In some examples, the system 100 may acquire knowledge to perform a task from a first SME and a second SME by obtaining a first video data of a first SME performing a task, a first audio data of the first SME narrating the performance of the task, a second video data of a second SME performing the task, and a second audio data of the second SME narrating the performance of the task. The machine learning system 112 correlates the first video data of the first SME performing the task with the first audio data of the first SME narrating the performance of the task. The machine learning system 112 further correlates the second video data of the second SME performing the task with the second audio data of the second SME narrating the performance of the task. The machine learning system 112 further processes the correlated video data and audio data of each SME performing a task to identify semantic differences or inconsistencies between the performance of the task by the first SME and the performance of the task by the second SME. The computation engine 130 may query the first SME and the second SME for an explanation of the semantic difference, and the computation engine 130 may update the domain model 114 with the resulting explanation of the semantic difference.
[0070] In some examples, the system 100 may acquire knowledge to perform a task from a first user and a second user (e.g., an SME and a novice user, respectively) by obtaining first video data of an SME performing a task, first audio data of the SME narrating the performance of the task, second video data of a novice user performing the task, and second audio data of the novice user narrating the performance of the task. The computational engine 130 may identify semantic differences or inconsistencies between the performance of the task by the SME and the second user in a manner similar to the above examples. Additionally, the training unit 210 may generate feedback describing the identified semantic differences between the performance of the task by the SME and the performance of the task by the second user, and the output device 254 may output for display to the second user to guide the second user in performing the task.
[0071] In some examples, the system 100 may acquire knowledge to perform a task from a first user who is an SME and a second user who is a novice by obtaining a first video data of an SME performing a task, a first audio data of an SME narrating the performance of the task, a second video data of a novice performing the task, and a second audio data of the novice narrating the performance of the task. The computational engine 130 may identify inconsistencies or differences between the performance of the task by the SME and the performance of the task by the novice. Further, the computational engine 130 may generate queries to the SME for explanations of inconsistencies or differences between actions performed by the SME and actions performed by the novice, as well as receive annotations from the SME of errors made by the SME or the novice user. In some examples, the computational engine 130 may generate an output indicating differences in the maps of the first SME versus the second SME versus the novice user. For example, the computing engine 130 may generate one or more maps showing the differences between the first user and the second user to visualize one or more differences between the performance of the task by each of the first user and the second user. Such maps may include, for example, the pressure applied by each user, the hand movements of each user, the body movements of each user, etc. Such maps may enable visualization of the differences between the performance of the task by an SME and a novice user, and may be useful in assisting the novice user in learning to perform the task.
[0072] The system 100 may receive a response, for example, from the first user 102 in response to a query to clarify the ambiguity as discussed above. Additionally, the system 100 may receive audio / video information from the SME during a post-task interview discussing the performance of the task. The machine learning system 112 updates (318) the domain model 114 for the task based on the received explanations and / or post-interview. The training unit 210 stores (316) the semantic information of the data model 114 in the knowledge database 116 in the form of training information 117.
[0073] In some examples, after the training information 117 is stored in the knowledge database 116, the system 100 may repeat the above operations of Figure 3 to iteratively perform active and passive knowledge acquisition. In this manner, the system 100 may use the existing knowledge acquisition information to improve the performance of the machine learning system 112 and to increase the accuracy and detail of the training information 117 in the knowledge database 116.
[0074] 4 is a diagram illustrating an example system 400 for generating training information in accordance with the techniques of this disclosure. In some examples, FIG. 4 depicts fusion 414 of information from multiple modalities of FIG. 3 (312).
[0075] As depicted in FIG. 4, the video device 106 includes a camera 402 that captures a wide-angle first-person perspective of the first user 102, a camera 404A that captures a third-person perspective of the first user 102 from a position to the left of the first user 102, and a camera 404B that captures a third-person perspective of the first user 102 from a position to the right of the first user 102.
[0076] The video recognition unit 214 receives the video data acquired from the cameras 402, 404A, and 404B and performs object recognition to identify one or more objects depicted in the video data. For example, the video recognition unit 214 identifies an instruction manual 408 captured by the camera 402 and a working area machine space 410 captured by the camera 404B. Furthermore, the video recognition unit 214 labels the video data from the cameras 404A and 404B with the human pose data 412 of the first user 102.
[0077] The machine learning system 112 fuses information from multiple modalities to update the domain model 114 of FIG. 1. For example, as shown in FIG. 4, the machine learning system 112 processes video data acquired from the cameras 402, 404A, and 404B with, for example, the audio data 109, the sensor data 121, and text data acquired from the domain documents 104 of FIG. 1 to update the domain model 114. For example, the machine learning system 112 correlates portions of the video data to portions of the audio data by identifying objects depicted in the video data, identifying references to the objects from the audio data 109, and correlating the objects depicted in the video data to the references to the objects in the audio data 109. The machine learning system 112 may then define the domain model 114 based on the correlations between the objects identified from the video data and the references to the objects identified from the audio data 109, for example, by using the correlations to define rules of the domain model 114 that define an ontology, an entity, an action, an event, or the performance of a task.
