AI digital twin robot practical training platform for vocational training

Abstract

The application discloses an AI digital twin robot practical training platform for vocational technical training and relates to the technical field of digital twin practical training, which comprises the following steps: sequentially performing the following steps: constructing a task semantic contract and a target environment capability portrait; performing execution mapping and authentication mapping on semantic atoms, generating a target execution description and a double-axis loss matrix; generating a work step evidence template according to the double-axis loss matrix, collecting and forming standardized execution traces and synchronous quality scores; generating a consistency certificate based on the foregoing results, and determining a comparable mode and a release range according to the consistency certificate. The scheme can unify the task expression, execution judgment and evidence judgment of cross-brand workstations, reduce the scoring deviation caused by the differences in log openness, time asynchronization and the absence of structured program trees, improve the comparability, traceability and consistency of the release boundary of the practical training results, and enhance the stable operation of the training closed loop.

Description

Technical Field

[0001] This invention relates to the field of digital twin training technology, specifically to an AI digital twin robot training platform for vocational and technical training. Background Technology

[0002] It is mainly used in vocational schools, school-enterprise cooperative training bases, and enterprise skills certification platforms for scenarios such as robot loading and unloading, visual inspection, assembly, palletizing, and human-robot collaborative safety drills. Existing applications generally adopt the form of offline programming or virtual controllers, first establishing a 3D workstation model, and then conducting coordinated training with units such as programmable logic controllers, manufacturing execution systems, visual inspection, automated guided transportation, and conveying and warehousing. According to ABB's official robotStudio documentation, it can be used for modeling, programming, simulation, and cloud collaboration based on virtual controllers, indicating that the field already has the technical background of offline programming + digital twin + remote collaboration.

[0003] Two existing technologies that are similar to this application are: the intelligent manufacturing training platform and training method described in Chinese patent document CN114550570A. The platform described in this document includes a programmable logic controller, a manufacturing execution system control system, a vision inspection system, an automatic guided transportation system, a machine tool processing system, a warehousing system, a transmission system, and a digital twin system. It uses a virtual mirror modeling physical training platform for virtual training.

[0004] US Patent Document US20210138651A1 describes a robot digital twin control system for simulating an industrial environment. Its control design and testing system includes a user interface component, a simulation component, a controller simulation component, and a model import component. It integrates robot digital models from various manufacturers into a common simulation environment and communicates with manufacturers' proprietary simulation platforms to execute proprietary robot programs. US Patent Document US20240189997A1 describes a robot task orchestration method based on a universal graphical interface (GUI). This method receives universal commands through a robot-independent graphical interface, translates these universal commands into proprietary commands suitable for different manufacturers' hardware through a driver program, records task files, and identifies common tasks in different task files.

[0005] The three technologies mentioned above improve training organization methods from the aspects of training platform integration, manufacturer simulation reuse, and general task orchestration, respectively. However, they are not applicable to cross-brand vocational skills assessments. CN114550570A emphasizes the integration of physical equipment and digital twin workstations. When the same course task is transferred to different brand robots and different simulators, the process sequence, coordinate system definition, input / output timing, and safety status description may differ. Although Us20210138651A1 allows robots from different manufacturers to perform simulations together, its core is still performed by manufacturer-specific programs and control algorithms. Although US20240189997A1 can convert general commands into specific commands through drivers to find common tasks, the translatable command set is not necessarily equivalent to the same course task that can be used for examination and assessment. This is because different controllers have inconsistent log openness, time resolution, state observability, and safety mode visibility. High-frequency external synchronization may also cause packet skipping due to resource competition between controllers.

[0006] Therefore, when the same course task is moved between different workstations, the execution process may be operable, but the evaluation criteria may be inconsistent and the safety constraints may be difficult to verify in a consistent manner, thus affecting the fairness of scoring, the boundaries of training release, and the efficiency of equipment utilization.

[0007] Therefore, the technical problem that needs to be solved is: how to ensure the comparability and verifiability of process evidence and safety constraints when migrating the same course task from a vocational skills assessment cloud platform to different brands of robots, different simulators and different physical workstations. Summary of the Invention

[0008] (a) Technical problems to be solved

[0009] To address the shortcomings of existing technologies, this invention provides an AI digital twin robot training platform for vocational skills training. It generates step evidence templates based on a dual-axis loss matrix, collects and forms standardized execution traces and synchronization quality scores, and generates a consistency certificate based on the aforementioned results, thereby determining comparable models and release ranges. This solution can unify task expression, execution judgment, and evidence judgment across different brand workstations, reduce scoring biases caused by differences in log openness, time asynchrony, and missing structured program trees, improve the comparability of training results, and enhance the stable operation of the training closed loop; thus solving the technical problems described in the background art.

[0010] (II) Technical Solution

[0011] To achieve the above objectives, the present invention provides the following technical solution:

[0012] The AI ​​digital twin robot training platform for vocational and technical training includes receiving course tasks and target workstation environment information, breaking down the course tasks into multiple work steps, and further breaking down each work step into semantic atoms containing course semantics, process semantics, control semantics and safety semantics, and constructing a task semantic contract.

[0013] Construct a target environment capability profile that includes execution capabilities, observability capabilities, time capabilities, and security capabilities;

[0014] Based on the task semantic contract and the target environment capability profile, execution mapping and authentication mapping are performed on each semantic atom to generate a target execution description and a biaxial loss matrix containing execution retention and authentication retention. Based on the biaxial loss matrix, a process step evidence template is generated and standardized execution traces are collected. Based on the biaxial loss matrix, the process step evidence template and standardized execution traces, a consistency certificate is generated, and the comparable patterns and release range are determined based on the consistency certificate.

[0015] Furthermore, the construction of the task semantic contract includes: extracting the process step boundaries based on the course title, process description and scoring rules, and establishing semantic atoms corresponding to course semantics, process semantics, control semantics and safety semantics for each process step, so that each semantic atom is associated with preconditions, action content, post-state, time window and evidence slot.

[0016] Furthermore, the construction of the target environment capability profile includes: reading the controller configuration, simulator interface list and safety status mapping of the target workstation to form a set of declared capabilities, and performing active detection on the action primitive invocation capability, trajectory and input / output observability capability, alarm and safety status readability capability, and timestamp source stability capability to form a set of verified capabilities.

[0017] Furthermore, when there is a conflict between the declared capability set and the verified capability set, the target environment capability profile is corrected using the verified capability set, and a capability profile version identifier is written into the corrected target environment capability profile; when critical security states are not visible, the target environment capability profile is limited to correspond only to the simulation training path.

[0018] Furthermore, the execution mapping includes matching the objects, actions, preconditions, poststates, and security boundaries of each semantic atom, and writing them into the mapping types corresponding to the original mapping, split mapping, replacement mapping, and failed mapping.

[0019] The authentication mapping involves channel matching of the evidence slots corresponding to each semantic atom and generating cause codes and recoverable tags in the biaxial loss matrix.

[0020] Furthermore, when the security semantics corresponding to a certain semantic atom are not fully preserved, a high-priority restriction flag is written into the dual-axis loss matrix, and the high-priority restriction flag is used as the basis for determining the range of access when generating a consistency certificate in the future; when a certain semantic atom is executable but not authenticable, it is restricted to entering only the training path.

[0021] Furthermore, the work step evidence template includes mandatory evidence slots, alternative evidence slots, optional evidence slots, time alignment rules, missing data handling rules, and degradation rules corresponding to each work step, and binds the action segment, precondition segment, and poststate segment in the target execution description to the corresponding evidence slots.

[0022] Furthermore, by coupling to the robot controller, programmable logic controller, video acquisition unit and sensor edge acquisition gateway, trajectory data, input / output logs, alarm logs, operation events and video data are collected according to the work step evidence template, and the acquisition results are written with work step number, semantic atom number, channel source, event type, time stamp and sequence number.

[0023] Furthermore, when the target controller cannot provide a structured program tree, an alternative control structure is constructed based on the program text, trajectory data, input / output logs, alarm logs, and operation events. After matching the action keywords in the program text with the time sequence of the trajectory data and input / output logs, a standardized execution trace corresponding to the semantic atom is generated.

[0024] Furthermore, the synchronization quality score is obtained by performing sequential verification and cross-source fitting verification on the precondition events, action events and poststate events in the normalized execution trace under a unified time axis; when either the required evidence integrity rate or the synchronization quality score does not meet the corresponding preset conditions, the current step is limited to the corresponding mode in the partially comparable mode or the training mode according to the degradation rules.

[0025] Furthermore, the conformance certificate includes at least the task version identifier, capability profile version identifier, evidence template version identifier, execution retention score, authentication retention score, synchronization quality score, comparable mode, release range, and reason code summary. The scoring engine and release interface both perform subsequent processing based on the conformance certificate and standardized execution traces.

[0026] Furthermore, when the conformance certificate is characterized as fully comparable, standardized scores are output and restricted entity trial operation paths are opened; when the conformance certificate is characterized as partially comparable, only partial scores with restriction labels are output; when the conformance certificate is characterized as training-only mode, only training feedback is output and entry into the entity workstation release path is blocked.

[0027] Furthermore, if the same reason code appears repeatedly in consecutive tasks, the reason code will be written back to the corresponding object in the capability profile registry and the work step evidence template according to the corresponding type of capability deficiency and evidence deficiency. If the target workstation version change causes the capability profile version identifier to be inconsistent with the current version, the existing consistency certificate will be invalidated, and the construction of the task semantic contract and the construction of the target environment capability profile will be retried.

