CN-121973215-A - Robot task generation method, system and computer equipment

CN121973215ACN 121973215 ACN121973215 ACN 121973215ACN-121973215-A

Abstract

The embodiment of the disclosure provides a robot task generating method, a system and computer equipment, wherein the method comprises the steps of determining a task space tuple for describing a robot task, carrying out data sampling on preset task description items based on historical task data of corresponding description data of the preset task description items in the task space tuple to obtain a task description data set at least comprising the corresponding sampling data of the preset task description items, and providing the task description data set as constraint conditions for a proposal generator to generate a target task proposal. According to the method provided by the disclosure, task space tuples are predefined for the robot task proposal, and the forced proposal generator generates corresponding target task proposals based on the constraint of the task description data set through a set data sampling mechanism. The task proposal forcedly generated by the method explores uncovered long tail scenes, objects and skill combinations, and solves the problem of uneven distribution of the existing training data.

Inventors

WU KUN
ZHANG YIXUE
CHE ZHENGPING
TANG JIAN

Assignees

北京人形机器人创新中心有限公司

Dates

Publication Date: 20260505
Application Date: 20260204

Claims (14)

1. A robot task generation method, comprising: Determining a task space tuple for describing a robot task, wherein the task space tuple comprises a plurality of task description items; Based on historical task data of corresponding description data of preset task description items in the task space tuple, performing data sampling on the preset task description items to obtain a task description data set at least containing the corresponding sampling data of the preset task description items; And providing the task description data set as constraint conditions for a proposal generator to generate a target task proposal.
2. The method according to claim 1, wherein the performing data sampling on the preset task description item based on the historical task data of the corresponding description data of the preset task description item in the task space tuple to obtain a task description data set at least including the sampled data corresponding to the preset task description item includes: And based on the use frequency of the corresponding description data of the preset task description items in the task space tuple in the historical task data, carrying out data sampling on the preset task description items to obtain a task description data group at least comprising the corresponding sampling data of the preset task description items, wherein the use frequency is lower than the preset frequency.
3. The method of claim 1, wherein the preset task description items include a scene category set that characterizes a scene of use of the robot, a task object library that characterizes objects of robot interactions, and a task skill library that characterizes ways of robot interactions; based on the use frequency of the corresponding description data of the preset task description items in the task space tuple in the historical task data, carrying out data sampling on the preset task description items to obtain a task description data group at least comprising the corresponding sampling data of the preset task description items, wherein the method comprises the following steps: according to the frequency of executing tasks by the robot in each scene category in the scene category set, determining the scene category with the use frequency lower than a preset first threshold value as a target scene category; selecting task objects logically associated with the target scene category from the task object library based on the semantics of the target scene category to obtain a task object set; Based on the semantics of the target scene category, selecting task skills logically associated with the target scene category from the task skill library to obtain a task skill set; according to the historical interaction frequency of the robot and the task objects, task objects with the use frequency within a first preset low-frequency order range are screened from the task object set to obtain a candidate task object set; According to historical use frequency of the robot on task skills, task skills with use frequency within a second preset low-frequency order range are screened from the task skill set, and a candidate task skill set is obtained; And determining the target scene category, the candidate task object set and the candidate task skill set as the task description data set.
4. The method of claim 1, wherein providing the task description data set as constraints to a proposal generator to generate a target task proposal comprises: invoking a generating model, determining a prompt word based on the task description data set, and generating an initial task proposal for indicating the target robot to interact; calling a checking model to check the initial task proposal to obtain a checking result; invoking a disjunct optimizer to determine whether the initial task proposal needs to be corrected according to the retrieved history interaction record, the examination result and the initial task proposal; And when the correction is needed, generating a task proposal correction instruction, and correcting the initial task proposal to obtain a target task proposal.
5. The method of claim 1, wherein providing the task description data set as constraints to a proposal generator to generate a target task proposal comprises: invoking a generating model, determining a prompt word based on the task description data set, and generating an initial task proposal for indicating the target robot to interact; Taking the initial task proposal as the current task proposal, executing the following proposal examination step: Invoking an inspection model to inspect the current task proposal to obtain an inspection result; Invoking a disjunct optimizer to determine whether the current task proposal needs to be corrected according to the retrieved history interaction record, the current examination result and the current task proposal; under the condition that the current task proposal needs to be corrected, generating a task proposal correction instruction, correcting the current task proposal, and obtaining a corrected task proposal; And taking the corrected task proposal as a new current task proposal, and returning to execute the proposal examination step until the current task proposal is determined to be not required to be corrected, so as to obtain a target task proposal.
6. The method of claim 4 or 5, wherein the audit model includes at least one of a physical feasibility evaluator, a novelty evaluator, and a constraint consistency evaluator; The method comprises the following steps of calling a checking model to check the task proposal to be checked to obtain checking results, wherein the checking method comprises the following steps of: And respectively calling the evaluator, and inspecting the task proposal to be inspected from the corresponding evaluation dimension and respectively obtaining the corresponding inspection result, wherein the task proposal to be inspected comprises an initial task proposal and/or a current task proposal.