[0078] As an illustrative example, the video recognition unit 214 recognizes objects by detecting human pose data (e.g., skeleton detection and / or joint detection) and performing object detection from the video data acquired from the cameras 402, 404A, and 404B. The video recognition unit 214 performs action sequence recognition to detect actions based on a sequence of recognized human poses. In some examples, the video recognition unit 214 categorizes similar actions that are close in time to each other by ignoring very short time intervals between similar actions. In some examples, the video recognition unit 214 generates a list of recognized actions and generates a confidence that the video recognition unit 214 correctly identified the action. In some examples, the speech recognition unit 212 may normalize the detected object or tool by checking for equivalent names in the population of the knowledge database, for example, using narrative information from the first user 102 or the domain documents 104. For example, an instruction manual may refer to a part as “roller upper” whereas an SME may refer to the same part as “upper roller”, or two different documents may refer to the same tool as “dial gauge” or “dial indicator”. The machine learning system 112 may improve the accuracy of the video extraction results obtained by the video recognition unit 214 by leveraging the ontological knowledge of the domain model 114, thereby removing inaccurate or inconsistent hypotheses. For example, the machine learning system may determine whether a recognized action sequence including actions, objects and tools / equipment is consistent or permitted by the domain model 114. An action recognized by the video recognition unit 214 as, for example, “cleaning with a hammer” is not possible and will be discarded by the machine learning system 112. Furthermore, the machine learning system 112 may fuse similar actions of longer time intervals from the video data 107.
[0079] Continuing with the above example, the machine learning system 112 determines whether one or more references to objects obtained from the audio data 109 correspond to objects recognized in the video data 107. For example, if a portion of the audio data 109 describes an activity, an object, a tool, and a location, and that portion of the audio data 109 is in proximity to a portion of the video data 107, and semantic information extracted from that portion of the video data 107 matches the portion of the audio data 109 (e.g., the activity, object, and tool are recognized in the portion of the video data 107), then the machine learning system 112 forms a correlation between the portion of the audio data 109 and the portion of the video data 107.
[0080] As another example, if a portion of the audio data 109 describes an action, an object, a tool, and a location, and semantic information extracted from a portion of the video data 107 does not match that portion of the audio data 109 (e.g., the action, object, and tool are not recognized in that portion of the video data 107), the machine learning system 112 examines the objects recognized in the video data 107 in order according to the confidence of the object recognition. If the object described in the audio data 109 appears in the video data 107 with a level of certainty that exceeds a predetermined threshold, the machine learning system 112 forms a correlation between said portion of the audio data 109 and said portion of the video data 107.
[0081] As another example, if a portion of the audio data 109 describes actions, objects, tools, and locations, and the semantic information extracted from the video data 107 does not match that portion of the audio data 109 (e.g., the actions, objects, and tools are not recognized in that portion of the video data 107), but the objects appearing in the video data 107 are recognized with a high level of certainty, the machine learning system 112 determines that the objects were correctly recognized in the video data 107, and the machine learning system uses the video data 107 in building the domain model 114.
[0082] As another example, if a portion of the audio data 109 describes actions, objects, tools and locations and the semantic information extracted from the video data 107 does not match that portion of the audio data 109 (e.g., the actions, objects and tools are not recognized in that portion of the video data 107), but the objects appearing in the video data 107 are recognized with a low level of certainty, the machine learning system 112 determines that the objects may not be correctly recognized in the video data 107 and the computational processing engine 130 generates a query to the first user 102 (e.g., an SME) to resolve the ambiguity.
[0083] As another example, if a portion of the audio data 109 describes an action, object, tool, and location, but there is no corresponding portion of the video data 107 for that portion of the audio data 109, the machine learning system 112 may treat that portion of the audio data 109 as containing tacit knowledge. The machine learning system 112 may then use this portion of the audio data 109 in constructing the domain model 114.
[0084] 5 is a diagram illustrating an example system 500 for generating training information in accordance with the techniques of this disclosure. In some examples, the system 500 illustrates an example of creating the knowledge database 116 of FIG 1. For convenience, FIG 5 is depicted with respect to FIG 1 and FIG 2.