[0028] (III) Beneficial Effects

[0029] This invention provides an AI digital twin robot training platform for vocational and technical training, which has the following beneficial effects:

[0030] The course tasks are transformed into task semantic contracts, and each work step is decomposed into semantic atoms corresponding to course semantics, process semantics, control semantics, and safety semantics. Combined with the target environment capability profile, constraints are imposed on execution capability, observability capability, time capability, and safety capability, establishing a unified task benchmark and environment benchmark before migration and eliminating differences between workstations of different brands. Execution mapping and authentication mapping are performed on each semantic atom separately, generating a dual-axis loss matrix. This separately determines whether an action has been implemented and whether the evidence is valid. The target execution description retains both the action chain and the source of constraints and incomparable boundaries to prevent scoring from starting solely based on whether the program is runnable.

[0031] By generating step-by-step evidence templates, the edge acquisition gateway collects multi-source signals according to step number, semantic atom number, and evidence slot. Trajectory, input / output, alarm, video, and sensor information are unified into standardized execution traces. Subsequent judgments are all based on the same object index and time foundation, avoiding evidence gaps caused by scattered multi-source data. By constructing alternative control structures when the structured program tree is missing, synchronous quality scores and evidence closed-loop are used together, allowing low-openness workstations to continue unified processing. Log out-of-order issues, time drift, and channel gaps are used as boundaries to prevent raw acquisition noise from directly reaching the scoring stage.

[0032] By uniformly writing the dual-axis loss matrix, process evidence template, standardized execution trace, and synchronous quality score into the consistency certificate, and simultaneously writing the consistency certificate into the scoring engine, release interface, and reason code, the results of performance output, comparable mode, release range, and closed-loop correction can revolve around a single adjudication object, thereby achieving synergy at each step. Attached Figure Description

[0033] Figure 1 This is a schematic diagram of the overall architecture of the AI ​​digital twin robot training platform for vocational and technical training according to the present invention;

[0034] Figure 2 This is a schematic diagram illustrating the construction of the task semantic contract and the target environment capability profile in this invention;

[0035] Figure 3 This is a schematic diagram of the process of performing mapping, authentication mapping, and generating the biaxial loss matrix in this invention;

[0036] Figure 4 This is a schematic diagram illustrating the generation of process evidence templates and the issuance of collection contracts in this invention;

[0037] Figure 5 This is a schematic diagram illustrating the standardized execution trace construction and synchronous quality score generation in this invention;

[0038] Figure 6 This is a schematic diagram illustrating the closed-loop writeback process of consistency certificate generation, comparable mode determination, and score release in this invention. Detailed Implementation

[0039] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0040] Please see Figures 1-6 This invention provides an AI digital twin robot training platform for vocational and technical training, comprising:

[0041] This is geared towards training environments where the same course task requires continuous movement between different brands of robots, different controller versions, different simulators, and different physical workstations. Teachers do not submit abstract problems, but rather training tasks that include process actions, work sequences, on-site safety requirements, and grading criteria. Students are not faced with a single software screen, but rather a composite workstation comprised of robot controllers, programmable logic controllers, cameras, fixtures, video acquisition devices, and edge acquisition gateways.

[0042] Step 1: Convert the course tasks from teaching texts into task semantic contracts that can be understood by different robot workstations, and convert the target workstation from equipment list into a target environment capability profile that can be directly called by subsequent migration compilation, so as to provide task benchmarks and environment benchmarks under the same terminology for subsequent steps.

[0043] In vocational training, course assignments submitted by instructors typically include action requirements, process requirements, deduction rules, and on-site safety requirements. However, these elements are scattered across course titles, process cards, operating procedures, and assessment instructions. If this text is directly passed to subsequent migration compilers, different controllers can only read the parts that closely resemble their own program syntax, failing to reliably identify the relationships between process sequences, action triggering conditions, and safety prerequisites. This results in the same course assignment being broken down into program fragments of varying granularity at different workstations. To avoid this, course assignments are first converged into a unified task semantic contract, ensuring that all subsequent modules operate around the same set of steps and semantic atoms.

[0044] The task semantic contract service consists of a course parser, a step preformer, a semantic atom generator, and a contract encapsulator, sequentially coupled. After receiving the course title, process description, and scoring rules, the course parser first identifies the main actions with clear start and end points and process boundaries. The step preformer then concatenates these main actions into a sequence of steps, following the direction of pre-action-post-state. The semantic atom generator further expands each step along four directions: course semantics, process semantics, control semantics, and safety semantics. Finally, the contract encapsulator encapsulates the expanded results into a task semantic contract, reserving time windows, tolerances, and evidence slots for subsequent steps within each semantic atom. After this processing, subsequent modules no longer deal with natural language questions but with a well-structured, clearly defined, and comparable contract object.

[0045] When the course parser receives a course task, it does not directly use periods, semicolons, or paragraphs as the basis for segmentation. Instead, it first looks for process trigger points and physical trigger points in the task. Process trigger points refer to states that change the legality of subsequent actions, such as the workpiece being in place, the fixture being closed, the inspection being completed, and the safety door being reset. Physical trigger points refer to spatial states that change the motion path and equipment coordination mode, such as the robotic arm's starting position, approach position, gripping position, and withdrawal position.

[0046] After the course parser identifies these trigger points, the step normalizer merges the actions between adjacent trigger points into a single step, ensuring that each step retains both the action's purpose and the process boundary. Subsequently, the semantic atom generator performs four-level naming on this step, obtaining the... Each step of the semantic quad:

[0047]

[0048] Among them, the semantic quadruple of the process step : No. A complete semantic container for each process step This represents the work step number, which is a positive integer and increments sequentially within a single task; course semantics This represents the teaching objectives, deduction points, and completion conditions for this process step in the course assessment; its value is a finite ordered set. (Process semantics) : Represents the workpiece state, fixture state, pose constraints, and pre- and post-process relationships corresponding to this step, and its value is a finite ordered set;

[0049] Control semantics This step requires that the motion primitives, input / output trigger conditions, and action duration intervals be called, and their values ​​be a finite, ordered set; safety semantics. This step requires the invocation of a safe zone, mode switch, confirmation action, and prohibition action, with values ​​belonging to a finite ordered set. By assembling these four-tuples, requirements scattered across different documents are aggregated into a single step container. Subsequent modules only need to focus on the semantic four-tuples of this step. The explanation of the original text will not be repeated here.

[0050] In one specific embodiment, after the teacher submits a training task involving visually guided grasping and placement outside the collaborative area, the course parser first reads out the series of on-site actions: camera recognizing the workpiece position, robotic arm approaching the workpiece, gripper closing, robotic arm retracting, placement on the designated tray, and resetting after confirming the collaborative area is unoccupied. The step normalizer does not abruptly separate camera recognition and robotic arm approach into isolated segments; instead, it locks the logical relationship between the completion of recognition and the generation of the grasping pose and the robotic arm's movement from the waiting position to the approach position into the same main thread. At this point, the course semantics focuses on whether the student has completed the continuous operation from recognition to grasping, the process semantics focuses on whether the workpiece pose is correctly received, the control semantics focuses on the order of motion primitives and gripper triggering, and the safety semantics focuses on the collaborative area occupancy status and reset conditions. After these four layers of naming, the on-site teacher, the edge acquisition gateway, and the subsequent migration compiler are facing the same step semantic quadruple, instead of different text fragments as they each understand.

[0051] After obtaining the semantic quadruple for each work step, the semantic atom generator further refines the semantics to prevent subsequent steps from only making coarse-grained judgments on the entire work step. Specifically, it extracts the smallest indivisible judgment units within the course semantics, process semantics, control semantics, and safety semantics, and then strings these judgment units together into a single chain through cross-layer constraints. This single chain is not a simple list, but requires each semantic atom to simultaneously carry preconditions, action content, post-action state, time window, and evidence slot, thus expressing not only what to do, but also when to do it, under what state to do it, and what should be seen after completion. Each semantic atom can be characterized as:

[0052]

[0053] Among them, semantic atoms : indicates the first The smallest judgment unit; preconditions : Represents the workpiece, fixture, controller, or safety state that must be satisfied before this semantic atom takes effect; its value is a set of Boolean constraints; Action content : Represents the action or state transition to be performed, and its value is a set of action descriptions; post-state : Represents the set of states that should be reached after an action is executed; time window : Indicates the allowed start and end interval of the action relative to the previous semantic atom, and the value is an ordered time interval;

[0054] Evidence slot : Indicates the location of logs, trajectories, inputs / outputs, alarms, video, or sensor evidence that needs to be filled in subsequent steps, and the value is a finite set of slots.

[0055] Furthermore, the task semantic contract is written as a fixed field structure. For example, each semantic atom should include at least: atom number, process step number, semantic level type, preconditions, action content, post-state, time window, tolerance, required evidence slot, alternative evidence slot, scoring weight, and key marker. The semantic atom generation rules are as follows: if an action requires a separate precondition check, then a separate semantic atom is generated; if a post-state change must be observed after an action is triggered, then the action and the post-state together constitute a semantic atom; if the safety condition is only used as an outer constraint of the entire process step, then the safety semantic atom can cover multiple control semantic atoms; if there are two non-interchangeable triggering events within the same action, then it is split into two semantic atoms.

[0056] Through semantic atoms By definition, the course evaluation conditions, process boundaries, control triggering conditions, and safety constraints within the same process step are bound to a minimum unit, and the subsequent migration compiler only needs to compare semantic atoms one by one. This allows us to determine whether the target workstation has the necessary execution and certification foundations.

