A method, device and equipment for grading determination of success of robot task execution, and a storage medium

Abstract

This invention discloses a method, apparatus, device, and storage medium for classifying the success of robot task execution. The method, through a task type identification module, classifies tasks as autonomous, immediately verifiable, or delayed-experience tasks based on whether the instruction includes a service recipient and whether the execution result can be objectively verified within a preset time window. Autonomous tasks are judged according to physical criteria; immediately verifiable tasks are judged by collecting the service recipient's immediate confirmation signal after the physical criteria are met; delayed-experience tasks are set to pending judgment, and an evaluation score is calculated upon receiving the service recipient's evaluation signal, classifying them as successful, failed, or questionable. A preference anchor writing module solidifies the judgment results into positive or negative preference anchors, or temporarily leaves questionable results unsolvable. This invention prevents the robot from self-declaring success, returning the judgment authority for subjective experiential tasks to the human, and avoiding the solidification of erroneous experiences.

Description

Technical Field

[0001] This invention relates to the fields of robot control and artificial intelligence technology, and specifically to a method, apparatus, electronic device, and computer-readable storage medium for classifying the success of robot task execution. Background Technology

[0002] Current service robots and embodied intelligent robots generally equate "action completed" with "execution successful" when executing user commands. That is, once the robot completes the physical action corresponding to the command, the task is considered successfully completed. This approach has the following problems:

[0003] First, physical completion does not equate to true success. "Completing" a task does not equate to "doing it correctly." For example, "pouring a glass of water" objectively confirms success once it's done; however, the success of "cooking a meal" is subjective and delayed, requiring feedback from the person being served. Current technology cannot distinguish between these two types of tasks, treating "delayed experience" tasks the same as "instantly verifiable" tasks, directly determining success based on "action completion."

[0004] Secondly, the machine self-declares success, lacking a judgment step. Existing robots assume success upon completion, without any "judgment" process. For instructions whose literal meaning does not match the actual request (for example, a user says "Hey, let's keep going," which literally means "keep going," but the actual request might be "want to rest"), the robot will "do the literal thing correctly and the request incorrectly" without realizing it.

[0005] Third, erroneous results become entrenched, leading to a vicious cycle. If unresolved execution results are treated as "successful experiences" and solidified, erroneous experiences will continue to accumulate, undermining the robot's reliability.

[0006] In existing technologies, some solutions have attempted to address the problem of "how to determine execution success." For example, CN115060273A discloses a method to determine robot execution success by ensuring that both the "time difference and path distance difference are less than a threshold." However, this method is only applicable to the physical layer determination of navigation tasks and does not address task classification, delayed experience tasks, or returning the determination authority to humans. KR1020260131753A discloses a method to determine task completion by generating a task completion detection function using a large language model, but this method still relies on the machine itself for determination and does not address "returning subjective tasks to humans for determination" or "evaluation feedback." In addition, some solutions (such as CN119658692A) first eliminate instruction ambiguity at the task environment level and call the user feature memory module for preference filtering, and then collect user feedback on task success or failure during execution to update the memory module. However, its ambiguity elimination is aimed at the environmental dissolution of specific references (such as dissolving "beverage" into a specific category existing in the environment). It neither quantifies "whether the instruction can be uniquely interpreted" into a calculable score and triggers follow-up questions when it falls below a threshold, nor does it distinguish the verification method of the task. The success or failure of the task is determined by the machine based on preset rules or model outputs, without the step of "suspending subjective experience-type tasks that cannot be objectively verified in real time, handing them over to the service recipient for evaluation, and solidifying preference anchors according to three levels: directionality and doubt".