[0085] As depicted in the example of FIG. 5, the machine learning system 112 fuses information from multiple modalities to update the domain model 114. Training information 117 may be obtained from the domain model and stored in the knowledge database 116. For example, the multiple modalities may include video data 107. From the video data, the machine learning system 112 acquires knowledge of activities, activity sequences, tools, etc. related to performing a task. As another example, the multiple modalities may include domain documents 104 including instruction manuals, list of tools, etc. From the domain documents, the machine learning system 112 acquires knowledge of the sequence of activities related to the task, expected duration, worker expertise, etc. As another example, the multiple modalities may include audio data (e.g., provided via narration by an SME such as the first user 102). From the audio data, the machine learning system 112 acquires knowledge of the approximate chronological or sequential timing of activities, order of activities, measurements, tools, procedures, etc. As yet another example, the multiple modalities may include sensor data 121 (e.g., provided by one or more accelerometers, pressure sensors, force sensors, and / or motion sensors), from which the machine learning system 112 acquires knowledge of micro-movements, such as hand movements, by the first user 102, the actions of the first user 102, physical forces (e.g., rotational or translational forces) applied to one or more objects, such as a tool or a workpiece, or surface features of one or more objects, such as the texture or roughness of the workpiece. The machine learning system 112 processes the video data 107, the audio data 109, the sensor data 121, and text data obtained from the domain documents 104 to update the domain model 114. The training unit 210 applies the domain model 114 to generate training information 117 for the task. The training unit 210 stores in the form of a knowledge database 116, which may be a fused knowledge database that can be queried by a trainee for training information regarding the execution of a task or the implementation of one or more steps of a plurality of steps to accomplish a task.
[0086] 6 is a diagram of labeled video data 600 for use in generating training information in accordance with the techniques of this disclosure. In some examples, the video recognition unit 214 of FIG. 2 processes video data 107 captured via one or more video devices 106 to generate the labeled video data 600. As depicted in FIG. 6, the labeled video data 600 includes a single video frame labeled with human pose data 602, including, for example, skeletal pose data, joint recognition, hand gesture recognition, etc., of the first user 118 of FIG. 1.
[0087] 7 is a diagram of labeled video data 700 for use in generating training information in accordance with the techniques of this disclosure. In some examples, the video recognition unit 214 of FIG. 2 processes video data 107 captured via one or more video devices 106 to generate the labeled video data 700. As depicted in FIG. 7, the labeled video data 700 includes a single video frame labeled with human pose data 702, including, for example, skeletal pose data, joint recognition, hand gesture recognition, etc., of the first user 118 of FIG. 1.
[0088] 8 is a diagram of an example user interface 800 depicting training information 117 generated in accordance with the techniques of this disclosure. The training information 117 may be, for example, training information 117 generated by training unit 210 of FIG. 2. As shown in the example of FIG. 8, user interface 800 presents training information 117 in the form of multiple modalities of data, such as video data 107, audio data 109, sensor data 121, and portions of domain documents 104, where each type of data is correlated based on the presence of the same object, concept, or step in the performance of a task.
[0089] As an illustrative example, user interface 800 includes displays 802, 804, 806, 808, 810, 812, and 814. Display 802 depicts a time series graph of video data 107. Display 804 depicts a transcription of a portion of audio data 107 (e.g., a narration by first user 102) that correlates to a respective portion of video data 107. Display 806 identifies a portion of video data 107 as corresponding to a recognized action sequence (e.g., one step of multiple steps for performing a task). Display 808 depicts a label of the recognized action sequence for that portion of video data 107 identified by display 806. Display 810 identifies a portion of video data 107 as depicting a recognized object. Display 812 identifies a location of a recognized object within a workspace depicted in video data 107. Display 814 identifies a portion of video data 107 as depicting a recognized tool.
[0090] Each of the displays 802, 804, 806, 808, 810, 812, and 814 is informed by the domain documents 104, the video data 107, the audio data 109, and / or the sensor data 121. In some examples, the system 100 of FIG. 1 acquires the video data 107, the audio data 109, and / or the sensor data 121 simultaneously such that the computation engine 130 can correlate portions of the video data 107, the audio data 109, and / or the sensor data 121 with each other based on time correspondence in a chronological order. In some examples, the system 100 of FIG. 1 acquires the video data 107, the audio data 109 and / or the sensor data 121 asynchronously or at different times, such that the computational processing engine 130 may correlate portions of the video data 107, the audio data 109 and / or the sensor data 121 with each other, for example, based on recognition of objects recognized in the video data 107, references to such objects recognized in the audio data 109 and / or measurements obtained from the sensor data 121.
[0091] In some examples, the user interface 800 may output a representation of the differences in the map of a first SME performing a task versus a second SME performing the same task versus a novice user performing the same task. For example, the representation may include one or more highlighted portions that indicate the changes in behavior or differences in actions performed by each user. Additionally, the user interface 800 may allow the user to zoom in to view a particular area of interest that represents the differences between the behavior or actions of the different users.
[0092] 9 is a flow chart illustrating example operations for generating training information in accordance with the techniques of this disclosure. For convenience, FIG. 9 is described with respect to FIG.
[0093] As depicted in the example of FIG. 9, the video device 106 acquires (902) video data 107 of the first user 102 performing a task. The video data 107 may include multiple camera sources. For example, the video device 106 includes a first video device and a second video device. The first video device is configured to acquire video data of the first user 102 performing a task from the perspective of a first-person perspective of the first user 102. The second video device is configured to acquire video data of the first user 102 performing a task from the perspective of a third-person perspective of the first user 102. The video recognition unit 214 receives the video data 107 from the video device 106 of FIG. 1 and performs object recognition to identify one or more objects depicted in the video data 107. In some examples, the video recognition unit 214 identifies the first user 102, one or more tools used by the first user 102, one or more workpieces with which the first user 102 interacts, etc. In some examples, the video recognition unit 214 processes the video data 107 to generate video data labeled with annotations of human poses, objects, or activity sequences.