[0057] To ensure that cross-layer constraints don't remain merely descriptive, the step planner also chains semantic atoms into a single chain, prioritizing the anchoring of the target by course semantics, locking the object state by process semantics, defining the action path by control semantics, and limiting the permissible boundaries by safety semantics. For example, in the action of gripper closure, course semantics doesn't just require the student to press the button, but rather to trigger the gripper at the correct step; process semantics requires the workpiece to be in the gripping position; control semantics requires the gripper closure signal to appear after the specified action; and safety semantics requires the collaborative area to be unoccupied or in an allowed mode. Only when all four layers simultaneously write the same semantic atom will the action be considered a complete action in subsequent steps. Conversely, if only the gripper closure command is recorded without recording the preceding and following states, subsequent migration compilation will only yield a translatable command fragment, unable to determine whether the command still holds true within the course context.

[0058] In practice, course tasks are stably decomposed into semantic quadruplets of work steps, preventing different devices or teachers from making different granular breakdowns of the same task. Within each work step, course objectives, process states, control triggers, and safety boundaries are compressed into semantic atoms, allowing subsequent steps to continue processing along a unified semantic object without needing to reinterpret natural language. This combination enables the task semantic contract to possess a transferable, verifiable, and adjudicable subordinate structure from the outset.

[0059] Furthermore, even after the course tasks are organized into task semantic contracts, they still cannot be directly transferred to the target workstation for execution because different brands of robots, different controller versions, and different simulators do not support action primitives, log accessibility, timestamp sources, and safety states in a consistent manner. If only the equipment list or interface manual is read, the system can often only know what is theoretically supported; however, the training site is more concerned with whether, at this very moment, in the current version, and under the current wiring status, it can actually be called, seen, and confirmed.

[0060] Therefore, by establishing a capability profile, the target workstation is transformed into a target environment capability profile that can be directly invoked by subsequent steps. This makes subsequent migration and compilation no longer dependent on vendor promotion or static configuration, but on the field capability results formed by passive reading and active detection.

[0061] As a supplement: active detection is only performed in maintenance mode, no-load mode, or simulation linkage mode; for motion capabilities, only short-range, low-amplitude test trajectories that do not cross process positions are allowed; for I / O capabilities, only loopback readings or no-load triggers that do not drive high-risk actuators are allowed; for safety capabilities, only status bits and confirmation links are read, and no actual out-of-bounds actions are triggered; after detection, the edge acquisition gateway issues a reset confirmation and then writes the capability profile version identifier.

[0062] The capability profile building unit consists of a capability declaration reader, an active detection subunit, a conflict corrector, and an acquisition contract encapsulator, arranged sequentially. The capability declaration reader first extracts declared capability items from the controller configuration file, simulator interface list, programmable logic controller variable table, and safety signal mapping table. The active detection subunit then issues micro-benchmark detection tasks to the target workstation to verify the usability of key capability items. The conflict corrector compares the declared capabilities with the detection results to obtain a set of verified capability items. The acquisition contract encapsulator then binds this set to the input / output channels of the edge acquisition gateway, generating a target environment capability profile and a capability profile version identifier. The resulting capability profile is not merely a documentation document, but an environmental benchmark object with source, status, and boundary conditions.

[0063] In the best implementation, the capability reader does not copy all vendor documentation verbatim into the capability profile. Instead, it first categorizes the capabilities into four types that will actually be used in subsequent steps: execution capabilities, observability capabilities, timing capabilities, and security capabilities. The target environment capability profile can be represented as follows:

[0064]

[0065] Among them, target environment capability profile : Indicates the target workstation The set of capabilities under the current version and current wiring conditions; execution capabilities. : Represents the ability to invoke action primitives, handle coordinate systems, drive input / output, and switch modes; its value is a set of discrete capability items; observable capability. : Represents the available states of trajectory, input / output, alarm, operation event, program text, and structured program tree, with values ​​being a set of discrete capability items;

[0066] Time Capacity : Represents the timestamp source, sampling tick, rearrangement tolerance boundary, and local cache order; values ​​are a set of discrete capability items; security capabilities. : Indicates the visibility of the safe zone status, confirmation action, mode status and stop reset status, and takes the value of a discrete set of capability items.

[0067] After the capability declaration reader completes the merging, the active detection subunit targets the execution capabilities. Observability Time management ability Security capabilities Micro-reference probing tasks are issued item by item. These micro-reference probing tasks are not complete training programs, but rather short, small-scale verification actions that can be executed when the workstation is idle. Examples include reading whether the controller returns a response to a certain motion primitive, triggering an input-output loop that does not drive an actuator, checking the visibility of the alarm register, comparing the timestamps returned by the edge acquisition gateway and the controller to ensure they remain monotonically consistent, and checking for readable status bits after a safety mode switch. The purpose of this design is to prevent the probing actions from altering the tooling state in the training environment while simultaneously confirming the actual existence of the capability items.

[0068] Taking a gripping station containing a robot controller, a programmable logic controller (PLC), an industrial camera, and an edge acquisition gateway as an example, the capability reader first reads the motion instruction list from the controller's project file, the gripper's position and the workpiece's position from the PLC's variable table, and the cooperative area occupancy signal and reset button mapping from the safety mapping table. Subsequently, the active detection subunit sends a short-range test program to the controller to observe whether the controller returns the corresponding program text, trajectory fragments, and alarm words; it sends a loopback read request to the PLC to observe whether the input / output bits return according to the contract; and it requests the frame header timestamp from the video acquisition unit to observe whether the timestamp monotonically increases. After the detection is completed, the system obtains not the nominal functions of the equipment, but rather clear information about which capabilities of the current station are truly callable, observable, and usable for subsequent judgment in the current state.

[0069] Furthermore, when the declared capabilities are inconsistent with the results of active detection, the conflict corrector corrects the capability profile in the order of detection priority, declaration logging, and conflict marker retention. That is, if a capability item is declared in the vendor documentation but not verified in field detection, it will not be added to the set of verified capabilities; conversely, if a capability item is not explicitly stated in the documentation but can be reliably obtained through active detection, it will be added as a verified capability item with the detection source attached. After conflict correction, the acquisition contract encapsulator binds the capability profile to the acquisition channels of the edge acquisition gateway, eliminating the need for subsequent steps to search for channel sources again. The edge acquisition gateway preferably uses a dual-port industrial control computer or equivalent device, connecting the robot controller and programmable logic controller on one side and the private cloud exchange network on the other. Local power-off retention storage is used to save short-term caches, and a unified time reference is maintained at the system level. The acquisition contract is emphasized here because subsequent steps do not abstractly call logs, but specifically call which trajectory, input / output, alarm, or video channel.

[0070] If the channel binding is not completed in step one, the subsequent steps will lose continuity even with a task semantic contract due to the uncertainty of the evidence collection entry point.

[0071] In a parallel extended implementation, if the target workstation lacks complete metadata, the system directly enters the detection priority path. The active detection subunit sequentially verifies action primitives, trajectory visibility, input / output visibility, alarm visibility, and safety status visibility. Then, the conflict corrector writes the pass items into the target environment capability profile. If critical safety status is not visible, the acquisition contract encapsulator adds a training restriction flag to the capability profile, ensuring that subsequent steps can only enter simulation training and not enter the entity release candidate. If detection fails after a version change, the old version capability profile is frozen, and the new version workstation retains only the minimum training capability set. The above alternative paths and the best implementation use the same terminology system, that is, they always revolve around the target environment capability profile, edge acquisition gateway, declared capabilities, active detection subunit, and acquisition contract, and will not cause object mixing due to different workstation openness.

[0072] In practice, the execution boundaries, observation boundaries, time boundaries, and safety boundaries of the target workstation are converged into the same target environment capability profile. Subsequent steps can be directly invoked by capability item, instead of repeatedly parsing device documentation. The difference between the declared information and the on-site available status is pre-processed and fixedly bound to the specific channel of the edge acquisition gateway, thus providing subsequent steps with continuous data entry points and clear degradation entry points. In this way, the environmental benchmark upon which cross-brand migration relies is no longer limited to theoretical support but is grounded in the actual capabilities of the current workstation, current version, and current wiring.

[0073] Step 2: Based on the task semantic contract and target environment capability profile output in Step 1, complete the execution mapping and authentication mapping respectively, and condense the mapping results into target execution description and dual-axis loss matrix. Then, in Step 3, process evidence templates can be generated based on the determined execution boundary and evidence collection boundary.

[0074] Step one has already compressed the course tasks into task semantic contracts and the on-site workstations into target environment capability profiles. However, these two types of objects are still in a parallel state of task requirements and equipment status, and a one-to-one correspondence has not yet been established. If the system directly calls the target controller generation program at this point, it will often only obtain a sequence of actions that can be accepted by the controller, without being able to confirm whether the sequence of actions still retains the process sequence, control triggering conditions, and safety boundaries in the original course tasks. Therefore, an execution mapping chain is first established, connecting each semantic atom with the execution capabilities and safety capabilities in the target environment capability profile item by item, and then generating the target execution description, so that all subsequent actions are based on the premise that the original semantics are preserved.

[0075] The migration compiler consists of a semantic atom reader, an execution mapper, a constraint marker, and a target execution description wrapper connected sequentially. The semantic atom reader reads semantic atoms in the order of steps in the task semantic contract. and together with the semantic quadruple of the process steps The higher-level information is sent to the execution mapper; the execution mapper then obtains the target environment capability profile. The process reads execution and security capabilities, comparing the action content, preconditions, poststates, and security boundaries of semantic atoms item by item. The constraint marker determines whether the semantic atom needs to be split, replaced, extended in series, or have constraint markers added based on the comparison results. Finally, the target execution description encapsulator reassembles all processed semantic atoms into a target execution description for the target workstation. This description is not a natural language description, but a structured sequence containing action order, triggering conditions, alternative actions, constraint conditions, and source mapping. The subsequent step three will directly reference the relationship between action segments and evidence slots within this description.

