CN-122021712-A - Task processing method and device based on robot, computer equipment and medium

CN122021712ACN 122021712 ACN122021712 ACN 122021712ACN-122021712-A

Abstract

The application belongs to the technical field of artificial intelligence, and relates to a task processing method, a device, computer equipment and a medium based on a robot, wherein the task processing method comprises the steps of carrying out demand analysis on an application scene of the robot to obtain scene demand information; the method comprises the steps of collecting hardware parameters of a robot based on a preset tool, carrying out association analysis on scene demand information and the hardware parameters to generate a modeling demand result, carrying out model construction processing on the basis of a target model frame and the modeling demand result to obtain a large model, carrying out model optimization on the large model based on a lightweight compression strategy to obtain a target large model, compiling the target large model based on a compiler to obtain a model file, deploying the model file into the robot based on a deployment strategy to obtain a target robot, and processing tasks to be processed based on the target robot. The application can be applied to the task processing scenes of robots in the fields of financial science and technology and digital medical treatment, and effectively improves the task processing efficiency of the robots.

Inventors

WANG JIANZONG
SUN AOLAN

Assignees

平安科技（深圳）有限公司

Dates

Publication Date: 20260512
Application Date: 20260109

Claims (10)

1. The task processing method based on the robot is characterized by comprising the following steps of: performing demand analysis on an application scene of the robot to obtain corresponding scene demand information; Acquiring hardware parameters of the robot based on a preset tool; performing association analysis on the scene demand information and the hardware parameters to generate a corresponding modeling demand result; performing model construction processing based on a target model frame and the modeling demand result to obtain a corresponding large model; Performing model optimization processing on the large model based on a preset lightweight compression strategy to obtain a corresponding target large model; compiling the target body large model based on a preset compiler to obtain a corresponding model file; And deploying the model file into the robot based on a preset deployment strategy to obtain a deployed target robot, and processing a task to be processed based on the target robot.
2. The robot-based task processing method according to claim 1, wherein the step of performing association analysis on the scene demand information and the hardware parameters to generate a corresponding modeling demand result specifically includes: Invoking a preset mapping engine to map the scene demand information and the hardware parameters to generate a corresponding mapping result; generating a corresponding requirement specification based on the mapping result; Processing the mapping result based on a preset modeling strategy to generate a corresponding modeling result; Integrating the requirement specification and the modeling result to obtain corresponding integrated data; and taking the integrated data as the modeling requirement result.
3. The robot-based task processing method according to claim 1, wherein the step of performing model construction processing based on the target model frame and the modeling demand result to obtain a corresponding large model comprises: Calling an initial model established based on a target model architecture, wherein the initial model at least comprises a multi-mode fusion front-end module, a selective state space module and a task processing back-end module; Performing model adjustment processing on the initial model based on the modeling demand result to obtain a corresponding first generation model; Acquiring a scene data set collected in advance; performing fine tuning training processing on the first generation model based on the scene data set to obtain a corresponding second generation model; And taking the second generative model as the large body model.
4. The robot-based task processing method according to claim 3, wherein the step of performing model adjustment processing on the initial model based on the modeling requirement result to obtain a corresponding first generated model specifically includes: acquiring task priority from the modeling demand result; based on the task priority, carrying out adjustment processing on the network layer of the initial model to obtain a corresponding first processing model; collecting self-knowledge of the robot; performing knowledge embedding processing on the first processing model based on the self knowledge to obtain a corresponding second processing model; And taking the second processing model as the first generating model.
5. The robot-based task processing method according to claim 1, wherein the step of performing model optimization processing on the large model based on a preset lightweight compression strategy to obtain a corresponding target large model specifically comprises: performing feature layer mixing precision quantization processing on the large model to obtain a corresponding first large model; performing structural layer task perception pruning treatment on the first large model to obtain a corresponding second large model; Performing end-side inference optimization processing on the second large model to obtain a corresponding third large model; and taking the third large model as the target large model.
6. The robot-based task processing method of claim 5, wherein the step of performing feature layer hybrid precision quantization processing on the large model to obtain a corresponding first large model specifically comprises: obtaining hardware constraint parameters from the modeling demand result; Determining a corresponding quantization strategy based on the hardware constraint parameters; Based on a preset quantization tool, performing quantization operation on the large model with body according to the quantization strategy to obtain a corresponding first appointed large model; Performing accuracy verification on the first specified large model based on a preset evaluation strategy; And if the first specified large model passes the accuracy verification, taking the first specified large model as the first large model.
7. The robot-based task processing method of claim 5, wherein the step of performing structural layer task aware pruning on the first large model to obtain a corresponding second large model specifically comprises: acquiring task priority from the modeling demand result; Determining a corresponding pruning strategy based on the task priority; Based on a preset pruning algorithm, pruning the structure of the first large model according to the pruning strategy to obtain a corresponding second designated large model; migrating knowledge of the first large model into the second specified large model based on a knowledge distillation technology to obtain a processed third specified large model; and taking the third specified large model as the second large model.
8. A robot-based task processing device, comprising: The analysis module is used for carrying out demand analysis on the application scene of the robot to obtain corresponding scene demand information; The acquisition module is used for acquiring hardware parameters of the robot based on a preset tool; the generating module is used for carrying out association analysis on the scene demand information and the hardware parameters to generate a corresponding modeling demand result; The building module is used for carrying out model building processing based on a target model frame and the modeling demand result to obtain a corresponding large model; the optimization module is used for carrying out model optimization processing on the large model based on a preset lightweight compression strategy to obtain a corresponding target large model; The compiling module is used for compiling the target large model based on a preset compiler to obtain a corresponding model file; The processing module is used for deploying the model file into the robot based on a preset deployment strategy to obtain a deployed target robot and processing a task to be processed based on the target robot.
9. A computer device comprising a memory and a processor, the memory having stored therein computer readable instructions which when executed by the processor implement the steps of the robot-based task processing method of any of claims 1 to 7.
10. A computer readable storage medium, characterized in that it has stored thereon computer readable instructions, which when executed by a processor, implement the steps of the robot-based task processing method according to any of claims 1 to 7.