[0094] The audio device 108 obtains (904) audio data 109 of the first user 102 performing the task. In some examples, the audio data 109 includes speech by the first user 102 describing the first user's actions while performing the task. In other examples, the audio data 109 includes speech by the first user 102 describing how to perform the task while the first user is not performing the task, such as during an interview with the first user 102 before or after performing the task. The speech recognition unit 212 receives the audio data 109 from the audio device 108 of FIG. 1 and performs speech recognition to identify references to one or more objects or concepts present in the audio data 109.
[0095] The sensor device 120 acquires (906) sensor data 121 of the first user 102 performing the task. The sensor 120 may include, for example, one or more motion sensors, pressure sensors, force sensors, or acceleration sensors. In some examples, the sensor 120 is incorporated into an article worn by or worn by the first user 102, such as a motion tracking glove that detects the motion and / or force of the user's fingers, hand, and / or arm. In some examples, the sensor 120 is incorporated into one or more tools used by the first user 102, such as a smart tool that incorporates one or more pressure sensors to detect the motion and force of the tool being used by the user. In some examples, the sensor 102 is an external sensor that senses data related to the first user 102, the workspace of the first user 102, or an object with which the first user 102 interacts, such as a force pad that detects the force the user applies to a surface, such as a work surface, workpiece, tool, or an IMU that detects the acceleration of the first user 102.
[0096] The text recognition unit 202 receives (908) text data in the form of domain documents 104 related to the performance of a task. Examples of domain documents 104 include instruction manuals for performing the task, parts lists of parts needed to perform the task, tools lists of tools needed to perform the task, or other written guides. The text recognition unit 202 performs text recognition to obtain text data suitable for use by the machine learning system 112.
[0097] The machine learning system 112 correlates (910) the video data 107 with the audio data 109, the sensor data 121, and the text data obtained from the domain documents 104. In some examples, the machine learning system 112 correlates at least a portion of the video data 107 with at least a portion of the audio data 109, at least a portion of the sensor data 121, and at least a portion of the text data obtained from the domain documents 104. Based on the correlation of the video data 107 with the audio data 109, the sensor data 121, and the text data, the machine learning system 112 may identify at least a portion of the video data depicting a step of a plurality of steps for performing a task, at least a portion of the audio data describing the same step of the plurality of steps for performing a task, at least a portion of the sensor data describing the same step of the plurality of steps for performing a task, and at least a portion of the text data describing the same step of the plurality of steps for performing a task. For example, the machine learning system 112 extracts first semantic information from the video data 107, second semantic information from the audio data 109, third semantic information from the sensor data 121, and fourth semantic information from the text data obtained from the domain documents 104. The machine learning system 112 correlates the first semantic information with the second, third, and fourth semantic information to identify portions of each of the video data 107, the audio data 109, the sensor data 121, and the text data obtained from the domain documents 104 that depict the same step in performing a task.
[0098] The machine learning system 112 processes the correlated video data 107, audio data 109, sensor data 121, and text data obtained from the domain documents 104 to update the domain model 114 to more accurately or more comprehensively describe the performance of the task (912). The domain model 114 provides a model of the task performed by the first user 102. In some examples, the domain model 114 models the task as multiple steps performed to achieve a particular goal. In some examples, the domain model 114 models the task by specifying at least one of an ontology, a cluster, an entity, an action, an event, or a rule (e.g., a semantic rule) associated with the performance of the task. For example, the machine learning system 112 may correlate a portion of the video data to a portion of the audio data by identifying an object depicted in the video data 107, identifying a reference to the object from the audio data 109, and correlating the object depicted in the video data 107 to the reference to the object in the audio data 109. The machine learning system 112 may then update the domain model 114 based on correlations between objects identified from the video data 107 and references to objects identified from the audio data 109, for example by using the correlations to define rules in the domain model 114 that define an ontology, entity, action, event, or task performance.
[0099] The training unit 210 applies the domain model 114 to generate (914) training information 117 for the task. The training information 117 may include, for example, video data, audio data, sensor data, and / or textual data related to the task, one or more of the steps that make up the task, or objects (e.g., tools or workpieces) related to each type of data cross-referenced to other types of data.
[0100] The training unit 210 outputs (916) training information 117 for use in training the second user 118 to perform the task. For example, the training information 117 may take the form of an interactive multimedia manual in the form of textual, audio, and / or video information with which the second user 118 may interact. As another example, the training information 117 may take the form of augmented reality content. For example, the training unit 210 may output such augmented reality content, for example, to an HMD worn by the second user 118 to provide an experiential first-person perspective of the performance of the task by the SME. In some examples, the augmented reality content may include a relevant portion of the audio data 109, such as a narration by the first user 102, and a relevant portion of the video data 107, such as a perspective of the first user 102 when performing the task.
