CN-121973189-A - Robot correction method, robot correction device, robot and computer readable storage medium

CN121973189ACN 121973189 ACN121973189 ACN 121973189ACN-121973189-A

Abstract

The application discloses a robot correction method, a robot correction device, a robot and a computer readable storage medium, wherein the robot correction method comprises the steps of collecting multidimensional original sensing data through a plurality of heterogeneous sensors; the method comprises the steps of obtaining synchronous and aligned multi-source sensing information based on multi-dimensional original sensing data, constructing a multi-directional diagram based on the multi-source sensing information, generating a node sequence based on the multi-directional diagram by adopting a random walk sampling strategy with bias, dynamically adjusting sampling probability of a neighborhood node according to data quality and timestamp freshness of the multi-source sensing information associated with each side by the random walk sampling strategy, training a Skip-gram model according to the node sequence to obtain a low-dimensional embedded vector of each node, and correcting a motion track or an operation strategy of a robot based on the low-dimensional embedded vector. Thus, by embedding nodes more prone to reflect high quality, recent operational context, reliability and adaptability of trajectory or policy correction is improved.

Inventors

CHEN LEI
HUANG JINYE
YAN SHENGKAI

Assignees

深圳市旗扬特种装备技术工程有限公司

Dates

Publication Date: 20260505
Application Date: 20260115

Claims (10)

1. A robot calibration method, wherein a plurality of heterogeneous sensors are configured in the robot body and its environment, the method comprising: acquiring multidimensional raw sensing data through the heterogeneous sensors in the process of executing tasks by the robot; performing space-time calibration and noise filtering on the multidimensional original sensing data to obtain synchronous and aligned multisource sensing information; constructing a multi-directional graph based on the multi-source sensing information, wherein nodes of the multi-directional graph represent the operation state of the robot, edges represent state transition examples, and each edge is attached with the corresponding multi-source sensing information; Generating a node sequence for graph embedding training by adopting a random walk sampling strategy with bias based on the multiple directed graphs, wherein the random walk sampling strategy dynamically adjusts sampling probability of a neighborhood node according to data quality and timestamp freshness of multi-source sensing information associated with each edge; training a Skip-gram model according to the node sequence to obtain a low-dimensional embedded vector of each node; and correcting the motion trail or the operation strategy of the robot based on the low-dimensional embedded vector.
2. The robot correction method of claim 1, wherein the generating a node sequence for graph embedding training based on the multiple directed graphs using a random walk sampling strategy with bias comprises: respectively calculating a first transition probability based on the data quality of the multi-source sensing information associated with each side and a second transition probability based on the freshness of the timestamp of the multi-source sensing information associated with each side aiming at a target node and a neighborhood node of the target node; according to a preset weight parameter, fusing the first transition probability and the second transition probability to obtain a final transition probability; And executing offset random walk according to the final transition probability to generate the node sequence.
3. The robot correction method according to claim 2, wherein the step of calculating a first transition probability based on data quality of the multi-source sensing information associated with each edge for a target node and a neighborhood node of the target node, comprises: Aiming at the target node and the neighborhood nodes of the target node, calculating a first transition probability through a first transition probability formula; wherein the first transition probability formula is expressed as: Wherein, the To be from the target node Neighborhood nodes to the target node Is selected from the group consisting of a first transition probability, For the target node Neighborhood nodes to the target node All multi-source sensory information sets associated therewith, Is the first in the collection The quality score of the group data is calculated, For the target node Is a set of all of the neighborhood nodes, For the target node And a neighborhood node of the target node All of the multi-source sensory information associated therewith, For the target node And a neighborhood node of the target node All multisource sensor information associated therewith Quality score of group data.
4. The robot correction method according to claim 2, wherein the calculating, for a target node and a neighborhood node of the target node, a second transition probability based on freshness of a timestamp of multi-source sensing information associated with each edge includes: aiming at the target node and the neighborhood nodes of the target node, calculating a second transition probability through a second transition probability formula; the second transition probability formula is expressed as: Wherein, the For the target node Neighborhood nodes to the target node Is selected from the group consisting of a first transition probability, For the target node And a neighborhood node of the target node A set of all timestamp freshness scores in between, For the target node And a neighborhood node of the target node Between the first The timestamp freshness score corresponding to the group multisource sensing information, For the target node Is a set of all of the neighborhood nodes, For the target node And a neighborhood node of the target node A set of all timestamp freshness scores in between, For the target node And a neighborhood node of the target node Between the first The timestamp freshness score corresponding to the group multisource sensing information.
5. The robot correction method according to claim 2, wherein the fusing the first transition probability and the second transition probability according to a preset weight parameter to obtain a final transition probability includes: According to the preset weight parameter and a weighted fusion formula, fusing the first transition probability and the second transition probability to obtain the final transition probability; The weighted fusion formula is expressed as: Wherein, the For the preset weight parameter to be used, For the first transition probability of the first transition, And the second transition probability.
6. The robot calibration method according to claim 1, wherein training the Skip-gram model according to the node sequence results in a low-dimensional embedded vector for each node, comprising: Extracting subsequences only containing the robot operation state nodes from the node sequences to serve as sampling sequences; constructing a first corpus and a second corpus based on the node sequence and the sampling sequence respectively; Training a first Skip-gram model based on the first corpus to obtain a first embedded vector of each node; training a second Skip-gram model based on the second corpus to obtain a second embedded vector of each node; and aiming at each robot operation state node, splicing or weighting fusion is carried out on the first embedded vector and the second embedded vector corresponding to each robot operation state node, so as to obtain a low-dimensional embedded vector of the node.
7. The robot correction method of claim 6, further comprising: Respectively training the first Skip-gram model and the second Skip-gram model through an optimized objective function to obtain the first embedded vector and the second embedded vector; The objective function is expressed as: Wherein, the Is the first A corpus of dimensions, As a function of the index of the values, Is a node , Is a node , Is a node Is used to determine the embedded vector of (c), Is a node Is used to extract the vector.
8. A robot calibration device, wherein a plurality of heterogeneous sensors are disposed in the robot body and its environment, the device comprising: the acquisition module is used for acquiring multidimensional original sensing data through the heterogeneous sensors in the process of executing tasks by the robot; the preprocessing module is used for carrying out space-time calibration and noise filtering on the multidimensional original sensing data to obtain synchronous and aligned multisource sensing information; The construction module is used for constructing a multi-directional graph based on the multi-source sensing information, wherein nodes of the multi-directional graph represent the operation states of the robot, edges represent state transition examples, and each edge is attached with the corresponding multi-source sensing information; The generation module is used for generating a node sequence for graph embedding training by adopting a random walk sampling strategy with bias based on the multiple directed graphs, wherein the random walk sampling strategy dynamically adjusts the sampling probability of the neighborhood nodes according to the data quality and the time stamp freshness of the multi-source sensing information associated with each side; The training module is used for training the Skip-gram model according to the node sequence to obtain a low-dimensional embedded vector of each node; and the execution module is used for correcting the motion trail or the operation strategy of the robot based on the low-dimensional embedded vector.
9. A robot comprising a robot body, a robot body and a robot body, characterized by comprising the following steps: One or more processors; A memory; One or more applications, wherein the one or more applications are stored in the memory and configured to be executed by the one or more processors, the one or more applications configured to perform the robot correction method of any of claims 1-7.
10. A computer readable storage medium, characterized in that the computer readable storage medium has stored therein a program code, which is callable by a processor for performing the robot correction method according to any one of claims 1-7.