[0007] In summary, existing technologies have not solved the technical problem of "how to classify and determine whether the execution was truly successful according to task type and reliably feed back human evaluations". Summary of the Invention

[0008] I. Technical problems to be solved

[0009] In existing technologies, objectively verifiable tasks can be judged using physical criteria, but subjective experience-based tasks cannot be judged by physical quantities. If all tasks are judged automatically by the machine without distinction, subjective tasks may be misjudged as successful, and erroneous experiences will be solidified. The technical problem this invention aims to solve is how to differentiate task types, classify and judge whether the "execution was truly successful" for various types of tasks where there is no unified judgment standard, and reliably feed back the judgment results, thereby avoiding false successes of "physical completion but failure to meet real requirements" and the vicious cycle of "erroneous experiences being solidified."

[0010] II. Technical Solution

[0011] To address the aforementioned technical problems, this invention provides a hierarchical method for determining the success of robot task execution. Its core idea can be summarized by two deterministic principles: execution must be definite (if the intent is not unique, execution will not proceed; the unique intent must be determined first); success must be determined (execution cannot be self-declared successful; it must undergo hierarchical determination).

[0012] The method includes the following steps:

[0013] Step S1: Task Type Identification. The task type identification module determines the current task as an "autonomous task", "instantly verifiable task", or "delayed experience task" based on whether the instruction contains a service object and whether the execution result can be objectively verified within a preset time window.

[0014] Specifically, the determination is made using a first feature and a second feature: the first feature is whether the instruction semantics contain a service object or an interaction object; the second feature is whether the execution result can be objectively observed by sensors or environmental feedback within a preset time window. When the first feature is "no", it is determined to be an autonomous task; when the first feature is "yes" and the second feature is "yes", it is determined to be an immediately verifiable task; when the first feature is "yes" and the second feature is "no", it is determined to be a delayed experience task.

[0015] Step S2: Classification and determination of autonomous tasks. The execution result signal (such as force feedback signal, pose signal, position switch signal, sensor reading or one or more) is collected through the physical execution verification module. The execution result signal is compared with a preset reference, and the consistency score is calculated. When the consistency score is not lower than the preset physical threshold, the execution is determined to be successful.

[0016] Step S2′: Grading and determination of immediately verifiable tasks. After the physical execution verification module determines that the consistency between the execution result and the preset reference is not lower than the preset physical threshold, the immediate confirmation signal of the person being served (such as verbal confirmation, nodding, etc.) is collected, and the confirmation score is calculated by the evaluation and analysis module; when the confirmation score is not lower than the preset confirmation threshold, the execution is determined to be successful.

[0017] Step S3: Suspension and Manual Judgment of Delayed Experience Tasks. For tasks judged as delayed experience tasks, the task status is set to "Pending Judgment," without self-declaring success. After receiving the evaluation signal from the service recipient, the evaluation score is calculated through the evaluation parsing module: explicit satisfaction is assigned the first positive value (e.g., +1.0), explicit dissatisfaction is assigned the first negative value (e.g., -1.0), and neutral or no explicit expression is assigned zero. Execution is considered successful when the evaluation score is not lower than the preset satisfaction threshold; execution is considered unsuccessful when it is not higher than the preset dissatisfaction threshold; otherwise, it is considered questionable.

[0018] The evaluation score is calculated by the evaluation parsing module based on the semantic polarity words in the evaluation signal. An example of the parsing rules is as follows: a first positive value is assigned to words that explicitly satisfy (e.g., "good," "satisfied," "not bad," "okay," "just right"); a first negative value is assigned to combinations of negative and satisfying words (e.g., "bad," "unsatisfied," "too bland," "too salty"); and zero is assigned to only neutral responses (e.g., "um," "oh"). The semantic polarity classification is a well-known text sentiment analysis method in the field, and this invention does not limit its specific implementation.

[0019] Step S4: Evaluate the reflux and solidify the anchor points. The judgment result is solidified into a preference anchor point through the preference anchor point writing module: a successful judgment is solidified as a positive preference anchor point, a failed judgment as a negative preference anchor point, and a judgment in doubt is not solidified for the time being. Each preference anchor point includes a source identifier and a confidence weight. The confidence weight is determined by the clarity of the evaluation signal and decays over time.