[0101] The techniques described in this disclosure may be implemented, at least in part, in hardware, software, firmware, or any combination thereof. For example, various aspects of the techniques described may be implemented in one or more processors, including one or more microprocessors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs), or any other equivalent integrated or discrete logic circuitry, as well as any combination of such components. The term "processor" or "processing circuitry" may generally refer to any of the foregoing logic circuitry, alone or in combination with other logic circuitry, or to any other equivalent circuitry. A control unit including hardware may also implement one or more of the techniques of this disclosure.
[0102] Such hardware, software, and firmware may be implemented within the same device or within separate devices to support the various operations and functions described in this disclosure. Furthermore, any of the described units, modules, or components may be implemented together or separately as discrete but interoperable logical devices. The depiction of various functions as modules or units is intended to highlight various functional aspects and does not necessarily imply that such modules or units must be realized by separate hardware or software components. Rather, functionality associated with one or more modules or units may be performed by separate hardware or software components or integrated within a common or separate hardware or software component.
[0103] The techniques described in this disclosure may also be embodied or encoded in a computer readable medium, such as a computer readable storage medium containing instructions. The instructions embedded or encoded in the computer readable storage medium may, for example, cause a programmable processor or other processor to perform the method when executed. The computer readable storage medium may include a random access memory (RAM), a read only memory (ROM), a programmable read only memory (PROM), an erasable programmable read only memory (EPROM), an electronically erasable programmable read only memory (EEPROM), a flash memory, a hard disk, a CD-ROM, a floppy disk, a cassette, a magnetic medium, an optical medium, or other computer readable medium. The above-mentioned embodiment can be described as follows, but is not limited to the following. [Configuration 1] 1. A system for acquiring knowledge for performing a task, the system comprising: A domain model that specifies the execution of said tasks; a video input device configured to obtain video data of a first user performing the task; a voice input device configured to obtain voice data describing the performance of the task; one or more sensors configured to generate sensor data during performance of the task; A computing engine, comprising: correlating at least a portion of the video data with at least a portion of the audio data and at least a portion of the sensor data; a computational processing engine configured to process the at least a portion of the correlated video data, the at least a portion of the audio data, and the at least a portion of the sensor data to update the domain model that governs the execution of the task; a training unit configured to apply the updated domain model to generate training information for performing the task; an output device configured to output the training information for use in training a second user to perform the task. [Configuration 2] To correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data, the computational processing engine is a machine learning system, 2. The system of claim 1, comprising: a machine learning system configured to correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data to identify at least a portion of the video data depicting one step of a plurality of steps for performing the task, at least a portion of the audio data describing the same step of the plurality of steps for performing the task, and at least a portion of the sensor data describing the same step of the plurality of steps for performing the task. [Configuration 3] To correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data, the machine learning system comprises: extracting from the at least a portion of the video data first semantic information related to performance of the task; extracting second semantic information from the at least a portion of the speech data that is related to performance of the task; extracting third semantic information related to performance of the task from the at least a portion of the sensor data; 3. The system of claim 2, configured to identify from the first semantic information, the second semantic information, and the third semantic information: the at least a portion of the video data depicting the one step of the plurality of steps for performing the task, the at least a portion of the audio data describing the same step of the plurality of steps for performing the task, and the at least a portion of the sensor data depicting the same step of the plurality of steps for performing the task. [Configuration 4] the machine learning system configured to correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data comprises a first machine learning system configured to correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data; to process the at least a portion of the correlated video data, the at least a portion of the audio data, and the at least a portion of the sensor data to update the domain model governing the execution of the task, the computing engine comprising a second machine learning system configured to process the at least a portion of the correlated video data, the at least a portion of the audio data, and the at least a portion of the sensor data to update the domain model governing the execution of the task; To correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data, the first machine learning system: identifying, from the at least a portion of the video data, one or more objects used to perform the task; identifying, from the at least a portion of the audio data, references to the one or more objects used to perform the task; identifying, from the at least some of the sensor data, one or more physical measurements of the one or more objects used to perform the task; configured to correlate the one or more objects identified from the video data with the references to the one or more objects identified from the audio data and with the one or more physical measurements of the one or more objects identified from the sensor data; 3. The system of claim 2, wherein the second machine learning system is further configured to: define at least one of an ontology, an entity, an action, an event, or a rule governing execution of the task of the domain model based on correlations between the one or more objects identified from the video data, the references to the one or more objects identified from the audio data, and the one or more physical measurements of the one or more objects identified from the sensor data to process the at least a portion of the correlated video data, the at least a portion of the audio data, and the at least a portion of the sensor data to update the domain model governing execution of the task. [Configuration 5] the video data of the first user performing the task includes first video data of the first user performing the task; the audio data describing performance of the task includes first audio data of the first user narrating performance of the task; the sensor data generated during performance of the task includes first sensor data generated during performance of the task by the first user; the video input device is further configured to receive second video data of at least one of the second user or a third user performing the task; the audio input device is further configured to receive second audio data of the at least one of the second user or the third user narrating the performance of the task; the one or more sensors are further configured to generate second sensor data during performance of the task by the at least one of the second user or the third user; The arithmetic processing engine includes: correlating at least a portion of the first video data with at least a portion of the first audio data and with at least a portion of the first sensor data; correlating at least a portion of the second video data with at least a portion of the second audio data and with at least a portion of the second sensor data; 2. The system of claim 1, configured to process at least a portion of the correlated first video data, at least a portion of the first audio data, and at least a portion of the first sensor data, and at least a portion of the correlated second video data, at least a portion of the second audio data, and at least a portion of the second sensor data, to identify