Description

Task processing method and device based on robot, computer equipment and medium Technical Field The application relates to the technical field of artificial intelligence, which can be applied to the fields of financial science and technology, digital medical treatment and the like, in particular to a task processing method, a task processing device, computer equipment and a storage medium based on a robot. Background Under the current depth fusion of the robot technology and the artificial intelligence, the large model endows the robot with strong environment sensing and decision making capability, so that the robot can execute diversified tasks in various complex scenes. The existing large model is mostly designed based on a transducer architecture, and the architecture can effectively capture the association relation among elements at different positions in an input sequence by means of a self-attention mechanism, so that excellent performance is shown in various tasks. However, the self-attention mechanism has a significant computational complexity problem that grows in square steps with the length of the input sequence. In the case of limited calculation of the end-side equipment, this characteristic causes serious delay in the end-side thrust of the robot, resulting in low task processing efficiency. In the field of financial insurance, taking an insurance claim investigation robot as an example, the insurance claim investigation robot needs to perform rapid and accurate information acquisition and analysis on complex claim settlement sites, including identifying damaged articles, positioning damaged positions and the like. However, the existing large model has too high reasoning delay in processing on-site multi-mode data due to the problem of computational complexity, can not respond to investigation requirements in time, influences the efficiency and accuracy of claim settlement, and brings barriers to smooth development of insurance business. In the field of digital medical treatment, the surgical auxiliary robot needs to process multi-mode data such as visual images, medical images, doctor voice instructions and the like in real time in the surgical process so as to accurately assist a doctor in completing surgical operations. However, due to the restriction of the calculation complexity of the self-attention mechanism, the robot is seriously delayed when pushing the end side, so that the extremely high requirement of the operation on the real-time performance is difficult to meet, and the operation safety and effect can be adversely affected. Therefore, a large model-related technology capable of effectively reducing the efficiency of end-side thrust is needed to promote the wide application and performance improvement of robots in various fields. Disclosure of Invention The embodiment of the application aims to provide a task processing method, a device, computer equipment and a storage medium based on a robot, which are used for solving the technical problem that the task processing efficiency is low due to serious delay when the existing robot is pushed to the end side. In a first aspect, a task processing method based on a robot is provided, including: performing demand analysis on an application scene of the robot to obtain corresponding scene demand information; Acquiring hardware parameters of the robot based on a preset tool; performing association analysis on the scene demand information and the hardware parameters to generate a corresponding modeling demand result; performing model construction processing based on a target model frame and the modeling demand result to obtain a corresponding large model; Performing model optimization processing on the large model based on a preset lightweight compression strategy to obtain a corresponding target large model; compiling the target body large model based on a preset compiler to obtain a corresponding model file; And deploying the model file into the robot based on a preset deployment strategy to obtain a deployed target robot, and processing a task to be processed based on the target robot. In a second aspect, a robot-based task processing device is provided, including: The analysis module is used for carrying out demand analysis on the application scene of the robot to obtain corresponding scene demand information; The acquisition module is used for acquiring hardware parameters of the robot based on a preset tool; the generating module is used for carrying out association analysis on the scene demand information and the hardware parameters to generate a corresponding modeling demand result; The building module is used for carrying out model building processing based on a target model frame and the modeling demand result to obtain a corresponding large model; the optimization module is used for carrying out model optimization processing on the large model based on a preset lightweight compression strategy to obtain a corresponding target l