[0020] Furthermore, a preliminary auxiliary step may be included before step S1:

[0021] Step S0: Instruction Parsing and Ambiguity Determination. The instruction parsing module calculates a uniqueness score that allows the instruction to be uniquely interpreted. When the uniqueness score is lower than a preset uniqueness threshold, the instruction is determined to be ambiguous, triggering a burden reduction inquiry. The user is prompted with follow-up questions by narrowing down the options until the instruction is uniquely determined or a preset maximum number of follow-up questions is reached. Task type identification is only performed after the instruction is uniquely determined. The uniqueness score is a weighted composite of four sub-features: polarity conflict score, degree adverb polysemy score, negation hesitation score, and multimodal contradiction score.

[0022] The determination rules for the four sub-features are as follows: the polarity conflict score is the conflict measure taken when affirmative appeal words and negative contexts co-occur in the instruction, for example, a high value is taken when "want" and negative contexts such as "don't / no" co-occur; the degree adverb polysemy score is the degree uncertainty measure caused by polysemous degree adverbs (such as "slightly", "a little", "very") in the instruction, and a high value is taken when polysemous adverbs are hit; the negation hesitation score is the hesitation measure taken when negation words and hesitation words (such as "not..." "maybe not") co-occur in the instruction; the multimodal contradiction score is the inconsistency between the linguistic signal and non-linguistic signal output by the multimodal consistency model. The uniqueness score is synthesized according to U = 1 - max(polarity conflict score, degree adverb polysemy score, negation hesitation score, multimodal contradiction score), and the scores of the four sub-features are all normalized to [0, 1], with larger values ​​indicating a greater likelihood of ambiguity.

[0023] Furthermore, it may also include instruction completeness checks: calculating the missing score of parameters required for instruction execution; when the missing score is not lower than a preset missing threshold, or when any security level parameter is missing, it is determined that the instruction information is insufficient, and a prompt is made to fill in the missing parameters; wherein the security level parameters are triggered by a single vote.

[0024] Furthermore, it may also include contextual conflict handling: when the conflict score between the literal appeal of the instruction and the tone polarity is detected to be no less than a preset conflict threshold, it is determined to be a contextual conflict task. The task is then solicited through a reduced-burden solicitation sentence and executed after the instruction is redefined based on the user's confirmed wishes.

[0025] III. Beneficial Effects

[0026] 1. Avoid pseudo-success when "physical completion fails to meet the needs": By classifying tasks into "immediately verifiable" and "delayed experience" types, subjective experience tasks are no longer judged as successful based on "action completion". Instead, they are suspended and await the feedback from the person being served, thus fundamentally avoiding "doing the literal thing right but fulfilling the needs wrong".

[0027] 2. The power of judgment is returned to humans, and the machine does not declare success on its own: The machine only judges "objectively verifiable" (immediately verifiable), and returns "subjective, delayed, and uncertain" (delayed experience) to humans, making the boundaries of certainty clear.

[0028] 3. Reliable feedback of evaluations, becoming more accurate with use: Human evaluations are solidified into three levels: positive anchor point, negative anchor point, and questionable. This avoids the solidification of erroneous experiences and allows for the accumulation of preferences, reducing the need for repeated follow-up questions in subsequent tasks.

[0029] 4. Pre-interception of ambiguity and missing parameters: Before execution, uniqueness score and missing value score are checked to eliminate execution failures caused by "insufficient information" and "multiple interpretations".