semantic differences between the performance of the task by the first user and the performance of the task by the at least one of the second user or the third user. [Configuration 6] the at least one of the second user and the third user includes the third user; the output device is configured to query at least one of the first user and the at least one of the second user or the third user for an explanation of an identified semantic difference between the performance of the task by the first user and the performance of the task by the at least one of the second user or the third user; 6. The system of claim 5, wherein the computing engine is further configured to update the domain model that governs the performance of the task based on an explanation of the identified semantic differences between the performance of the task by the first user and the performance of the task by the at least one of the second user or the third user. [Configuration 7] the at least one of the second user and the third user includes the second user; To generate the training information for performing the task, the training unit is configured to generate feedback to the second user describing the identified semantic differences between the performance of the task by the first user and the performance of the task by the second user; 6. The system of claim 5, wherein to output the training information for use in training the second user to perform the task, the output device is configured to output the feedback to the second user describing the identified semantic differences between the performance of the task by the first user and the performance of the task by the second user. [Configuration 8] 2. The system of claim 1, wherein the video data includes video data annotated with at least one of human pose data or object detection data of the first user. [Configuration 9] The system further comprises a document processing unit configured to receive text data describing the performance of the task; to correlate the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data, the computational engine is configured to correlate the at least a portion of the video data with the at least a portion of the audio data, the at least a portion of the sensor data, and at least a portion of the text data; 2. The system of claim 1, wherein the computational processing engine processes at least a portion of the correlated video data, at least a portion of the audio data, at least a portion of the sensor data, and at least a portion of the text data to update the domain model that governs the execution of the task. [Configuration 10] 10. The system of claim 9, wherein the text data describing the performance of the task includes one or more instructions for performing the task or a list of tools used during the performance of the task. [Configuration 11] 2. The system of claim 1, wherein the one or more sensors are configured to generate the sensor data based on input from one or more of the first user, a workspace of the first user, one or more tools used by the first user while performing the task, or one or more objects used in performing the task. [Configuration 12] The sensor data is Data relating to at least one of micro-movements or actions of the first user; data relating to one or more objects with which the first user interacts during performance of the task; data relating to one or more finger or hand movements of the first user, rotation of the wrist of the first user, or hand or finger pressure applied to one or more objects of the first user; or 2. The system of configuration 1, including one or more sensor data relating to one or more of: an angle between one or more tools and the one or more objects, a pressure on the one or more objects, a surface feature of the one or more objects, or an acceleration of the one or more objects. [Configuration 13] The video input device includes: a first video input device configured to capture first video data of the first user performing the task from a first-person perspective of the first user; a second video input device configured to obtain second video data of the first user performing the task from a third-person perspective of the first user; 2. The system of claim 1, wherein the video data of the first user performing the task includes the first video data and the second video data. [Configuration 14] To generate training information for performing the task, the training unit: augmented reality content for training the second user to perform the task; or 10. The system of claim 1, further comprising: a) a user interface for providing a user interface for a user to communicate with the user; and b) an interactive technical manual for training the second user to perform the task. [Configuration 15] The training unit is further configured to receive a query from the second user for instructions to perform the task; 2. The system of claim 1, wherein the training unit is further configured to apply the updated domain model to generate the training information for performing the task in response to receiving the query for instructions to perform the task. [Configuration 16] the computational processing engine is configured to process the at least a portion of the video data, the at least a portion of the audio data, and the at least a portion of the sensor data that are correlated to identify semantic differences between the at least a portion of the video data, the at least a portion of the audio data, or the at least a portion of the sensor data; the output device is configured to query the first user for an explanation of the identified semantic difference between the at least a portion of the video data, the at least a portion of the audio data, or the at least a portion of the sensor data; 2. The system of claim 1, wherein the computational processing engine is configured to update the domain model that governs the execution of the task based on an explanation of the identified semantic differences between the at least a portion of the video data, the at least a portion of the audio data, or the at least a portion of the sensor data. [Configuration 17] the audio data describing the performance of the task includes audio data of an interview with the first user; a model of decision making during the performance of said task; and identifying one or more procedural skills in at least one of mechanical, technical or artisanal domains. [Configuration 18] 1. A method for acquiring knowledge for performing a task, the method comprising: acquiring video data of a first user performing the task by a video input device; obtaining, by a voice input device, voice data describing the performance of said task; generating sensor data during performance of the task by one or more sensors; correlating, with a processing engine, at least a portion of the video data with at least a portion of the audio data and at least a portion of the sensor data; processing, by the computing engine, the at least a portion of the correlated video data, the at least a portion of the audio data, and the at least a portion of the sensor data to update a domain model that governs the execution of the task; applying, by a training unit, the updated domain model to generate training information for performing the task; and outputting, by an output device, the training information for use in training a second user to perform the task. [Configuration 19] Correlating the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data comprises: 19. The method of claim 18, comprising: correlating the at least a portion of the video data with the at least a portion of the audio data and the at least a portion of the sensor data to identify at least a portion of the video data depicting one step of a plurality of steps for performing the task, at least a portion of the audio data describing the same step of the plurality of steps for performing the task, and at least a portion of the sensor data describing the same step of the plurality of steps for performing the task. [Configuration 20] A non-transitory computer-readable medium containing instructions that, when executed, cause a processing circuit to: acquiring video data of a first user who executes the task; obtaining speech data describing the performance of said task; generating sensor data during execution of said task; correlating at least a portion of the video data with at least a portion of the audio data and at least a portion of the sensor data; processing the at least a portion of the correlated video data, the at least a portion of the audio data, and the at least a portion of the sensor data to update a domain model that governs the performance of the task; applying the updated domain model to generate training information for performing the task; and outputting the training information for use in training a second user to perform the task.
