CN-122008216-A - Double-arm cooperative control method of wheel type humanoid robot

CN122008216ACN 122008216 ACN122008216 ACN 122008216ACN-122008216-A

Abstract

The invention relates to the technical field of robot control, and discloses a double-arm cooperative control method of a wheel type humanoid robot, which comprises the steps of firstly acquiring global data of an environment, a task and the robot, and establishing a double-arm DH model; the tasks are decomposed through a layered task network algorithm, a master arm and slave arm or double-arm cooperative mode is selected by combining a deep learning algorithm, and a corresponding master-slave mode or parallel cooperative control method is adopted for different modes. The invention realizes the accurate matching of the control strategy and the operation mode, improves the operation precision, the cooperative efficiency and the scene suitability, ensures the operation stability and the reliability, and is suitable for complex operation scenes.

Inventors

WANG KEXIN
WEI JUN
YANG HANGYU
Pan Guanchao
SONG CHENGLIN
GUO XIANGNAN
DONG JIANLONG
LIU GUANGUO
JIANG LIGUO
WANG SIHAI
ZHANG YU
DONG ZHI
ZHAO XIA
Tian Wenmao

Assignees

中铁十二局集团有限公司
南京智航科技发展有限公司
中铁建南方建设投资有限公司
苏交科集团股份有限公司

Dates

Publication Date: 20260512
Application Date: 20260305

Claims (8)

1. The double-arm cooperative control method of the wheel type humanoid robot is characterized by comprising the following steps of: s1, acquiring environment state, task information and global data of the state of a robot, and establishing a robot double-arm DH model; S2, decomposing the task through a hierarchical task network algorithm based on the global data acquired in the S1; s3, selecting behavior modes of the two arms according to a deep learning algorithm based on a task decomposition result, wherein the behavior modes are divided into a master arm and slave arm mode and a double arm cooperative mode; and S4, selecting a corresponding control method according to different behavior modes of the two arms, wherein the control method comprises a master-slave cooperative control method and a parallel cooperative control method, and the master-slave arm mode and the two-arm cooperative mode are respectively corresponding to each other.
2. The method for cooperatively controlling the two arms of the wheeled humanoid robot according to claim 1 is characterized in that the global data in S1 specifically comprises the environmental state including three-dimensional coordinates of obstacles, positioning information of work stations and size and material properties of a target object, the task information including task types, task completion time limit and precision requirements, and the robot state including angles of joints of the two arms, joint moments, positions and postures of a movable chassis and residual electric quantity.
3. The method for controlling the double-arm cooperative control of the wheeled humanoid robot according to claim 1, wherein the layering decomposition rule of the layering task network algorithm in the step S2 is that the whole task is sequentially disassembled into a top-layer task, an abstract subtask and an atomic subtask, wherein the top-layer task is a whole job target, the abstract subtask comprises a mobile subtask and a double-arm operation subtask, and the atomic subtask is a single action which can be directly executed by the robot.
4. The method for collaborative control of both arms of a wheeled humanoid robot according to claim 3, wherein the task layering decomposition process of S2 further comprises a subtask feasibility verification step of verifying whether the joint angle corresponding to the atomic subtask is within a preset motion limit according to a robot double-arm DH model.
5. The method for controlling the cooperative double arms of the wheeled humanoid robot according to claim 3, wherein the specific execution logic of the deep learning algorithm in S3 is as follows: S31, constructing a deep learning training data set, wherein a sample of the training data set comprises an atomic subtask type, the environmental state data acquired in the step S1, the state data of the robot and a corresponding optimal double-arm behavior mode label; s32, constructing a deep learning network model integrating space feature extraction and time sequence feature analysis, wherein the model comprises a convolution layer for extracting space features of an environment state and a state of a robot, a long-period memory layer for analyzing time sequence associated features of an atomic subtask, and a full-connection layer for outputting a behavior pattern classification result; s33, training the deep learning network model by adopting a supervised learning mode, and iteratively optimizing model parameters by a back propagation algorithm until the behavior mode classification accuracy of the model reaches a preset threshold; s34, inputting the atomic subtask combination information, the environment state data and the state data of the robot, which are acquired in real time, into a model which is trained and completed, and outputting a corresponding double-arm behavior mode decision result.
6. The method for collaborative control of both arms of a wheeled humanoid robot according to claim 5, wherein the optimal dual-arm behavior pattern label constructed in S31 is determined by combining simulation verification with real object calibration, and the optimal performance determination index priority of the label is a task suitability index, a motion safety index, an energy consumption and a stability index in sequence from high to low.
7. A wheeled humanoid robot double-arm cooperative control method according to claim 3, characterized in that the master-slave cooperative control method comprises the steps of: s411, based on the atomic subtask decomposition result of the step S2 and the global data obtained in the step S1, completing the match between the master arm function division and the initial role S412, performing core track planning for the main arm, wherein the planning process takes the precision requirement in the task information as core constraint, and constructs an obstacle avoidance constraint space by combining the three-dimensional coordinates of the obstacle in the environment state and the positioning information of the operation station; s413, establishing a linkage following mapping relation between the slave arm and the master arm, and taking real-time pose data of the end effector of the master arm as a datum reference of the movement of the slave arm; S414, building a master arm and slave arm real-time feedback control loop, wherein the control loop comprises a control unit, and the master arm and the slave arm respectively acquire the joint angle, the joint moment and the end effector pose data of the master arm and the slave arm in real time and synchronously upload the data to the control unit; and S415, after the atomic subtask is completed, the master arm sends an action end signal to the slave arm, and the slave arm releases the following state and executes a reset action.
8. A wheeled humanoid robot double-arm cooperative control method according to claim 3, characterized in that the parallel cooperative control method comprises the steps of: s421, completing double-arm associated task allocation and action collaborative planning based on the atomic subtask decomposition result in the step S2 and the global data acquired in the step S1; S422, constructing a double-arm collaborative track planning constraint space, fusing the environmental state data acquired in the step S1 with a robot double-arm DH model, determining the movement range of the double-arm end effector, and synchronously generating the movement track of the double-arm end effector by adopting a multi-target collaborative optimization algorithm; s423, establishing a double-arm motion synchronous control mechanism, and constructing a control framework integrating pose synchronization, speed synchronization and moment coordination; and S425, after the atomic subtask is completed, executing double-arm cooperative reset action, and after the double-arm synchronous reset track is generated, closing a cooperative control loop and waiting for a next round of task instruction.