[0076] The rules for performing mappings are as follows: If the source semantic atom's object, action, preceding and following states, and security boundaries can all be natively supported by the target environment, then the mapping type is native mapping; if the object and boundaries can be preserved, but the action needs to be split into multiple target action segments, then the mapping type is split mapping; if the action needs to be replaced by another object or another input source, but the preceding and following states can still be verified, then the mapping type is substitution mapping; if the preceding and following states or security boundaries cannot be established, then the mapping execution fails.

[0077] The smallest record unit of the target execution description must include at least: source semantic atom number, target action segment number, mapping type (native mapping / split mapping / merge mapping / alternate mapping), precondition binding item, poststate binding item, restriction flag, and evidence index entry. The original semantic atom can be split into multiple target action segments, but all target action segments must share the same source semantic atom number. Multiple original semantic atoms are only allowed to be merged if they have the same process goal, the same time window, and the same safety boundary. Once a split or merge occurs, the process evidence template in step three must use the source semantic atom number as the primary index, not the target action segment number.

[0078] In a preferred embodiment, the execution mapper does not directly allow the vendor controller to interpret the action content of semantic atoms. Instead, it processes the action in the following order: object alignment first, action implementation second, and boundary locking third. Object alignment first means checking whether the object acted upon by the semantic atom has a corresponding object at the target workstation, such as the workpiece coordinate system, gripper action, visual result input position, reset button, or collaborative area status position. Action implementation second means that after object alignment is established, it checks whether the corresponding action primitive and its combination exist at the target workstation. Boundary locking third means that after the action can be implemented, the process context and safety conditions are finally bound to the action segment. Execution retention is obtained using a coupled evaluation method.

[0079]

[0080] Among them, the retention rate : indicates the first The execution retention level of each semantic atom at the target workstation, with a value range of [value range missing]. When the value is closer The more completely the original semantics are implemented, the better.

[0081] Course Anchoring Coefficient : No. The degree of consistency between each semantic atom and the objectives of its respective course. The value range is... When a semantic atom is executable but has deviated from the course objectives, the course anchoring coefficient is... Take the smaller value; 1 indicates that the course objectives are fully retained; 0.5 indicates that the actions exist but the assessment granularity is reduced; 0 indicates that the course objectives are mismatched.

[0082] Process mapping value The feasibility of the process semantics at the target workstation, with a value range of [value range missing]. , indicates whether the workpiece state, fixture state, and conditions of the preceding and following processes are fully inherited; 1 indicates that the workpiece state and process boundaries are natively retained, 0.5 indicates that the replacement object or conversion rule is retained; 0 indicates that it cannot be retained.

[0083] Control mapping value : Controls the feasibility of semantic implementation at the target workstation, with a value range of . , indicates whether the action primitives, trigger order, and input / output linkage are fully inherited; 1 indicates that the action primitives and trigger order are natively retained; 0.5 indicates that the split action segments are retained; 0 indicates that they cannot be retained.

[0084] Security Mapping Value The feasibility level is the value representing the degree to which the safety semantics can be executed at the target workstation, with a value range of [value range missing]. 1 can indicate that the security zone, confirmation action, and prohibition conditions are fully inherited; 1 indicates that the security boundary and confirmation link are fully preserved; 0 indicates that they are restricted; 0 indicates that they are invisible and cannot be established.

[0085] The formula uses a three-way coupled fractional expression instead of nonlinear addition because, in the practical training environment, the process mapping value... Control mapping value Mapped values ​​with security The actions are not parallel and substitutable; if one of them is significantly reduced, the legitimacy of the action chain will be lowered. As a result, the execution mapper can identify action segments that seem executable but whose boundaries are actually broken in advance, and will not misjudge that the whole is fully preserved just because a certain individual ability is high.

[0086] For example, when a course task requires a robot to complete a grasping action along the approach, gripping, and withdrawal positions after recognizing a workpiece, the semantic atom reader first reads the sequence of semantic atoms: read visual pose, move to approach position, close gripper, and withdraw to safety position. If the robot controller in the target workstation does not support the continuous transition parameters used in the original workstation, but can execute linear and waiting actions in segments, the execution mapper will not directly determine that the semantic atom has failed. Instead, it will break down the original action into three action segments: enter approach position, pause confirmation, and enter gripping position, while writing the pause confirmation into the target execution description. The resulting action chain is visually longer than the original action, but because the process context, safety zone boundaries, and action order are still locked, the execution retention is maintained. It will not be directly reset to zero, but will be determined based on the process mapping value. Control mapping value and security mapping value The coupling result gives the corresponding value.

[0087] After the execution mapper obtains the execution preservation degree, it does not immediately regard the current result as the final migration result. Instead, the constraint marker continues to determine whether the semantic atom is a safe semantic atom, whether it triggers cross-layer constraints, and whether additional evidence is needed in subsequent step three. For ordinary semantic atoms, the constraint marker writes action replacement, action splitting, order preservation, and object change source into the target execution description. For safe semantic atoms, the constraint marker further writes high-priority constraint markers to inform subsequent steps three and four that even if the action of this semantic atom has been implemented, its evidence path and scope of release still need to be carefully checked. In a preferred embodiment, the target execution description consists of a target action segment, a precondition segment, a poststate segment, a constraint marker segment, and a mapping source segment in sequence. The target action segment describes the specific action that the target workstation should perform, the precondition segment describes the state before the target action begins, the poststate segment describes the state that should be reached after the target action ends, the constraint marker segment describes the constraints that need to be handled in subsequent steps, and the mapping source segment maps the target action segment to the original semantic atom one by one. With this encapsulation method, in the subsequent step three, when reading the target execution description, there is no need to trace back the teacher's original text. It is possible to directly know which semantic atom this action comes from, what restrictions it is subject to, and what its previous and subsequent states are.

[0088] In another parallel implementation, if the target workstation does not have a vision module installed, and a semantic atom in the course task requires reading a visual pose, the execution mapper will first check the execution and safety capabilities in the target environment capability profile, and then determine whether the semantic atom can obtain an alternative path through manual teaching pose input, a preset coordinate table, or a host computer forwarding table. If an alternative path exists, the marker is restricted from writing the source replacement into the target execution description, while maintaining the mapping relationship between the semantic atom and the original semantic atom; if no alternative path exists, the semantic atom is retained in the target execution description, but is marked as not allowed to be processed as a fully retained action in the future.

[0089] In practice, the migration compiler first establishes a single-chain mapping around objects, actions, and boundaries. Then, it binds the retention status of process constraints, control constraints, and security constraints to the same result through execution retention, avoiding incorrect judgments based solely on action statements. The target execution description is no longer an isolated program fragment, but a structured sequence carrying constraints and source mappings. Therefore, subsequent step three can directly perform forensics and downgrade processing around this structure.

[0090] Even with a generated target execution description, it still cannot be concluded that the semantic atom can be reliably confirmed by subsequent steps at the target workstation. In vocational skills training scenarios, scoring is not just about whether the robot completes the action, but also whether the action is completed in the correct sequence, in the correct state, and within the correct boundaries. If the target workstation can execute the action but cannot output corresponding logs, read the corresponding inputs and outputs, identify alarms, or obtain a continuous temporal trace, then the action segment can only be considered executed, not usable for cross-brand scoring.

[0091] Therefore, after performing the mapping, we continue to complete the authentication mapping, verify the evidence path of each semantic atom item by item, and write the results together with the execution preservation degree into the biaxial loss matrix.

[0092] After receiving the target execution description, task semantic contract, and target environment capability profile, the authentication mapper first locates the corresponding semantic atom according to the mapping source segment in the target execution description. Then, it checks whether the evidence slot required by that semantic atom can be filled by the existing channel of the target workstation. Here, the evidence slot originates from the task semantic contract in step one, and the channel originates from the target environment capability profile in step one. Therefore, each step of the authentication mapping corresponds one-to-one with the object in the previous step. When determining the evidence path, the authentication mapper not only checks whether a certain log exists, but also whether that log can form a closed loop with the preconditions, action segments, and post-states. If the closed loop is not established, even if a log file exists independently, it is not considered valid evidence.

[0093] Subsequently, the loss analysis unit writes the authentication mapping results and the completed execution retention into the dual-axis loss matrix, and adds a cause code and a recoverable marker to each loss. The authentication mapping procedure is as follows: read the mandatory evidence slots required by the semantic atom, match them with the corresponding observable channels in the target environment capability profile; check whether the channel has precondition evidence, action occurrence evidence, or poststate evidence. If so, find alternative evidence according to the work step evidence template. If not, the authentication mapping fails.

[0094] In the preferred embodiment, the authentication mapper checks the first... The evidence path for a semantic atom first checks mandatory evidence, then alternative evidence, and finally time closure. Mandatory evidence is evidence that the correct execution of the action cannot be guaranteed without it. Alternative evidence is evidence that can indirectly prove the action's validity from another source when mandatory evidence is unavailable. Time closure means that the occurrence of the precondition, the triggering of the action, and the confirmation of the subsequent state must be consecutive within the allowed time window. The authentication retention rate can be calculated using the following formula:

[0095]

[0096] Among them, certification retention Indicates the first The degree of evidence retention of each semantic atom at the target workstation, with a value range of [value range missing]. When the value is closer to 1, it means that the semantic atom is more likely to be confirmed by evidence loop-closing in subsequent steps.

[0097] Coefficient of closure of evidence : indicates the first Whether the required or alternative evidence for each semantic atom constitutes a complete closed loop is determined by a value of 0 or 1. If the closed loop is not valid, then... If all three conditions—precondition evidence, action occurrence evidence, and subsequent state evidence—are met, then the result is 1; otherwise, the result is 0.