[0030] 5. Each judgment method collaborates and divides labor according to task type, with each function being a prerequisite for the other, achieving technical effects that a single method cannot achieve independently: Task type identification divides tasks into autonomous, immediately verifiable, and delayed experience types, thereby defining the boundary for the machine to self-declare success—only autonomous tasks without service recipients are judged using pure physical criteria, while for tasks with service recipients, after the physical criteria are passed, the recipient's immediate confirmation is required, or the task is suspended and awaits the recipient's response; if no prior classification is performed, subjective experience tasks will be misjudged as successful by physical criteria; if delayed experience tasks are not suspended and the machine self-declare success, the existing technology's "machine self-declaration of success" will be repeated; if the evaluation only has two values, success / failure, without an intermediate state of "doubt," then the fuzzy evaluation will be solidified into an erroneous preference. The aforementioned grading, physical criteria boundary delineation, suspension for judgment, and three-level solidification are interdependent and indispensable, working together to form a complete closed loop of "objective tasks assigned to machines, subjective tasks assigned to humans, and prevention of the solidification of erroneous experience"; the overall technical effect is indivisible - no single means, or even a simple superposition of any part of the means, can achieve this effect independently. Attached Figure Description

[0031] Figure 1 is a flowchart illustrating a method for classifying the success of robot task execution according to an embodiment of the present invention.

[0032] Figure 2 is a schematic diagram of the entire process of completing a "cook a meal" task according to an embodiment of the present invention.

[0033] In Figure 1, S1 represents the task type identification step, S2 represents the physical criterion determination step for autonomous tasks, S2′ represents the physical criterion plus immediate confirmation step for immediately verifiable tasks, S3 represents the suspension pending determination step for delayed experience tasks, and S4 represents the evaluation feedback solidification preference anchor point step; L0, L1, and L2 represent autonomous tasks, immediately verifiable tasks, and delayed experience tasks, respectively. Detailed Implementation

[0034] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings. The following embodiments are used to illustrate this invention, but are not intended to limit the scope of the invention.

[0035] Example 1: Module Structure

[0036] The grading determination device of this embodiment can be deployed in the control system of service robots, humanoid robots, quadruped robots, wheeled or tracked mobile robots, bionic robots, or industrial robotic arms, and includes the following modules:

[0037] Task type identification module: Based on the first feature (whether it contains a service object) and the second feature (whether it can be objectively verified within a preset time window), the module determines whether the task type is autonomous, immediately verifiable, or delayed experience.

[0038] Instruction parsing module: Calculates uniqueness score and missing score, and performs ambiguity determination and instruction completeness check;

[0039] Physical execution verification module: Collects execution result signals and calculates consistency scores;

[0040] Evaluation and analysis module: Analyzes the immediate confirmation signal and evaluation signal of the service recipient, and calculates the confirmation score and evaluation score respectively;

[0041] Preference Anchor Point Writing Module: Solidifies the judgment result into a positive preference anchor point, a negative preference anchor point, or leaves it unsolidified;

[0042] The burden reduction consultation module generates burden reduction consultation sentences and performs follow-up questions by narrowing the options. The burden reduction consultation sentences include one or more of the following: packaging the request as an optional option, narrowing the open question to a choice between two options, packaging the refusal as a default to continue, and replacing the direct judgment of the user's state with first person or environmental attribution.

[0043] The multimodal consistency model is used to calculate the multimodal inconsistency score. Its inputs are feature vectors of linguistic and non-linguistic signals. The linguistic feature vectors are obtained from user speech through speech recognition and text encoding. The non-linguistic feature vectors are obtained from one or more observable features such as gait, speech rate, volume, pauses, and facial expressions through feature extraction. Its output is the inconsistency score corresponding to the consistency comparison between the linguistic and non-linguistic signals, with a value ranging from 0 to 1. A larger value indicates a greater inconsistency between the linguistic and non-linguistic signals. The multimodal consistency model includes a linguistic encoding branch, a non-linguistic encoding branch, and a similarity comparison layer. The linguistic and non-linguistic encoding branches output feature vectors of equal length. The similarity comparison layer calculates the similarity between the two vectors and normalizes it to obtain the inconsistency score. The multimodal consistency model can use a pre-trained multimodal alignment model (i.e., a readily available model known in the art) and fine-tune it for robot task interaction scenarios. This invention does not limit its specific network structure.

