CN-121973200-A - Redundant mechanical arm kinematics inverse solution method based on physical space constraint mechanism generation model

CN121973200ACN 121973200 ACN121973200 ACN 121973200ACN-121973200-A

Abstract

The disclosure relates to the technical field of mechanical arms, and discloses a redundant mechanical arm kinematics inverse solution method based on a physical space constraint mechanism generation model. The method comprises the steps of inputting the terminal target pose of a next moment into a generating type neural network for redundant mechanical arms with n degrees of freedom to obtain multiple groups of kinematic parameter solutions with r first degrees of freedom, determining n-r kinematic parameter solutions with the second degrees of freedom based on a conversion expression of the kinematic parameters with the second degrees of freedom, the terminal target pose and each group of kinematic parameter solutions with r first degrees of freedom, and selecting a group of optimal kinematic parameter solutions from the multiple groups of complete kinematic parameter solutions based on a target optimization function to reach the terminal target pose at the next moment. According to the method, a generating network capable of generating multiple groups of kinematic parameter solutions with r first degrees of freedom based on the terminal pose is established, so that the model operation complexity and the training cost can be reduced, and meanwhile, the accuracy of the terminal pose of the redundant mechanical arm is guaranteed by introducing a forward kinematic control equation.

Inventors

LI ZHIJIE
CHEN HAOLONG
YANG TIANLE
YI YUANLIN
ZHOU QIN
MA WEITAO
SUN LIBIN
XIE HENG
SHI LEI

Assignees

清华大学

Dates

Publication Date: 20260505
Application Date: 20260123

Claims (10)