Claims
1. 1. A system for acquiring knowledge for performing a task, comprising: A domain model defining a number of steps for performing the task; A computation engine having a processing circuit, receiving first video data depicting performance of the task by a first user and first audio data describing performance of the task by the first user; receiving second video data depicting performance of the task by a second user and second audio data describing performance of the task by the second user; processing the first video data, the first audio data, the second video data, and the second audio data to identify at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user; outputting a query to the first user for an explanation of the at least one semantic difference between the first user's performance of the task and the second user's performance of the task; receiving from the first user the description of the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user; processing the description of the at least one semantic difference between the first user's performance of the task and the second user's performance of the task to update the domain model defining the steps for performing the task; A calculation processing engine configured as above, A training unit comprising: generating training information for performing the task based at least in part on the updated domain model, the training information including an indication of the at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user; A training unit configured as follows:
1. An augmented reality content unit, comprising: converting the training information into augmented reality content for use in training the second user to perform the task, the augmented reality content including force feedback based on sensor data that replicates performance of the task by the first user; outputting the augmented reality content to the second user to guide the second user in performing the task. An augmented reality content unit configured as follows: A system comprising:
2. 2. The system of claim 1, wherein the display of the at least one semantic difference between the first user's performance of the task and the second user's performance of the task comprises a map visualizing the at least one semantic difference between the first user's performance of the task and the second user's performance of the task.
3. The map visualizing the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user comprises: a map visualizing the difference between a first pressure applied by the first user to at least one of a tool or a workpiece and a second pressure applied by the second user to at least one of a tool or a workpiece; a map visualizing the difference between a first hand movement of the first user and a second hand movement of the second user; a map visualizing the difference between a first body movement of the first user and a second body movement of the second user; The system of claim 2 , further comprising at least one of:
4. The arithmetic processing engine includes: receiving first sensor data during performance of the task by the first user; receiving second sensor data during performance of the task by the second user; 2. The system of claim 1, wherein the computing engine is further configured to process the first video data, the first audio data, the first sensor data, the second video data, the second audio data, and the second sensor data to identify the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user.
5. The system of claim 4 , wherein the first sensor data comprises sensor data of at least one of the first user's hand, a tool, or a workpiece during performance of the task by the first user.
6. The computing engine processes the first video data, the first audio data, the second video data, and the second audio data to identify the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user. extracting first semantic information related to the execution of the task by the first user from the first video data and the first audio data; identifying, from the first semantic information related to performance of the task by the first user, at least a portion of the first video data and at least a portion of the first audio data that depict one step of a plurality of steps for performing the task; extracting second semantic information related to the execution of the task by the second user from the second video data and the second audio data; identifying, from the second semantic information related to performance of the task by the second user, at least a portion of the second video data and at least a portion of the second audio data that depict the same step of the plurality of steps for performing the task; 2. The system of claim 1 , configured to identify at least one semantic difference between at least a portion of the first video data and at least a portion of the first audio data depicting the step of the plurality of steps for performing the task and at least a portion of the second video data and at least a portion of the second audio data depicting the same step of the plurality of steps for performing the task.
7. The computing engine processes the first video data, the first audio data, the second video data, and the second audio data to identify at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user. identifying, from at least a portion of the first video data depicting a step of a plurality of steps for performing the task, one or more first objects used in the performance of the task by the first user; identifying, from at least a portion of the first audio data depicting a same step of the plurality of steps for performing the task, a first reference to the one or more first objects used in performing the task by the first user; identifying one or more second objects used in the performance of the task by the second user from at least a portion of the second video data depicting the same steps of the plurality of steps for performing the task; identifying second references to the one or more second objects used in performing the task by the second user from at least a portion of the second audio data depicting the same steps of the plurality of steps for performing the task; 2. The system of claim 1, configured to identify at least one semantic difference between (1) the one or more first objects identified from the first video data and the first references to the one or more first objects identified from the first audio data, and (2) the one or more second objects identified from the second video data and the second references to the one or more second objects identified from the second audio data.