Description

Double-arm cooperative control method of wheel type humanoid robot Technical Field The invention relates to the technical field of robot control, in particular to a double-arm cooperative control method of a wheeled humanoid robot. Background In the operation scene of the wheel type humanoid robot, double-arm cooperative control is a core key for realizing complex operation tasks, and the control precision, cooperative fluency and the cooperation stability with a chassis directly determine the operation performance of the robot. The existing wheel type humanoid robot double-arm cooperative control technology has the defects that under most working scenes, the robot double arms usually play roles of different master arms, slave arms and cooperative arms, and the prior art generally uses the same control method to control the working modes of different logics of the master arm, the slave arms and the two cooperative arms of the double-arm robot, so that the matching degree of a control strategy and actual operation requirements is low, and the operation precision and the cooperative efficiency under different modes are difficult to be considered. Disclosure of Invention Aiming at the defects existing in the prior art, the invention aims to provide a double-arm cooperative control method of a wheeled humanoid robot, which aims to solve the technical problems. Aiming at the defects existing in the prior art, the invention aims to provide a double-arm cooperative control method of a wheeled humanoid robot, which aims to solve the technical problems. The double-arm cooperative control method of the wheel type humanoid robot comprises the following steps: A method for controlling the double arms of a wheel robot in a coordinated manner comprises the following steps of S1, obtaining environmental state, task information and global data of the state of the robot, establishing a robot double arm DH model, S2, decomposing tasks through a hierarchical task network algorithm based on the global data obtained in S1, S3, selecting behavior modes of the double arms according to a deep learning algorithm based on task decomposition results, wherein the behavior modes are divided into a master arm mode, a slave arm mode and a double arm coordinated mode, S4, selecting corresponding control methods according to different behavior modes of the double arms, and the control methods comprise a master-slave coordinated control method and a parallel coordinated control method, and the master-slave arm mode and the double arm coordinated mode correspond to the master arm mode and the slave arm mode respectively. Optionally, the global data in the step S1 specifically includes that the environment state includes three-dimensional coordinates of an obstacle, positioning information of a working station and size and material properties of a target object, the task information includes task types, task completion time limit and precision requirements, and the state of the robot itself includes angles of joints of two arms, joint moment, pose of a mobile chassis and residual electric quantity. Optionally, the layering decomposition rule of the layering task network algorithm in the step S2 is that the whole task is sequentially disassembled into a top-layer task, an abstract subtask and an atomic subtask, wherein the top-layer task is a whole job target, the abstract subtask comprises a mobile subtask and a double-arm operation subtask, and the atomic subtask is a single action which can be directly executed by a robot. Optionally, in the task layering decomposition process of the step S2, a subtask feasibility verification step is further included, wherein according to a robot double-arm DH model, whether the joint angle corresponding to the atomic subtask is within a preset motion limit is verified. The specific execution logic of the deep learning algorithm in the step S3 is that a deep learning training data set is built, a sample of the training data set comprises an atomic subtask type, environment state data acquired in the step S1, state data of the robot and a corresponding optimal double-arm behavior mode label, a deep learning network model integrating space feature extraction and time sequence feature analysis is built, the model comprises a convolution layer for extracting the environment state and the state space feature of the robot, a long-period memory layer for analyzing the time sequence related feature of the atomic subtask and a full-connection layer for outputting a behavior mode classification result, the deep learning network model is trained in a supervision learning mode, model parameters are iterated and optimized through a back propagation algorithm until the behavior mode classification accuracy of the model reaches a preset threshold value, and the corresponding double-arm behavior mode decision result is output by inputting the atomic subtask combination information acquired in real time, the environment st