[0098] Alignment Consistency Value : Indicates the degree of consistency between the evidentiary time sequence of this semantic atom and the contract time window, with a value range of . When the maximum observation interval does not exceed the allowed time window, Take 1; decrease proportionally if it exceeds; source mutual verification value : Indicates the degree of mutual corroboration of the same semantic atom among different sources of evidence, with a value of . Take the number of valid sources pointing to the same state transition / the number of sources already used;

[0099] The reason why this formula is adopted Together with the coupling fraction, this is because authentication mapping is not about having as much evidence as possible; rather, it's about first forming a closed loop, then comparing the consistency between time and source. If the closed loop doesn't exist, authentication retention... The authentication retention is directly reset to zero; if a closed loop exists but the time sequence or source verification is weak, then the authentication retention is... The level is lowered accordingly. This avoids the common misjudgment that the existence of a single path log constitutes authentication.

[0100] To illustrate with a real-world example: In a certain gripping station, the robot controller can output the trajectory, the gripper controller can output the closure position, and the programmable logic controller can output the workpiece's position. However, the video acquisition unit cannot continuously observe the gripping process due to lens obstruction. In this situation, the authentication mapper first checks the mandatory evidence slots of the semantic atom, finding that both the gripper closure position and the trajectory reaching the gripping position have been acquired. It then checks whether the workpiece's position changes within the allowed time window. If so, the evidence closure coefficient is determined. Pick The authentication mapper then compares the order of the trajectory, gripper position, and workpiece position to calculate the alignment value. Next, compare whether these sources all point to a crawl that has occurred, in order to calculate the source verification value. Even if the video is not present, it does not affect the authentication of the semantic atom, because the evidence path has already been closed by the structured channel. Conversely, if there is only video footage without the gripper position and workpiece position, the authentication retention will be compromised because the closed loop is not established. It will not be improved.

[0101] When performing retention Retention to certification Once both scores are obtained, the loss analysis unit writes them into the dual-axis loss matrix. The dual-axis loss matrix is ​​not simply a side-by-side arrangement of the two scores; instead, it records the execution retention state, authentication retention state, reason code, recoverable marker, source of the restriction marker, and target action segment index in the semantic atom order. This design aims to ensure that when step three reads the dual-axis loss matrix, it not only knows which item is deficient but also whether the deficiency stems from an inability to implement the action, a broken evidence path, a non-closed temporal sequence, or an invisible security state. In a preferred embodiment, when a semantic atom satisfies a high execution retention rate but a low authentication retention rate, the loss analysis unit directly writes that semantic atom as trainable only, not as a basis for cross-brand comparable scoring; when the semantic atom is a security semantic atom and no retention rate reaches full retention, a high-priority restriction marker is simultaneously added. The reason for writing this type of marker in this step is to avoid repeating the same legal boundary judgment in subsequent steps three and four, instead directly generating evidence templates and adjudicating consistency certificates based on the written restriction results.

[0102] In one implementation, if the target workstation outputs a structured program tree, the loss analysis unit inputs the node numbers and semantic atom numbers of the structured program tree into the biaxial loss matrix. If the target workstation cannot output a structured program tree, it writes the joint index of program text fragment position, trajectory fragment position, and input / output fragment position. Both methods aim to provide evidence for the next step, step three, ensuring that the correspondence is not lost due to different workstations. Combined with the source mapping of the target execution description, the biaxial loss matrix serves as the intermediate framework linking the task semantic contract, the target execution description, and the normalized execution trace.

[0103] In practice, the authentication mapping no longer treats log errors as valid evidence. Instead, it requires mandatory or alternative evidence to form a closed loop and further checks the time sequence and source verification. Therefore, it can more accurately distinguish between what can be executed and what can be confirmed by subsequent scoring. Execution retention and authentication retention are uniformly written into the dual-axis loss matrix, carrying the cause code, recoverable flag, and source of the limitation flag. Step three can then directly generate the work step evidence template and degradation strategy based on this matrix, without needing to parse the task text or device documentation again.

[0104] Step 3: Based on the biaxial loss matrix and target execution description generated in Step 2, construct the step evidence template for each step with closed-loop constraints. Then, at the edge acquisition gateway, transcribe the dispersed original signals into the standard execution traces that directly correspond to semantic atoms, and give the synchronization quality score and boundary degradation results so that Step 4 can generate a consistency certificate around the unified object.

[0105] Step two has already set the execution preservation degree for each semantic atom. Certification retention The reason code and limitation flags are written into the biaxial loss matrix, but these results are still judgment results and have not yet been transformed into a collection list that can be directly executed in subsequent collection actions. If the platform server selects a few logs as scoring criteria based solely on experience at this time, two problems are likely to occur: First, although some semantic atoms are identified as verifiable, it is not clearly constrained which evidence is mandatory and which is alternative evidence, resulting in different workstations having different evidence collection standards for the same step; Second, although some semantic atoms have high execution retention, their evidence paths have time gaps or source gaps. If they are not written into the evidence template in advance, the edge collection gateway will not know which signal to prioritize during on-site collection. Therefore, the biaxial loss matrix is ​​first transcribed into a step evidence template so that the evidence source, missing handling, and degradation conditions for each step are locked before collection.

[0106] The step evidence template generator in the platform server consists of a step splitter, an evidence slot allocator, an evidence level determiner, and a template encapsulator in sequence. The step splitter reads the action segment, precondition segment, poststate segment, and constraint marker segment sequentially according to the step boundaries in the target execution description, and then checks back the execution preservation degree of the corresponding semantic atoms in the dual-axis loss matrix. Certification retention The evidence slot allocator assigns mandatory, optional, and alternative evidence slots to each step based on the reason code. The evidence level determiner then assigns different weights to each slot according to the difficulty of evidence closure for each semantic atom within the step. Finally, the template encapsulator merges the step number, evidence level, time alignment rules, tolerance rules, and missing data handling rules into a step evidence template. This template is no longer an abstract explanatory text but a set of acquisition instructions for the edge acquisition gateway, upon which all subsequent acquisition actions are based.

[0107] The platform server comes pre-loaded with four types of dictionaries: course objective dictionary. Includes scoring objectives, compliance actions, and penalty actions; process status dictionary. Includes status terms such as workpiece in place, fixture closed, posture correction, and workpiece leaving; control action dictionary. : Contains action words such as move, wait, set, reset, clamp, release, etc.; Safety restriction dictionary This includes safety terms such as collaborative area occupation, protection stop, confirmation, and emergency stop reset.

[0108] After segmenting the course text into sentences, dictionary matching is performed sentence by sentence. If the same sentence matches multiple dictionary categories, it is written into four semantic layers according to the following priority: safety semantics > process semantics > control semantics > course semantics. The rule for determining the boundary of the process step is as follows: when a state completion predicate, action completion predicate, or object state significant change predicate appears in a sentence segment, a process step boundary is formed between the current sentence segment and the next sentence segment; otherwise, it is merged with the adjacent sentence segment, and each generated process step is further subdivided into semantic atoms.

[0109] In a preferred embodiment, the platform server does not include all visible signals within the same step in the formal forensics process. Instead, it first determines the authentication retention based on the authentication retention of each semantic atom in the biaxial loss matrix. The constraint markers determine which evidence is mandatory. Signals that, if missing, would prevent the formation of a closed evidentiary loop are written into mandatory evidence slots; signals that, even if missing, can still be indirectly proven by another source are written into alternative evidence slots; and remaining signals used to enhance explanatory power but do not determine the validity of the loop are written into optional evidence slots. To avoid template generation becoming a simple list, the platform server further calculates the... Weighting of evidence completeness for each step:

[0110]

[0111] Among them, the completeness of evidence in each step : indicates the first The evidence coverage level for each work step under the current template configuration, with a value range of [value range missing]. A higher value indicates a more complete formal data collection path for that step; semantic atom set : indicates belonging to the first The set of all semantic atoms for each process step, with a positive integer number of elements; execution preservation degree. Retention to certification All are inherited from step two, and the value range is... , respectively representing the first The degree of execution retention and evidence retention of each semantic atom; slot weights. Indicates the first The evidence priority of each semantic atom in the current step, with a value range of [value range missing]. Security semantic atoms and process key semantic atoms correspond to Take the higher value, and observe the semantic atom corresponding to the ordinary value. Take the lower value;

[0112] The function of this formula is not to generate an abstract score, but to drive the template generator to first identify the evidence sources most sensitive to loop closure at the step-by-step level. This is as long as the execution of a certain high-weight semantic atom maintains its degree of precision. Retention to certification Low, completeness of evidence in the work steps This will significantly reduce the requirement, so the template generator will prioritize adding stricter mandatory evidence requirements or a narrower range of alternative evidence to the semantic atom.

[0113] Taking a gripping station example: the robot controller outputs the trajectory segment, the gripper driver outputs the closure position, the programmable logic controller outputs the workpiece's position, and the video acquisition unit outputs the lateral view. For the step of closing the gripper after reaching the gripping position and confirming the workpiece has been lifted, the platform server writes the trajectory reaching the gripping position and the gripper closure position into the mandatory evidence slot, writes the workpiece's position change into either the mandatory evidence slot or an alternative evidence slot based on the on-site wiring, and writes the video image into the optional evidence slot. This way, the edge acquisition gateway will not treat all signals equally during subsequent acquisitions, but will prioritize the acquisition and caching links around the mandatory evidence. After the on-site action is completed, what is first visible are the trajectory segment, the gripper position change, and the workpiece position change, while the video image is retained as supplementary material for enhanced interpretation. This processing directly reflects the single-chain logic of first closing the loop and then enhancing.