[0044] The signal acquisition of each module follows privacy boundaries: only observable and non-intrusive signals (tone of voice, wording, degree adverbs, currently observable facial expressions / gait / speech rate, on-site task interaction habits) are collected, and no identity attribute privacy information (personality traits, medical history, unauthorized historical records, etc.) is collected.

[0045] Example 2: Optimal Threshold Value Selection

[0046] For ease of implementation, this embodiment provides preferred values ​​for each threshold (all of which can be adjusted within a reasonable range, and are not limited to these):

[0047] The preset missing threshold is 0.3, and the parameter risk weights are: security level 3.0, core experience level 2.0, and general level 1.0.

[0048] The maximum number of follow-up questions is 3;

[0049] The preset uniqueness threshold is 0.7; the uniqueness score is synthesized according to U = 1 - max(polarity conflict score, degree adverb polysemy score, negation hesitation score, multimodal contradiction score), and all four sub-feature scores are normalized to [0, 1]. Among them, the polarity conflict score, degree adverb polysemy score, and negation hesitation score are 1.0 when the corresponding tag is hit, and 0 when it is not hit; the multimodal contradiction score is a continuous value from 0 to 1 output by the multimodal consistency model.

[0050] The preset conflict threshold is 0.5; the conflict score F is the degree of conflict between the literal appeal and the tone polarity. When a tone word (such as "hey" or "never mind") is detected, which makes the tone polarity opposite to the semantic polarity of the literal instruction, F takes a high value close to 1. When the tone is consistent with the literal meaning, F takes a value close to 0. Specifically, it can be calculated by the polarity confidence output by the tone polarity classifier and the cosine similarity between the semantic polarity of the literal instruction and F = 1 - cos(literal appeal polarity, tone polarity).

[0051] The objective verification time window is set at 30 seconds;

[0052] The preset physical threshold (consistency) is set to 0.9;

[0053] The preset satisfaction threshold is set to +0.5, and the preset dissatisfaction threshold is set to -0.5.

[0054] Evaluation scores are set as follows: Clearly satisfied +1.0, Clearly dissatisfied -1.0, Neutral 0.

[0055] Example 3: A complete workflow for a "cook a meal" task (corresponding to Figure 2)

[0056] This embodiment uses an example of an elderly person saying to a robot, "Just get me some food," to illustrate the complete execution process of the grading determination method.

[0057] Step 1 (Instruction Completeness Check): The robot parses the instruction as a "cooking" task and checks the essential parameter cards for "cooking" (number of people / portions, flavor, dietary restrictions, dishes). It finds that none of the four items have default values ​​or observed habits, and the missing value score is not lower than the preset missing value threshold. It determines that the instruction information is insufficient and initiates active questioning.

[0058] Step 2 (Proactive Questioning to Complete the Questions): The robot narrows down the questions using a "reduced burden inquiry + two-choice" approach—"Is this meal for one person or for two?" The elderly person answers "one person"; "Would you like something light or something with rice?" The elderly person answers "light." The portion size and taste preferences are thus uniquely determined.

[0059] Step 3 (Ambiguity Resolution): "Lighter" still has multiple interpretations (porridge / noodles / steamed vegetables). The uniqueness score is lower than the preset uniqueness threshold, so the robot continues to narrow it down—"Would you like a bowl of vegetable porridge or a bowl of clear soup noodles?" The old man answers, "I'll make vegetable porridge." The dish is now uniquely determined.

[0060] Step 4 (Execution): With all parameters complete, intent unique, and unambiguous, the robot begins cooking.

[0061] Step 5 (Suspended for judgment): The porridge is cooked, but the robot does not announce success. The task status is set to "pending judgment" (corresponding to step S3).

[0062] Step 6 (Reducing Burden Consultation): After the elderly person finishes their meal, the robot asks in a "Reducing Burden Consultation" format—"Was the porridge salty enough? Should I make it thicker for you next time?"