Description

Robot correction method, robot correction device, robot and computer readable storage medium Technical Field The present application relates to the technical field of robot control, and more particularly, to a robot correction method, apparatus, robot, and computer-readable storage medium. Background When the industrial robot performs tasks such as precise assembly, splicing or grabbing, the operation precision is highly dependent on accurate perception and modeling of self state and environment interaction. Current methods typically construct an operational flow model based on sensor data and conduct behavioral analysis or control optimization based thereon. However, in practical applications, it is difficult to obtain consistent and reliable state characterization due to the lack of preliminary fusion of the sensing data. In addition, in the conventional modeling method, transitions between states are generally regarded as non-differential logic connections in the process of describing operations, and performance differences (such as successful completion and deviation-generating attempts) of the same transition in different execution examples cannot be effectively reflected. This makes behavior patterns learned from historical data lacking the ability to distinguish the effectiveness of the operation, thereby limiting its usefulness and robustness in real-time trajectory adjustment or policy optimization. Disclosure of Invention In view of the above problems, the present application provides a robot correction method, apparatus, robot and computer readable storage medium, which can perform preliminary fusion on sensing data, automatically identify and structurally characterize a high-quality and recent effective behavior pattern in historical operation data of the robot, and drive adaptive, semantic aware trajectory or strategy correction on a current task by using the method, apparatus, robot and computer readable storage medium, thereby significantly improving success rate and robustness of a precise operation task. According to the method, in the process of executing tasks by a robot, multi-dimensional original sensing data are collected through the plurality of heterogeneous sensors, space-time calibration and noise filtering are conducted on the multi-dimensional original sensing data, synchronous and aligned multi-source sensing information is obtained, a multi-directional graph is built based on the multi-source sensing information, nodes of the multi-directional graph represent operation states of the robot, edges represent state transition examples, and corresponding multi-source sensing information is attached to each edge, a random walk sampling strategy with bias is adopted to generate a node sequence for graph embedding training based on the multi-directional graph, the random walk sampling strategy dynamically adjusts sampling probability of neighborhood nodes according to data quality and time stamp freshness of the multi-source sensing information associated with each edge, training is conducted on a Skip-gram model according to the node sequence to obtain low-dimensional embedding vectors of each node, and motion tracks or operation strategies of the robot are corrected based on the low-dimensional embedding vectors. The embodiment of the application also provides a robot correction device, wherein a plurality of heterogeneous sensors are configured in a robot body and an environment of the robot body, the robot correction device comprises an acquisition module, a preprocessing module, a construction module, a generation module and a generation module, wherein the acquisition module is used for acquiring multidimensional original sensing data through the heterogeneous sensors in the process of executing tasks of the robot, the preprocessing module is used for carrying out space-time calibration and noise filtering on the multidimensional original sensing data to obtain synchronous and aligned multisource sensing information, the construction module is used for constructing a multi-directional graph based on the multisource sensing information, nodes of the multisource graph represent the operating state of the robot, edges of the multisource graph represent state transition examples, each edge is attached with the corresponding multisource sensing information, the generation module is used for generating a node sequence for graph embedding training based on the multi-directional graph, the random walk sampling strategy is used for dynamically adjusting the sampling probability of a neighborhood node according to the data quality and the time stamp freshness of the multisource sensing information associated with each edge, the training module is used for training a Skip-gram model according to the node sequence to obtain a low-dimensional embedding vector of each node, and the execution module is used for carrying out motion vector correction on the basis of the low-dimensional motion vec