[0114] After completing the evidence slot allocation, the template packager continues to distribute the process-level evidence templates to the edge acquisition gateway. For example, the edge acquisition gateway consists of a dual-port industrial control computer, an isolated digital acquisition board, a serial port expansion board, solid-state storage, and a local clock module. Its housing is an aluminum alloy heat sink fixed inside the robot control cabinet. Data input terminals are connected to the robot controller communication port, the programmable logic controller communication port, the video acquisition unit output port, and the sensor sampling board, respectively. The template packager does not distribute a set of general field names to the edge acquisition gateway; instead, it sends a complete set of acquisition contracts specifying the process step, the semantic atom, which path is mandatory evidence, which path is alternative evidence, the corresponding time alignment rules, and the missing data handling rules. Upon receiving the acquisition contract, the edge acquisition gateway establishes a process-level acquisition queue locally: mandatory evidence channels are prioritized for high-frequency polling or event-triggered listening; alternative evidence channels are subject to conditional listening; and optional evidence channels are subject to low-frequency sampling or post-processing readback. This acquisition queue corresponds one-to-one with the action segments in the target execution description, rather than simply corresponding to the device port table.

[0115] In a generalized implementation, if the target workstation uses an embedded edge box with equivalent functionality instead of an industrial control computer, then as long as the functional combination of multi-channel input, local buffering, unified time stamping, and template-driven acquisition remains unchanged, the same acquisition contract can still be executed. If the video acquisition unit of a workstation is replaced with a depth camera, the platform server does not change the upper-level structure of the work step evidence template. Instead, it registers the new channel source in the optional evidence slot or the alternative evidence slot, and continues to use the distribution format of work step number - semantic atom number - slot type - time alignment rule. Thus, the work step evidence template has parallel expansion capabilities for on-site hardware replacement, but its terminology system and action chain remain unchanged.

[0116] When in use, the step-by-step evidence template clearly separates mandatory, alternative, and optional evidence before collection, transforming subsequent collection from average collection to closed-loop collection. The step-by-step evidence template is further subdivided into collection contracts for edge acquisition gateways, ensuring that each signal on-site has a clear step and slot assignment, avoiding the need to reverse-engineer its use after collection.

[0117] Even if the evidence template for each step has been distributed to the edge acquisition gateway, the on-site acquisition results may still be affected by factors such as multi-source asynchronous processing, out-of-order logs, inconsistent controller output timing, and missing structured program trees. If the raw data is not encapsulated with the same standard, the subsequent step four, when calculating the consistency certificate, will still face a chaotic situation where the trajectory is recorded according to the controller time, the input and output are recorded according to the programmable logic controller scan time, and the video is recorded according to the acquisition card time. If the action context break caused by the missing structured program tree is not supplemented in this step, even if a lot of logs are obtained later, it will be impossible to clearly know which semantic atom a certain trajectory corresponds to.

[0118] Therefore, with the edge acquisition gateway as the center, the raw data is uniformly rearranged into standardized execution traces, and a synchronization quality score is further generated to provide stable input for step four.

[0119] The trace builder in the edge acquisition gateway consists of a raw channel access unit, a time stamper, a fragment merging unit, a control structure restorer, and a trace encapsulator in sequence. The raw channel access unit connects to the raw channels corresponding to each evidence slot according to the work step evidence template; the time stamper uniformly marks the data of each channel as event points or intervals on the local timeline of the edge acquisition gateway; the fragment merging unit merges trajectory fragments, input / output change fragments, alarm fragments, and video fragments according to the correspondence between work steps and semantic atoms; when the structured program tree is missing, the control structure restorer reconstructs the minimum control structure based on the program text, trajectory fragments, input / output fragments, and alarm fragments; the trace encapsulator ultimately forms a standardized execution trace. The platform server then reads the standardized execution trace, calls the synchronization quality evaluator to analyze the time sequence, source mutual verification, and rearrangement stability, and generates a synchronization quality score. The entire process proceeds in the order of time unification, action merging, structure restoration, and quality judgment to avoid signals from different sources being misjudged as evidence of the same action before unification.

[0120] In the preferred implementation, the edge acquisition gateway first prioritizes each channel according to the work step evidence template, and then writes a uniformly formatted event header to each original channel. The event header includes at least the work step number, semantic atom number, channel source, event type, local timestamp, and sequence number. For trajectory channels, the segment merging unit divides them into trajectory segments according to the motion start point, motion end point, and pause segment; for input / output channels, discrete event segments are formed according to state inversion points; for alarm channels, alarm segments are formed according to alarm setting, holding, and clearing; for video channels, corresponding image intervals are extracted according to the work step boundaries. If the target work station can output a structured program tree, the control structure restorer directly writes the structured program tree node number into the event header of the corresponding segment; if the target work station cannot output a structured program tree, the control structure restorer activates the trajectory segment graph rollback mechanism, splicing the action keywords in the program text, the spatial continuity of the trajectory segments, the input / output change sequence, and the alarm occurrence position into a minimum control structure graph. This minimum control structure diagram does not aim to reproduce the manufacturer's original syntax, but instead adds a chain relationship of previous action - current action - next action to each action segment, so that the subsequent step four can determine the sequence logic within the process step.

[0121] Taking an assembly station as an example: the robot controller continuously outputs three motion segments, the programmable logic controller sequentially outputs the clamping position, the position, and the release position, the alarm register remains silent throughout the entire operation, and the video acquisition unit records the workpiece being fed into the fixture. The edge acquisition gateway first divides the three motion segments into approach, insertion, and withdrawal segments according to the step boundaries, and then connects the clamping, position, and release positions between each segment according to their state flip points. If the structured program tree is not visible at this time, the control structure restorer uses the text segments "moved to insertion position," "fixture closed," "waiting for position," and "moved to withdrawal position" in the program text to string together the trajectory segments and input / output segments into a minimum control structure diagram of approach-closure-confirmation-withdrawal. In this way, what is read in the subsequent step four is not scattered motion data and input / output positions, but a standardized execution trace that can correspond to semantic atoms.

[0122] After the standardized execution traces are formed, the synchronization quality evaluator in the platform server continues to analyze the temporal consistency of segments from different sources. Instead of using a simple time mean, the synchronization quality evaluator examines sequential monotonicity, cross-source fit, and gap penalty separately. Sequential monotonicity is used to determine whether events within the same step maintain a unidirectional progression from preconditions to actions to subsequent states; cross-source fit is used to determine whether trajectory segments, input / output segments, and alarm segments are close to each other within the same permissible time window; gap penalty reflects whether long gaps occur in the required evidence channel. The overall synchronization quality score is obtained after comprehensive analysis.

[0123]

[0124] Among them, synchronous quality division : Indicates the stability of the current normalized execution trace over time, with a value range of . A higher value indicates a more stable cross-source time sequence; gap gating value : Indicates whether a time window that disrupts the closed-loop process occurs in the required evidence channel; the value is 0 or 1. When a disruptive window occurs, If any required evidence channel experiences consecutive gaps within the allowed time window of the process step, and these gaps disrupt the closed loop, then... ;otherwise .

[0125] Sequential monotonic values : Indicates the progress of the preconditions, action segments, and poststates within a work step on a unified time axis, with a value range of . Let the expected number of events for the current work step be... The number of events that satisfy the monotonic sequence of precondition → action → poststate is ,but Cross-source bonding value This indicates the degree of proximity between valid segments from the trajectory, input / output, alarms, and video within an allowed time window, with a value range of [value range missing]. Let the number of valid source logs participating in the comparison in the current step be . The number of source logs that fit within the allowed time window is ,but ;

[0126] The function of this formula is to first pass the gap gate value. Determine if a fatal gap exists, then use the sequential monotonic values. and cross-source bonding value This collectively reflects whether the multi-source data is sufficient to support the same process step. Therefore, even if one video stream has a significant delay, as long as the structured evidence can stably close the loop, the synchronization quality score will be satisfactory. It can still be kept within an acceptable range; conversely, as long as a long gap occurs in the required evidence channel, the synchronous quality score will be affected. This will lower the synchronization quality score, driving the subsequent step four to perform a degradation.

[0127] In a boundary scenario, the edge acquisition gateway detects continuous output of robot controller trajectory segments, but the gripper position of the programmable logic controller remains unchanged for a period of time, and the video acquisition unit fails to see the gripper's movements due to occlusion. In this situation, the trace builder can still generate normalized execution traces, but the synchronization quality evaluator will detect that the action segment has occurred but the required state segment has not kept up, thus reducing the cross-source fit value. If the gap continues to exceed the allowable time window specified in the work step evidence template, then the gap threshold value will be adjusted. Set as Subsequently, the platform server writes partially comparable or only training-available downgrade tags to the result at the process step level, instead of continuing to send the trace to the fully comparable path. This downgrade is done along the rules already in the process step evidence template, rather than through ad-hoc manual judgment, thus maintaining consistency in terminology and action chains.

[0128] In use, the edge acquisition gateway rewrites multi-source raw data into standardized execution traces under a unified header format and a unified process step index. Even when the structured program tree is missing, it can still recover the minimum control structure, ensuring that subsequent step four always revolves around the same object for certificate generation. The synchronous quality score compresses time sequence, cross-source alignment, and window conditions into a single constraint input, providing a clear basis for boundary degradation and preventing different modules from using different time standards. Step four generates a consistency certificate based on the process step evidence template, standardized execution traces, and synchronous quality score. This consistency certificate serves as the common basis for scoring, release, and closed-loop write-back, thus unifying the task benchmarks, environmental benchmarks, execution benchmarks, and evidence collection benchmarks formed in steps one through three at the final adjudication layer.