7. The method of claim 6, wherein the assessment dimensions of the physical feasibility assessor include a kinematic range, physical interaction logic and interaction capability requirements; In the case that the evaluator includes a physical feasibility evaluator, the respectively calling the evaluator, inspecting the task proposal to be inspected from the corresponding evaluation dimension, and respectively obtaining the corresponding inspection result, including: invoking a physical feasibility estimator, estimating the task proposal to be estimated from the estimation dimension corresponding to the physical feasibility estimator, and obtaining the estimation result of the physical feasibility estimator, comprising: invoking the physical feasibility estimator, analyzing the kinematic range of the robot in the task proposal to be inspected, and inspecting whether the kinematic range conflicts with the kinematic constraint of the target robot; Analyzing physical interaction logic of the robot in the task proposal to be inspected, and inspecting whether the physical interaction logic is reasonable; analyzing the interaction capability requirement of the robot in the task proposal to be inspected, and inspecting whether the interaction capability requirement conflicts with the hardware capability constraint of the target robot; and integrating the examination results of the physical feasibility estimator according to the examination results of the kinematic range, the physical interaction logic and the interaction capability requirements.
8. The method of claim 6, wherein the novelty evaluator's evaluation dimensions include complexity and redundancy; in the case that the evaluator includes a novelty evaluator, the invoking the evaluator respectively, inspecting the task proposal to be inspected from the corresponding evaluation dimension, and obtaining the corresponding inspection result respectively, includes: invoking a novelty evaluator, inspecting the task proposal to be inspected from an evaluation dimension corresponding to the novelty evaluator, and obtaining an inspection result of the novelty evaluator, including: invoking the novelty evaluator to analyze the interaction process of the robot in the task proposal to be inspected, and inspecting the complexity of the task proposal to be inspected according to a predetermined complexity indication standard, wherein the complexity indication standard at least comprises one of whether the task proposal comprises long-range planning, whether the task proposal comprises tool use or whether the task proposal comprises deformable object operation; searching in a preset task data set according to task logic of the robot in the task proposal to be inspected, and inspecting redundancy of the initial task proposal; and integrating the inspection results of the novelty evaluator according to the inspection results of the complexity and the redundancy.
9. The method of claim 6, wherein the evaluation dimension of the constraint consistency evaluator comprises content authenticity and scene logical suitability; In the case that the evaluator includes a constraint consistency evaluator, the invoking the evaluator respectively, inspecting the task proposal to be inspected from the corresponding evaluation dimension, and obtaining the corresponding inspection result respectively, includes: Invoking a constraint consistency evaluator, inspecting the task proposal to be inspected from an evaluation dimension corresponding to the constraint consistency evaluator, and obtaining an inspection result of the constraint consistency evaluator, wherein the method comprises the following steps: invoking a constraint consistency evaluator to detect whether task description items related in the task proposal to be inspected are contained in the task description data group, and inspecting the content authenticity of the task proposal to be inspected; invoking a constraint consistency evaluator to detect a logic relationship among task description items involved in the task proposal to be inspected, and inspecting scene logic suitability of the task proposal to be inspected; and integrating the examination results of the constraint consistency assessor according to the examination results of the content authenticity and the scene logic suitability.
10. The method of claim 1, wherein the data sampling the preset task description item based on historical task data of corresponding description data of the preset task description item in the task space tuple comprises: Based on historical count information of the category to which the corresponding description data of the preset task description items in the task space tuple belongs, carrying out data sampling on the preset task description items; The method further comprises the steps of: and modifying a corresponding counter value for the corresponding description data category of the preset task description item related in the target task proposal.
11. The method of claim 4 or 5, wherein the historical interaction record comprises task statement and feedback statement pairs in a historical task proposal; Further comprises: and under the condition that operation feedback data corresponding to the target task proposal is obtained, corresponding feedback sentences in the operation feedback data to task sentences in the corresponding target task proposal, and storing the feedback sentences in the operation feedback data into the history interaction record.
12. The method as recited in claim 1, further comprising: acquiring relevant image data of the task description data group; providing the task description data set as constraint conditions to a proposal generator to generate a target task proposal, comprising: And providing the task description data set as constraint conditions and the image data to a proposal generator to generate a target task proposal.
13. A robotic task generating system, comprising: the system comprises a determining module, a processing module and a processing module, wherein the determining module is used for determining a task space tuple for describing a robot task, and the task space tuple comprises a plurality of task description items; The sampling module is used for carrying out data sampling on the preset task description items based on historical task data of the corresponding description data of the preset task description items in the task space tuple to obtain a task description data group at least comprising the corresponding sampling data of the preset task description items; and the generation module is used for providing the task description data set as constraint conditions for the proposal generator to generate a target task proposal.
14. A computer device comprising a processor, a memory, and a bus; the memory stores machine-readable instructions executable by the processor; the processor and the memory communicate with each other through a bus when the computer device is running; the machine readable instructions, when executed by the processor, perform the steps of a robotic task generating method as claimed in any one of the claims 1-12.