1. A redundant mechanical arm kinematics inverse solution solving method based on a physical space constraint mechanism generation model is characterized in that the redundant mechanical arm comprises n degrees of freedom, n is a positive integer, and the method comprises the following steps: Under the condition that the terminal target pose of the redundant mechanical arm at the next moment is obtained, inputting the terminal target pose into a pre-trained generating neural network with a physical space constraint mechanism to obtain multiple groups of kinematic parameter solutions of r first degrees of freedom, wherein each group of kinematic parameter solutions comprises kinematic parameters of the r first degrees of freedom at the next moment; the generated neural network with the physical space constraint mechanism is obtained by training pre-generated training data, wherein the training data comprises a sample end pose and a plurality of groups of kinematic parameter sampling values capable of achieving r first degrees of freedom of the sample end pose; Determining a set of n-r second degree-of-freedom kinematic parameter solutions based on a conversion expression of the second degree-of-freedom kinematic parameters, the end target pose and each set of r first degree-of-freedom kinematic parameter solutions, wherein each set of r first degree-of-freedom kinematic parameter solutions and the corresponding n-r second degree-of-freedom kinematic parameter solutions form a set of complete kinematic parameter solutions, the conversion expression is determined based on a forward kinematic control equation, and the r first degree-of-freedom kinematic parameters are used for representing the n-r second degree-of-freedom kinematic parameters, so that a physical space constraint relation between n degrees of the redundant manipulator is established; and selecting a group of optimal kinematic parameter solutions from a plurality of groups of complete kinematic parameter solutions based on a target optimization function set according to operation requirements so as to control the redundant mechanical arm to operate according to the optimal kinematic parameter solutions, and reaching the tail end target pose at the next moment.
2. The method of claim 1, wherein the training data generation process comprises: sampling a plurality of kinematic parameter sampling values corresponding to each degree of freedom according to preset solving precision in an allowable variation range of the kinematic parameter of each degree of freedom, wherein one kinematic parameter sampling value corresponding to each degree of freedom forms a group of complete kinematic parameter solution samples; Determining terminal pose samples corresponding to each group of complete kinematic parameter solution samples based on the forward kinematic control equation, and establishing corresponding relations between the kinematic parameter sampling values of r first degrees of freedom in the complete kinematic parameter solution samples and the terminal pose samples to obtain initial training data; And cleaning the initial training data based on actual requirements to obtain the training data.
3. The method of claim 1, wherein training the generated neural network with physical space constraint mechanism using the training data comprises: Inputting each terminal pose sample and a plurality of groups of r first-degree-of-freedom kinematic parameter sampling values corresponding to the terminal pose samples into a to-be-trained generation type neural network to obtain a mapping of each group of r first-degree-of-freedom kinematic parameter sampling values into values in base distribution; inputting the values in the base distribution into a preset loss function to obtain loss values; And iteratively updating model parameters of the to-be-trained generated neural network based on the loss values, so that the to-be-trained generated neural network learns reversible mapping from the plurality of groups of r first-degree-of-freedom kinematic parameter sampling values to base distribution based on the terminal pose samples and the plurality of groups of r first-degree-of-freedom kinematic parameter sampling values corresponding to the terminal pose samples, and the trained generated neural network with a physical space constraint mechanism is obtained.
4. A method according to claim 3, wherein said inputting the end target pose into a pre-trained generative neural network with physical space constraint mechanism, resulting in multiple sets of kinematic parameter solutions of r first degrees of freedom, comprises: Inputting the tail end target pose into the trained generating neural network with the physical space constraint mechanism, so that the trained generating neural network with the physical space constraint mechanism samples a plurality of random variables from the base distribution, maps each random variable and the tail end target pose into a plurality of sets of kinematic parameter solutions with r first degrees of freedom based on the learned reversible mapping, and obtains a plurality of sets of kinematic parameter solutions with r first degrees of freedom, wherein the plurality of random variables correspond to the plurality of first degrees of freedom respectively.
5. The method of claim 3, wherein the generated neural network to be trained comprises a plurality of groups of transformation networks connected in sequence, each group of transformation networks comprising a coupling layer, a substitution layer and a coefficient network; the coupling layer is used for splitting input data corresponding to the terminal pose sample into a first part and a second part, and inputting the first part into the coefficient network; the coefficient network is used for processing the first part and the tail end pose sample to generate a scaling factor and an offset factor; The coupling layer is further configured to perform nonlinear transformation on the second portion based on the scaling factor and the offset factor, where the first portion and the nonlinear transformed second portion obtain a nonlinear transformation result of the coupling layer; the displacement layer is used for carrying out displacement on the dimension of the nonlinear transformation result to obtain output data so as to generate values in the base distribution based on the output data.
6. The method of claim 1, wherein the operational requirement can be represented by a theoretical expression, the operational requirement comprising at least one of: Optimal compliance requirements; Not exceeding a joint limit threshold; Obstacle avoidance requirements; energy optimal demand.
7. The method of claim 1, wherein the n degrees of freedom are each a revolute joint, or wherein the n degrees of freedom include a revolute joint and a telescopic joint.
8. The redundant mechanical arm kinematics inverse solution solving device based on the physical space constraint mechanism generation model is characterized in that the redundant mechanical arm comprises n degrees of freedom, n is a positive integer, and the device comprises: The first freedom degree kinematic parameter solution generating module is used for inputting the tail end target pose into a pre-trained generating type neural network with a physical space constraint mechanism under the condition that the tail end target pose of the redundant mechanical arm at the next moment is acquired, so as to obtain multiple groups of kinematic parameter solutions with r first freedom degrees, wherein each group of kinematic parameter solutions comprises the kinematic parameters of the r first freedom degrees at the next moment; The system comprises a first freedom degree kinematic parameter solution generating module, a second freedom degree kinematic parameter solution generating module, a forward motion control equation and a redundancy mechanical arm, wherein the first freedom degree kinematic parameter solution generating module is used for determining n-r first freedom degree kinematic parameter solutions based on a conversion expression of first freedom degree kinematic parameters, the end target pose and r first freedom degree kinematic parameter solutions, and the n-r first freedom degree kinematic parameter solutions, each r first freedom degree kinematic parameter solution and the corresponding n-r second freedom degree kinematic parameter solutions form a complete set of kinematic parameter solutions, the conversion expression is determined based on the forward motion control equation and represents the n-r second freedom degree kinematic parameters by using r first freedom degree kinematic parameters, so that a physical space constraint relation between the n freedom degrees of the redundancy mechanical arm is established, and the forward motion control equation is determined based on the structural parameters of the redundancy mechanical arm and is used for describing the relation between the end pose and the n freedom degrees; And the kinematic parameter solution determining module is used for selecting a group of optimal kinematic parameter solutions from a plurality of groups of complete kinematic parameter solutions based on a target optimization function set according to operation requirements so as to control the redundant mechanical arm to operate according to the optimal kinematic parameter solutions and reach the tail end target pose at the next moment.
9. A redundant manipulator kinematic inverse solution solving device based on a physical space constraint mechanism generation model, comprising a memory, a processor and a computer program stored on the memory, characterized in that the processor executes the computer program to implement the steps of the method of any one of claims 1 to 8.
10. A non-transitory computer readable storage medium having stored thereon a computer program, characterized in that the computer program when executed by a processor realizes the steps of the method according to any of claims 1 to 8.