[0063] Step 7 (Evaluation Feedback and Anchor Point Consolidation): The elderly person answers, "It's a bit bland; add more salt next time." The evaluation score is the first negative value, below the preset dissatisfaction threshold, indicating a failure. "Vegetable porridge should be slightly salty and thicker" is solidified as a negative preference anchor point (avoiding the "bland" approach next time and adjusting the flavor). Conversely, if the elderly person expresses satisfaction, it is solidified as a positive preference anchor point. The next time the robot asks for "just something to eat," it can directly default to "a slightly salty and thick vegetable porridge" based on the preference anchor point, reducing the need for repeated follow-up questions.

[0064] Example 4: Contextual Conflict Task (L3)

[0065] The user says to the robot, "Hey, let's keep going." The literal instruction is "keep going," but the interjection "hey" makes the tone polarity opposite to the literal instruction. If the conflict score is not lower than the preset conflict threshold, it is judged as a contextual conflict task. The robot does not directly execute "keep going," but instead asks in a "burden-reducing inquiry" style—"There's a pavilion ahead, would you like to sit down for a bit?"—giving the choice back to the user. The robot then re-determines the instruction based on the user's confirmed willingness before executing it.

[0066] Example 5: Distinguishing between autonomous, immediately verifiable, and delayed experience types

[0067] "Pour a glass of water": Includes the service recipient (first characteristic is yes), the water pouring result can be objectively verified within a preset time window (second characteristic is yes), it is judged as an instant verifiable task, and after the physical criteria are passed, it is supplemented by real-time human confirmation.

[0068] "Moving goods to the shelf" and "Tightening screws": No service object (first characteristic is no), so it is judged as an autonomous task and determined only by physical criteria.

[0069] "Cook a meal" or "Tidien a room to a satisfactory standard": If the service recipient is present (yes) but the result cannot be objectively verified within the preset time window (no), it is judged as a delayed experience task and is suspended pending the recipient's response.

[0070] Example 6: Experimental Verification of Determining the Attribution of Rights

[0071] To verify the effectiveness of the technology that "physical criteria are insufficient to replace human judgment for subjective experience-based tasks," a live-person labeled experiment was conducted on delayed experience-based tasks.

[0072] Four annotators from the same group independently (without prior consultation) categorized the two sets of tasks into three levels: "success / failure / questionable." The first set consisted of 41 delayed-experience tasks; the second set comprised 36 tasks where "the action was indeed completed, but it's unclear whether it was successful" (i.e., more subjective delayed-experience tasks). A paired comparison was performed based on the common number of 36 tasks in both sets. The results showed:

[0073] For the same batch of tasks, the judgments of different annotators were systematically inconsistent, and this inconsistency increased as the subjectivity of the task increased—the disagreement rate (the proportion of a question where no single majority opinion was formed) rose from 83.3% to 88.9%, the consistency coefficient κ fell from +0.17 to +0.04, and the proportion of questions with a "questionable" judgment rose from 58.3% to 88.9%; looking at each person individually, the proportion of "questionable" judgments increased for 3 out of 4 annotators, while it decreased for 1, and the consistency rate between that person and the majority opinion of the other annotators was only about 40%.

[0074] The above results indicate that for subjective experience-based tasks that cannot be objectively verified by physical quantities, people's judgments of "success" are inconsistent and unstable, and the uncertainty itself varies from person to person. If the machine declares success automatically with a fixed threshold, it will inevitably contradict the actual feelings of some users. This is precisely the technical basis for this application's "suspending and handing over the decision to the user" for delayed experience-based tasks—execution can be deterministic, but the right to determine "success" must be returned to the person, and that decision must be solidified as the user's personal preference anchor.

[0075] Example 7: Preference Anchor-Driven Recommendation and Customization

[0076] To avoid repeatedly asking users who have already accumulated preference anchors and thus degrading their user experience, this embodiment provides a recommendation and customization mechanism based on preference anchors.