Description

Robot task generation method, system and computer equipment Technical Field The disclosure relates to the field of personal intelligence, and in particular relates to a robot task generation method, a system and computer equipment. Background With the development of self-contained intelligence, the generalization ability of the visual-Language-Action (VLA) model is highly dependent on large-scale, high-quality training data. The current construction data is mainly based on manually designed and teleoperated paths. The scheme is high in physical and manpower cost, is limited by the cognitive deviation of human designers, and is limited by simple primitives such as grabbing, placing and the like. The data set presents serious long-tail distribution, lacks long-range planning, double-hand coordination and other complex interaction data, limits the generalization capability of the model in a novel scene, and cannot solve the problem of unbalanced distribution of robot training data. In summary, a method for generating robot tasks capable of automatically generating training data with balanced distribution is needed to solve the defects in the prior art. Disclosure of Invention The embodiment of the disclosure provides a robot task generation method, a system and computer equipment, which are used for solving the problem of uneven distribution of training data of the existing robot. Based on the above problems, in a first aspect, an embodiment of the present disclosure provides a method for generating a robot task, including: Determining a task space tuple for describing a robot task, wherein the task space tuple comprises a plurality of task description items; Based on historical task data of corresponding description data of preset task description items in the task space tuple, performing data sampling on the preset task description items to obtain a task description data set at least containing the corresponding sampling data of the preset task description items; And providing the task description data set as constraint conditions for a proposal generator to generate a target task proposal. With reference to the first aspect, in one possible implementation manner, the performing data sampling on the preset task description item based on the historical task data of the corresponding description data of the preset task description item in the task space tuple to obtain a task description data set at least including the corresponding sampling data of the preset task description item includes: And based on the use frequency of the corresponding description data of the preset task description items in the task space tuple in the historical task data, carrying out data sampling on the preset task description items to obtain a task description data group at least comprising the corresponding sampling data of the preset task description items, wherein the use frequency is lower than the preset frequency. With reference to the first aspect, in a possible implementation manner, the preset task description item includes a scene category set for characterizing a use scene of the robot, a task object library for characterizing an interaction object of the robot, and a task skill library for characterizing an interaction mode of the robot; based on the use frequency of the corresponding description data of the preset task description items in the task space tuple in the historical task data, carrying out data sampling on the preset task description items to obtain a task description data group at least comprising the corresponding sampling data of the preset task description items, wherein the method comprises the following steps: according to the frequency of executing tasks by the robot in each scene category in the scene category set, determining the scene category with the use frequency lower than a preset first threshold value as a target scene category; selecting task objects logically associated with the target scene category from the task object library based on the semantics of the target scene category to obtain a task object set; Based on the semantics of the target scene category, selecting task skills logically associated with the target scene category from the task skill library to obtain a task skill set; according to the historical interaction frequency of the robot and the task objects, task objects with the use frequency within a first preset low-frequency order range are screened from the task object set to obtain a candidate task object set; According to historical use frequency of the robot on task skills, task skills with use frequency within a second preset low-frequency order range are screened from the task skill set, and a candidate task skill set is obtained; And determining the target scene category, the candidate task object set and the candidate task skill set as the task description data set. With reference to the first aspect, in a possible implementation manner, the providing the task desc