Description

Redundant mechanical arm kinematics inverse solution method based on physical space constraint mechanism generation model Technical Field The disclosure relates to the technical field of mechanical arms, in particular to a redundant mechanical arm kinematics inverse solution method based on a physical space constraint mechanism generation model. Background The super-redundant mechanical arm has been widely applied in the fields of precision manufacture, medical operation, aerospace operation, nuclear industry, disaster relief and the like by virtue of high flexibility and obstacle avoidance capability, and can replace manual operation in complex environments in the future, so that the super-redundant mechanical arm becomes an important development direction of robot technology. For example, the wide-energy-spectrum ultrahigh-flux test stack is a large scientific device for carrying out material irradiation tests, rare isotope production and neutron scientific research, the internal space of the test stack is extremely narrow and complex, and the loading and unloading of irradiation targets (used for containing materials to be irradiated and isotope production samples) can be completed only by means of redundant mechanical arms. However, compared with a conventional non-redundant mechanical arm, the redundant mechanical arm kinematics inverse solution faces multiple challenges such as high computational complexity (the large amount of calculation caused by multiple degrees of freedom), difficult optimization of multiple solutions (the need of selecting the optimal solution from infinite solutions), joint overrun and potential deviation risks, insufficient real-time performance (the difficulty of meeting the requirement of quick response in the traditional method), and difficult multi-constraint collaborative optimization (the need of meeting the conditions of end precision, obstacle avoidance, energy consumption and the like). Compared with the limitation that only one set of solutions can be generated each time by a traditional numerical iteration or analytic method, the method for determining the generation type neural network has the remarkable prominent advantages that multiple groups of complete feasible solutions can be output in parallel through parameterized probability distribution, the solution space is covered on the whole, optimal scheme selection is provided for scenes such as complex assembly and obstacle avoidance planning, the inference speed of the method for determining the generation type neural network can reach millisecond (only 10ms is needed for generating 10 solutions), the requirements of continuous real-time adjustment under dynamic environment can be met by means of GPU acceleration efficiency, collision detection, joint limit, moment balance and other types of constraints can be naturally embedded in the training process, joint configuration which is free of collision and smooth in motion can be generated, tasks can be completed even when part of joints fail, the system robustness is improved, the single generation type neural network can be adapted to mechanical arms with different configurations and different degrees of freedom, the method can be applied to new scenes through small quantity of model parameter fine adjustment, the modeling migration capability is high, deployment is greatly reduced, the problem of the method can be solved by ten times, the problem of continuous real-time adjustment under dynamic environment can be solved compared with the traditional method, the problem of low-level precision of the position of the generation type neural network can be directly and the position of a linear position of a sub-level learning end can be avoided, and the problem of the vibration angle is solved. The advantages enable the determination method based on the generated neural network to play a key role in the fields of medical minimally invasive surgery robots (avoiding important organs, improving surgery safety), aerospace on-orbit services (zero-gravity environment multi-scheme planning, fault tolerance), industrial flexible manufacturing (flexible operation and micro assembly among dense equipment), disaster relief and special operation (obstacle avoidance in dangerous environments and unstructured scene adaptation), thoroughly change the solving paradigm of inverse solution of redundant mechanical arms, and open up new possibilities for autonomous flexible movement of intelligent robots in complex worlds. However, the application of the generated neural network to the inverse solution of the kinematics of the super-redundant mechanical arm still has a plurality of defects that firstly, training data is highly dependent, the training data is a high-quality data set covering a full working space, and the larger the number of degrees of freedom of the redundant mechanical arm is, the larger the solution space (namely, the set of all possible solutions of the kinematics pa