[0129] Step three has resulted in process evidence templates, standardized execution traces, and synchronized quality scores. However, these objects still serve template constraints, on-site trace representation, and time quality representation respectively, and have not yet been compressed into a unified object that can directly guide the scoring engine and the release interface. If the scoring engine continues to read the standardized execution traces for self-interpretation, and the release interface continues to read the dual-axis loss matrix for self-limitation, then the same process step is prone to split states where the score can be output but release is not allowed, or release is allowed but the score basis is insufficient. Therefore, step four sets up a consistency certificate service to uniformly encapsulate the aforementioned objects before performing comparable pattern determination, ensuring that all subsequent actions revolve around the same certificate object.

[0130] The consistency certificate service consists of a certificate aggregator, a pattern determiner, a release range generator, and a result encapsulator connected sequentially. The platform server first reads the biaxial loss matrix formed in step two from the certificate aggregator, then reads the work step evidence template, normalized execution traces, and synchronization quality score formed in step three. Following the work step order, it re-merges the execution retention, authentication retention, evidence loop closure status, time alignment status, and high-priority restriction markers for the same semantic atom. The pattern determiner then determines, based on the merged results, whether each work step and the entire task are fully comparable, partially comparable, or only usable for training. After determining the comparable patterns, the release range generator further limits the applicable boundaries of simulation execution, performance output, and restricted entity trial runs. The result encapsulator writes the task version identifier, capability profile version identifier, evidence template version identifier, total execution retention, total authentication retention, synchronization quality score, comparable patterns, and release range into the consistency certificate. After this encapsulation, subsequent modules no longer need to search for evidence among multiple intermediate objects; they only need to read the comparable patterns and release ranges based on the consistency certificate.

[0131] Preferably, the result encapsulator divides the consistency certificate into a set of sequential fields. The front end stores the task version identifier, capability profile version identifier, and evidence template version identifier. The middle section stores the total execution retention rate, total authentication retention rate, synchronization quality score, and high priority restriction summary. The back end stores the comparable pattern, release range, reason code index, and write-back target index. The sequential fields can be written to the platform server database or sent to the scoring engine and release interface through the message bus, thereby ensuring that the same decision call object is called in the same round.

[0132] In a preferred embodiment, the certificate aggregator does not simply average the execution retention, authentication retention, and synchronization quality scores of all steps. Instead, it first identifies key steps based on the high-priority restriction markers in step two and the step evidence templates in step three, and then generates a consistency total in the order of priority for key steps and subsequent for ordinary steps.

[0133] The certificate aggregation degree for the entire task is expressed by the following formula:

[0134]

[0135] Where: Certificate aggregation degree : Indicates the overall retention level of the current task before entering the final decision stage, with a value range of . A higher value indicates that the execution boundary, evidence collection boundary, and time boundary of the task are more consistent within the target workstation; security gate value : Indicates whether there are incompletely preserved restrictions in the security semantics, with a value of 0 or 1; when there are incompletely preserved security semantics and they are locked by a high-priority restriction flag, the security gating value is... Setting it to 0 blocks the task from entering the path that involves entity permission.

[0136] Overall performance retention : Represents the sum of the execution preservation results of all semantic atoms in the entire task, with a value range of . Overall performance retention The execution mapping results of semantic atoms directly from step two reflect the overall bearing of the target station on the original process constraints, original control constraints and original safety constraints;

[0137] Overall Certification Retention : Represents the sum of authentication retention scores for all semantic atoms in the entire task, with a value range of . Overall certification retention The authentication mapping result directly from step two reflects the level of evidence retention for the entire task at the current workstation; synchronous quality score. This indicates the stability of the multi-source time sequence output from step three, with a value range of [value range missing]. Synchronous quality analysis Its function is to characterize the degree of continuity between standardized execution traces under a unified timeline, and to prevent confusion in time boundaries from obscuring gaps in execution or evidence collection.

[0138] This formula employs a three-way coupling structure rather than linear superposition because the conformity certificate addresses not a single quality dimension, but rather the adjudication premise jointly constituted by execution reservation, evidence reservation, and time boundaries. If any one of these dimensions is significantly low, the certificate aggregation degree will decrease. It will be significantly lowered; once the security threshold is reached... If the value is set to zero, the entire task cannot enter the path that requires secure and complete preservation. Therefore, the certificate aggregator can intercept cases where the action is largely completed but the time boundaries are chaotic, or where the action and time boundaries are acceptable but the security semantics are insufficient, before the final decision.

[0139] In one field implementation, a student completes four steps at a handling station: workpiece grabbing, transfer, placement, and reset. Standardized execution traces show that the trajectory segments, input / output segments, and alarm segments for the first three steps have formed closed loops. However, the safety confirmation button for the reset step was not captured by the edge acquisition gateway due to changes in on-site wiring. After reading the dual-axis loss matrix, the certificate aggregator found that the safety semantic atom had a high-priority restriction marker. Simultaneously, the button signal was found to be a mandatory evidence slot in the step evidence template, thus triggering the safety gating value. It is placed in a blocked state. The consistency certificate service will not ignore the inadequacy of the last step simply because the first three steps were successful; instead, it will classify the entire task as a restricted path, ensuring that the subsequent allowance range is consistent with the on-site status. Teachers will not see a total score that deviates from the boundary on the platform interface, but rather a result object with a clear pattern description.

[0140] After certificate aggregation, the pattern determiner does not provide a final textual conclusion all at once. Instead, it makes determinations in the following order: first, steps; then tasks; first, restrictions; then results; first, scope; then display. For example, the pattern determiner first checks whether each step has a training-only limitation. If so, the step is not included in the fully comparable candidate. Then, it checks whether the entire task can only be included in the partially comparable category due to restrictions on one or more key steps. Finally, the release scope generator maps the pattern results to the simulation execution, result output, or restricted entity trial run scope.

[0141] Preferably, the release range generator presets three range objects: a simulation execution range to allow students to continue replaying and correcting within the simulator; a grade output range to allow teachers to view standardized grades or marked partial grades; and a restricted entity trial run range to allow the execution of action chains covered by restrictions at the physical workstation. Thus, full comparability does not directly equate to unrestricted release, but rather means that the entire task meets the conditions for entering the corresponding range within the current boundary; partial comparability does not equate to invalidating grades, but rather allows the output of partial grades within the retained set of work steps.

[0142] To ensure consistency between the action results and on-site observations, another example is provided below: In a vision-guided grasping scenario, the student completes camera recognition, approach, gripper closure, and withdrawal actions. The standardized execution traces show complete trajectory segments and gripper position changes, but workpiece position changes are only reflected laterally through the programmable logic controller's substitution signal. At this point, the pattern determiner reads that the authentication retention rate for this step is not fully retained, yet does not enter the training-only state; therefore, the entire task is judged to be partially comparable. Based on this, the release range generator allows the platform to output marked scores and allows the student to continue subsequent playback in the simulator, but does not directly push this result as a complete pass result on the physical workstation. This processing chain ensures that the score display and release boundary converge synchronously, preventing a disconnect where one interface declares pass while another rejects release.

[0143] When used, the consistency certificate service maintains overall execution consistency. Overall certification retention and synchronous quality Aggregation to the same certificate In, and use security gating values Security boundaries are pre-truncated to prevent different quality dimensions from obscuring each other. Comparable models and release ranges are generated in conjunction with the same certificate object, so the scoring engine and release interface share the same adjudication basis, and the results in the training session and the equipment release results converge in the same direction.

[0144] Furthermore, after the consistency certificate is generated, if the system only uses the certificate to display the results once and does not continue to apply it to subsequent corrections, the competency profile, work step evidence template, and collection contract formed in the previous three steps cannot be updated synchronously with changes in the situation. The platform will quickly degenerate into a state where manual checks are required for each step. Especially in vocational and technical training scenarios, the same workstation may be used repeatedly by different students, the same reason code may appear repeatedly in multiple tasks, and the evidence loop failure of the same work step may continue to occur. Therefore, step four should not only output the results and the scope of approval, but also write back the restriction information and failure sources in the certificate to the competency profile registry and work step evidence template. This allows the platform to gradually tighten unreliable paths and retain stable paths during repeated runs, and terminate the process or re-trigger the previous steps at the appropriate time.

[0145] The scoring engine, release interface, and write-back controller jointly receive the conformance certificate. Based on the comparable patterns and step-level restriction set in the certificate, the scoring engine decides whether to output a standardized score, a partial score with restriction markers, or only training feedback. The release interface sends the target execution description to the simulator or physical workstation according to the release scope in the certificate, or prevents it from entering the physical side. The write-back controller parses the reason code, high-priority restriction markers, and recoverable status recorded in the certificate to determine whether it should be written back to the capability profile registry in step one or to the step evidence template in step three.

[0146] To avoid terminological confusion caused by the write-back process itself, the platform server always executes the process in three stages: cause code - target object - write-back action. First, it identifies whether the cause code belongs to the capability deficiency category or the evidence template deficiency category. Then, it determines whether the target object is the capability profile registry or the process evidence template. Finally, it performs re-probing, strengthens the required evidence slots, or tightens the alternative evidence slots. This closed-loop feedback does not change the original terminology system but continuously updates along the original object.

[0147] In a preferred embodiment, the write-back controller does not write back for all exceptions, but first calculates the write-back directionality:

[0148]

[0149] In the formula: write-back direction : Indicates whether the write-back action corresponding to a certain reason code should be triggered immediately, with a value range of . A higher value indicates a greater need to modify the preceding object through a write-back; the frequency of recurrence value... : Indicates the degree of repetition of the same reason code in consecutive tasks, with a value range of . When the same reason code appears repeatedly, the recurrence frequency value is... The purpose of elevation is to distinguish between occasional problems and persistent problems;

[0150] Closed-loop blocking value : Indicates the degree of influence of this reason code on the effectiveness of the work step evidence closure or capability profile, with a value range of . When this reason code directly causes the process step to fail to form a closed loop or the capability item to fail, the closed loop blocking value is... Elevating the process highlights the real source of the problem that is blocking the flow.