8. the first voice data describing performance of the task by the first user comprises first voice data of the first user narrating performance of the task while the first user is not performing the task; 2. The system of claim 1, wherein the second audio data describing the performance of the task by the second user comprises second audio data of the second user narrating the performance of the task while the second user is not performing the task.
9. the first video data having first video data annotated with at least one of first human pose data or first object detection data of the first user; The system of claim 1 , wherein the second video data comprises second video data annotated with at least one of second human pose data or second object detection data of the second user.
10. The system of claim 1, wherein the training information further includes an interactive technical manual for guiding the second user in performing the task.
11. 1. A method for acquiring knowledge for performing a task, comprising: a computing engine executed by a processing circuit receives first video data depicting performance of the task by a first user and first audio data describing performance of the task by the first user; the computing engine receives second video data depicting performance of the task by a second user and second audio data describing performance of the task by the second user; the processing engine processes the first video data, the first audio data, the second video data, and the second audio data to identify at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user; the computing engine outputs a query to the first user for an explanation of the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user; the computing engine receiving from the first user the description of the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user; the computing engine processes the description of the at least one semantic difference between the first user's performance of the task and the second user's performance of the task to update a domain model defining a plurality of steps for performing the task; a training unit executed by the processing circuitry generates training information for performing the task based at least in part on the updated domain model, the training information including an indication of the at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user; an augmented reality content unit executed by the processing circuitry converts the training information into augmented reality content for use in training the second user to perform the task, the augmented reality content including force feedback based on sensor data that replicates performance of the task by the first user; The method of claim 1, further comprising: outputting the augmented reality content unit to the second user to guide the second user in performing the task.
12. 12. The method of claim 11, wherein the representation of the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user comprises a map visualizing the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user.
13. The map visualizing the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user comprises: a map visualizing the difference between a first pressure applied by the first user to at least one of a tool or a workpiece and a second pressure applied by the second user to at least one of a tool or a workpiece; a map visualizing the difference between a first hand movement of the first user and a second hand movement of the second user; a map visualizing the difference between a first body movement of the first user and a second body movement of the second user; The method of claim 12 , comprising at least one of:
14. the computing engine receives first sensor data during performance of the task by the first user; the computing engine receives second sensor data during performance of the task by the second user; The method further comprises the steps of:
12. The method of claim 11, wherein identifying the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user comprises the computing engine processing the first video data, the first audio data, the first sensor data, the second video data, the second audio data, and the second sensor data to identify the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user.
15. The method of claim 14 , wherein the first sensor data comprises sensor data of at least one of a hand, a tool, or a workpiece of the first user during performance of the task by the first user.
16. the first voice data describing performance of the task by the first user comprises first voice data of the first user narrating performance of the task while the first user is not performing the task; 12. The method of claim 11, wherein the second audio data describing the performance of the task by the second user comprises second audio data of the second user narrating the performance of the task while the second user is not performing the task.
17. the first video data having first video data annotated with at least one of first human pose data or first object detection data of the first user; The method of claim 11 , wherein the second video data comprises second video data annotated with at least one of second human pose data or second object detection data of the second user.
18. A non-transitory computer-readable medium having instructions that, when executed, cause a processing circuit of a system to acquire knowledge to perform a task: receiving first video data depicting performance of the task by a first user and first audio data describing performance of the task by the first user; receiving second video data depicting performance of the task by a second user and second audio data describing performance of the task by the second user; processing the first video data, the first audio data, the second video data, and the second audio data to identify at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user; outputting a query to the first user for an explanation of the at least one semantic difference between the first user's performance of the task and the second user's performance of the task; receiving from the first user the description of the at least one semantic difference between the performance of the task by the first user and the performance of the task by the second user; processing the description of the at least one semantic difference between the first user's performance of the task and the second user's performance of the task to update a domain model defining a plurality of steps for performing the task; generating training information for performing the task based at least in part on the updated domain model, the training information including an indication of the at least one semantic difference between a performance of the task by the first user and a performance of the task by the second user; converting the training information into augmented reality content for use in training the second user to perform the task, the augmented reality content including force feedback based on sensor data that replicates performance of the task by the first user; outputting the augmented reality content to the second user to guide the second user in performing the task. A non-transitory computer-readable medium that causes tasks to be performed.
Citation Information
Patent Citations
Human motion trajectory learning device in virtual environment
JP2000504854A
Working motion analysis method, working motion analysis apparatus and working motion analysis program
JP2007323222A
Work support device and work support method
JP2012234406A
Symbiotic helper
JP2013257565A
Work support system and work support program
JP2019109844A