[0077] When a user issues a task instruction of the same type as a previous task, the task type identification module determines the task type. If the user already has a corresponding preference anchor, it does not repeatedly ask for the same instruction. Instead, it generates a recommended default value (i.e., a default parameter determined based on the user's positive or negative preference anchor) based on the user's positive or negative preference anchor. For example, if a user previously indicated "too bland" and this preference anchor was solidified as "too salty and thick," and then says "just something to eat," the robot directly recommends the default value "a salty and thick vegetable porridge."

[0078] The recommended default values ​​can be presented through a simplified query format, allowing users to confirm and execute with a single click, or customize and modify individual parameters (such as taste, portion size, and dietary restrictions). The user's confirmation or modification results serve as a new evaluation feedback, updating the corresponding preference anchor points (successfully fixed as positive preference anchor points, unsuccessfully fixed as negative preference anchor points, and those in doubt are not fixed for now).

[0079] Therefore, for tasks executed for the first time, follow-up questions are asked as needed (see Example 3), while for repetitive tasks with accumulated preference anchors, direct recommendations and one-click confirmation are provided. The number of follow-up questions is reduced from multiple times to one or even zero times, avoiding repeated follow-up questions that would degrade the user experience. At the same time, preference anchors are continuously updated and become more accurate with use.

[0080] The above embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made by those skilled in the art within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for classifying the success of robot task execution, characterized in that, include: The task type identification module determines the current task as an autonomous task, an immediately verifiable task, or a delayed experience task based on whether the instruction contains a service recipient and whether the execution result can be objectively verified within a preset time window. For tasks determined to be autonomous tasks, the physical execution verification module collects the execution result signal and compares it with a preset reference. If the consistency is not lower than a preset physical threshold, the execution is considered successful. For tasks determined to be immediately verifiable tasks, after the physical execution verification module determines that the consistency between the execution result and the preset reference is not lower than the preset physical threshold, the immediate confirmation signal of the service recipient is collected, and the immediate confirmation is calculated by the evaluation and analysis module. The confirmation score of the signal is used to determine successful execution when the confirmation score is not lower than a preset confirmation threshold. For tasks identified as delayed experience tasks, the task status is set to pending judgment. After receiving the evaluation signal from the service recipient, the evaluation parsing module calculates the evaluation score of the evaluation signal based on the semantic polarity words in the evaluation signal. If the evaluation score is not lower than a preset satisfaction threshold, the execution is determined to be successful. If the evaluation score is not higher than a preset dissatisfaction threshold, the execution is determined to be unsatisfactory. Otherwise, it is determined to be questionable. The judgment result is solidified into a preference anchor point through the preference anchor point writing module: successful judgments are solidified into positive preference anchor points, unsuccessful judgments are solidified into negative preference anchor points, and questionable judgments are not solidified temporarily.

2. The method according to claim 1, characterized in that, Before the task type identification, the method further includes: parsing the instruction through the instruction parsing module and calculating the uniqueness score of the instruction; when the uniqueness score is lower than a preset uniqueness threshold, it is determined that the instruction is ambiguous, triggering a burden reduction consultation, and asking the user follow-up questions by narrowing the options until the instruction is uniquely determined or the preset maximum number of follow-up questions is reached; the task type identification is only performed after the instruction is uniquely determined.

3. The method according to claim 2, characterized in that, The instruction parsing module also calculates the missing score of the parameters required for instruction execution; when the missing score is not lower than the preset missing threshold, or when any security level parameter is missing, it is determined that the instruction information is insufficient, and the module actively asks for the missing parameters; wherein the security level parameters are triggered by a single vote, and a question is forced regardless of whether the missing score reaches the threshold.