[0151] This formula allows the write-back controller to consider both whether the cause code recurs and whether it truly blocks the loop closure. If a cause code appears occasionally but does not block the loop closure, the write-back direction... Keep it low; if a certain reason code appears repeatedly and continuously blocks the loop closure, then write back the pointer degree. Upon raising the level, the write-back controller initiates the corresponding write-back action. Therefore, the platform will not frequently reset previous objects due to occasional noise, nor will it ignore persistent problems.

[0152] In one specific embodiment, a gripping station repeatedly exhibited the cause code "gripper closure position missing," while the normalized execution traces showed that both the robot trajectory segment and the workpiece position change had occurred. After writing back to the controller, the recurrence frequency of this cause code was determined. The value continuously increases, and because the clamp closure position is a mandatory evidence slot in the process evidence template, the closed-loop blocking value... It is also at a high level. Therefore, the write-back pointer degree is... Upon being pushed up, the write-back controller directs the cause code to the evidence template in step three, tightens the range of alternative evidence slots, and requires the edge acquisition gateway to prioritize checking the write-back action of the gripper driver communication link, rather than directly modifying the grade result. The next time the student runs this step again, the platform will first perform acquisition along the updated evidence template, thereby changing the evidence collection link before the on-site action occurs.

[0153] Furthermore, after the scoring engine, release interface, and write-back controller have all completed the current round of processing, the process does not automatically end. Instead, the exit controller determines whether to terminate, wait for manual review, or re-trigger the previous steps. For example, if the consistency certificate indicates that the entire task is fully comparable and there are no recoverable limitations, the exit controller terminates the current round of processing and writes the consistency certificate, score result, and release result to the corresponding task record; if the consistency certificate indicates that the entire task is only usable for training and the relevant reason code does not have a recoverable state, the exit controller stops the automatic chain and issues a manual review prompt; if the consistency certificate indicates that a recoverable state exists, the exit controller triggers step one or step three depending on the reason type.

[0154] When the cause falls under the category of capability profiling, the active probing subunit in step one is triggered to re-verify the corresponding capability item; when the cause falls under the category of evidence closure, the process evidence template generator in step three is triggered to readjust the required evidence slots and alternative evidence slots. This exit control is not a simple termination or retry, but rather a targeted re-trigger that maintains a one-to-one correspondence with the preceding object. For example, the exit controller stores the task record index, certificate record index, and write-back record index in the platform server, all three referencing the same task version identifier. After the scoring engine completes its output, it first writes to the task record index; after the release interface completes its execution, it writes to the certificate record index; and finally, the write-back controller writes to the write-back record index. If any of these indexes fails to write, the exit controller places the process in a waiting state instead of terminating directly, allowing subsequent manual review to pinpoint the failed step.

[0155] As a supplement: if a declared capability exists and the probe passes, fill in the verified capability item; if a declared capability is not declared and the probe fails, fill in the set of capabilities that are declared non-existent and not declared, which cannot be included in the comparable scoring chain; if a declared capability is not declared and the probe passes, fill in the set of newly detected capability items, and indicate that the source is probe.

[0156] In a generalized implementation, if the software version of a workstation is updated, and the edge acquisition gateway finds that the capability profile version identifier is inconsistent with the current version when reconnecting, the exit controller directly invalidates the existing consistency certificate and triggers the active probing subunit in step one to reread the declared capabilities and reconstruct the target environment capability profile. If the version change does not affect the execution capability but causes a change in the time stamping method of the video acquisition unit, the exit controller does not restart the entire process, but only triggers step three to regenerate the work step evidence template and refresh the synchronization quality score. The above differentiated exit control ensures that each closed loop converges to the appropriate level, neither expanding small problems into a full-chain reconstruction nor compressing large problems into local repairs.

[0157] In practice, the reason code in the conformance certificate is transformed into a directional write-back pointer. This allows the platform to address recurring issues by resolving them through targeted corrections of preceding objects, rather than remaining at the result display level. Process exit and re-trigger conditions are explicitly bound to two preceding objects: the capability profile registry and the work step evidence template. This ensures that the closed-loop write-back truly returns to the technical objects that can change subsequent results, rather than remaining at the manual annotation level. Therefore, step four not only outputs results but also continuously tightens the preceding chain.

[0158] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0159] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and units described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0160] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0161] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0162] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. An AI digital twin robot training platform for vocational and technical training, characterized by: include, Receive course tasks and target workstation environment information, break down the course tasks into multiple work steps, and further break down each work step into semantic atoms containing course semantics, process semantics, control semantics and safety semantics, and construct a task semantic contract; Construct a target environment capability profile that includes execution capabilities, observability capabilities, time capabilities, and security capabilities; Based on the task semantic contract and the target environment capability profile, execution mapping and authentication mapping are performed on each semantic atom to generate a target execution description and a biaxial loss matrix containing execution preservation degree and authentication preservation degree; based on the biaxial loss matrix, a process step evidence template is generated and standardized execution traces are collected; A consistency certificate is generated based on the biaxial loss matrix, step evidence templates, and standardized execution traces. The comparable patterns and release range are determined based on the consistency certificate.

2. The AI ​​digital twin robot training platform according to claim 1, characterized in that: The construction of the task semantic contract includes: extracting the process step boundaries based on the course title, process description and scoring rules, and establishing semantic atoms corresponding to course semantics, process semantics, control semantics and safety semantics for each process step, so that each semantic atom is associated with preconditions, action content, post-state, time window and evidence slot.

3. The AI ​​digital twin robot training platform according to claim 2, characterized in that: The construction of the target environment capability profile includes: reading the controller configuration, simulator interface list and safety status mapping of the target workstation to form a set of declared capabilities, and performing active detection on the action primitive calling capability, trajectory and input / output observability capability, alarm and safety status readability capability and timestamp source stability capability to form a set of verified capabilities.

4. The AI ​​digital twin robot training platform according to claim 3, characterized in that: When there is a conflict between the declared capability set and the verified capability set, the target environment capability profile is corrected with the verified capability set, and a capability profile version identifier is written to the corrected target environment capability profile. When critical security conditions are not visible, the target environment capability profile is limited to the simulation training path only.

5. The AI ​​digital twin robot training platform according to claim 4, characterized in that: The execution mapping includes matching the objects, actions, preconditions, poststates, and security boundaries of each semantic atom, and writing them into the mapping type corresponding to the native mapping, split mapping, replacement mapping, and failed mapping; The authentication mapping involves channel matching of the evidence slots corresponding to each semantic atom and generating cause codes and recoverable tags in the biaxial loss matrix.

6. The AI ​​digital twin robot training platform according to claim 5, characterized in that: When the security semantics corresponding to a certain semantic atom are not fully preserved, a high-priority restriction flag is written into the dual-axis loss matrix, and the high-priority restriction flag is used as the basis for determining the range of access when generating a consistency certificate in the future; when a certain semantic atom is executable but not authenticated, it is restricted to entering only the training path.

7. The AI ​​digital twin robot training platform according to claim 6, characterized in that: The work step evidence template includes mandatory evidence slots, alternative evidence slots, optional evidence slots, time alignment rules, missing data handling rules, and degradation rules corresponding to each work step, and binds the action segment, precondition segment, and poststate segment in the target execution description to the corresponding evidence slots.

8. The AI ​​digital twin robot training platform according to claim 7, characterized in that: By using an edge acquisition gateway coupled to the robot controller, programmable logic controller, video acquisition unit and sensors, trajectory data, input / output logs, alarm logs, operation events and video data are collected according to the work step evidence template, and the acquisition results are written with work step number, semantic atom number, channel source, event type, time stamp and sequence number.

9. The AI ​​digital twin robot training platform according to claim 8, characterized in that: When the target controller cannot provide a structured program tree, an alternative control structure is constructed based on the program text, trajectory data, input / output logs, alarm logs, and operation events. After matching the action keywords in the program text with the time sequence of the trajectory data and input / output logs, a normalized execution trace corresponding to the semantic atom is generated.

10. The AI ​​digital twin robot training platform according to claim 9, characterized in that: The synchronization quality score is obtained by performing sequential verification and cross-source fitting verification on the precondition events, action events and poststate events in the normalized execution trace under a unified time axis. When either the required evidence integrity rate or the synchronization quality score does not meet the corresponding preset conditions, the current step is limited to the corresponding mode in the partially comparable mode or the training mode according to the degradation rules.

11. The AI ​​digital twin robot training platform according to claim 10, characterized in that: The conformance certificate includes at least the task version identifier, capability profile version identifier, evidence template version identifier, total execution retention score, total authentication retention score, synchronization quality score, comparable mode, release range, and reason code summary. The scoring engine and release interface perform subsequent processing based on the conformance certificate and standardized execution traces.

12. The AI ​​digital twin robot training platform according to claim 11, characterized in that: When the conformance certificate is fully comparable, output standardized scores and open restricted entity trial paths; when the conformance certificate is partially comparable, output only the partial scores with restriction tags. When the conformance certificate is characterized as training-only mode, only training feedback is output, and entry into the physical workstation release path is blocked.

13. The AI ​​digital twin robot training platform according to claim 12, characterized in that: If the same reason code appears repeatedly in consecutive tasks, the reason code will be written back to the corresponding object in the capability profile registry and the work step evidence template according to the corresponding type of capability deficiency and evidence deficiency. If the target workstation version change causes the capability profile version identifier to be inconsistent with the current version, the existing consistency certificate will be invalidated, and the construction of the task semantic contract and the construction of the target environment capability profile will be retried.

Images