4. The method according to claim 2, characterized in that, The uniqueness score is a weighted composite of four sub-features: polarity conflict score, degree adverb polysemy score, negation hesitation score, and multimodal contradiction score. Specifically, the polarity conflict score is a conflict measure calculated when affirmative appeal words co-occur with negative context in the instruction; the degree adverb polysemy score is a degree uncertainty measure caused by polysemous degree adverbs in the instruction; and the negation hesitation score is a hesitation measure calculated when negative and hesitant words co-occur in the instruction. The multimodal contradiction score is the inconsistency degree output by a multimodal consistency model after comparing the feature vectors of the linguistic signal and the feature vectors of the non-linguistic signal. The inconsistency value is between 0 and 1. The multimodal consistency model includes a language encoding branch, a non-language encoding branch, and a similarity comparison layer. The language encoding branch and the non-language encoding branch output feature vectors of equal length. The similarity comparison layer calculates the similarity between the two vectors and normalizes it to obtain the inconsistency value. The multimodal consistency model is obtained by using a pre-trained multimodal alignment model and fine-tuning it for robot task interaction scenarios. The non-language signals are one or more of the observable gait, speech rate, volume, pauses, and facial expressions, and no identity attribute privacy information is collected.

5. The method according to claim 1, characterized in that, The task type identification module determines the task type based on a first feature and a second feature: the first feature is whether the instruction semantics contain a service object or an interaction object; the second feature is whether the execution result can be objectively observed by sensors or environmental feedback within a preset time window; when the first feature is not present, it is determined to be the autonomous task; when the first feature is present and the second feature is present, it is determined to be the immediately verifiable task; when the first feature is present and the second feature is not present, it is determined to be the delayed experience task.

6. The method according to claim 1, characterized in that, Also includes: When the conflict score between the literal appeal and the tone of the instruction is detected to be no less than a preset conflict threshold, it is determined to be a contextual conflict task. The task is then asked to reduce the burden of the user by asking a different question, and the instruction is re-determined and executed based on the user's confirmed wishes.

7. The method according to claim 1, characterized in that, The evaluation score is determined based on the direction of the evaluation signal: a positive value is assigned to explicit satisfaction, a negative value is assigned to explicit dissatisfaction, and zero is assigned to neutral or no explicit stance. The preference anchor is accompanied by a source identifier and a confidence weight, the confidence weight being determined by the explicitness of the evaluation signal and decaying over time.

8. The method according to claim 2 or 6, characterized in that, The proposed burden reduction consultation includes one or more of the following methods: packaging the request as an optional option, narrowing open-ended questions into two choices, packaging rejection as a default to continue, and replacing direct judgment of the user's state with first-person or environmental attribution.

9. The method according to claim 1, characterized in that, The robot is selected from one of the following: service robot, humanoid robot, quadruped robot, wheeled or tracked mobile robot, bionic robot, and industrial robotic arm; the method is applied to scenarios where the robot performs tasks for the person being served.

10. A graded determination device for successful robot task execution, characterized in that, include: The task type identification module is used to determine the current task as an autonomous task, an immediately verifiable task, or a delayed experience task based on whether the instruction contains a service object and whether the execution result can be objectively verified within a preset time window. The physical execution verification module is used to collect the execution result signal of the autonomous task, compare it with a preset reference, and determine that the execution is successful when the consistency is not lower than a preset physical threshold; and to trigger the collection of the immediate confirmation signal of the service recipient after determining that the consistency between the execution result and the preset reference is not lower than the preset physical threshold for the immediate verifiable task. The evaluation and analysis module is used to calculate the confirmation score of the instant confirmation signal for the instant verifiable task, and determine the execution success when the confirmation score is not lower than a preset confirmation threshold; and to calculate the evaluation score after receiving the evaluation signal from the service recipient for the delayed experience task, and determine the execution success, failure or doubt based on the comparison result of the evaluation score with a preset satisfaction threshold and a preset dissatisfaction threshold. The preference anchor writing module is used to solidify the judgment result into a positive preference anchor, a negative preference anchor, or to temporarily not solidify the doubtful result.

11. An electronic device, characterized in that, It includes a memory, a processor, and a computer program stored on the memory; when the processor executes the computer program, it implements the method as described in any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, It stores a computer program thereon, which, when executed by a processor, implements the method as described in any one of claims 1 to 9.

Images