Intelligent self-reconfiguration robot collaborative configuration optimization method

Abstract

The invention discloses a collaborative configuration optimization method for an intelligent self-reconfiguration robot. The invention relates to the field of modular robot configuration-motion collaborative optimization in extraterrestrial planet surface inspection and detection tasks, constructs a multi-constraint dynamic graph optimization model for unified representation of configuration and motion control, and aims to improve the configuration reconstruction efficiency of a modular robot and obtain an optimal configuration-motion reconstruction scheme. The method comprises the following steps: carrying out restrictive configuration characterization on a robot by utilizing a rooted tree structure; constructing a sensing-driving integrated neural network to perform robot motion control representation; constructing a multi-constraint dynamic graph optimization model to uniformly represent configuration and motion control; constructing a fitness function for reconstruction optimization according to the evaluation function corresponding to the task; and performing configuration-motion optimization solution by using a non-dominated sorting genetic algorithm with an elitist strategy.

Description

Technical Field

[0001] The present invention relates to the technical field related to cooperative control methods and graph optimization theory, and in particular to an intelligent self-reconfigurable robot cooperative configuration optimization method. Background Art

[0002] Currently, both the fixed-configuration robot platform and the independent detection method of astronauts are difficult to meet the detection requirements of the unstructured complex lunar environment. The intelligent self-reconfigurable robot changes its own configuration driven by the environment and tasks, establishes a reconfiguration strategy based on environmental perception and intention decision-making, transforms into a configuration that fits the complex terrain, and has extremely strong terrain adaptability and structural variability.

[0003] Traditional modular robot reconfiguration methods generally do not consider environmental factors and are divided into two categories: reconfiguration methods with target configurations and those without target configurations. The target configuration is usually based on specific task requirements, usually task-driven, and the external environment is not a design factor. The operation task corresponds to the configuration of the modular robot's manipulator. Wu and Yang et al. optimized the combination of modules with different geometric parameters to obtain the modular manipulator configurations with the optimal dexterity and end-effector reachability respectively; the mobile task corresponds to the configuration of the mobile robot. Tang Shufeng optimized the configuration of the multi-legged robot to obtain the configuration with the optimal mobility. There is no general configuration design criterion for general tasks. In complex scenarios or when the specific task is not clear, it is necessary for the cognitive decision-making system to determine the task flow plan online, and then optimize the design of the target configuration on this basis.

[0004] On the other hand, the reconfiguration without a target configuration generally means that the modular robot generates quasi-periodic motions through continuous configuration reorganization to achieve various motion functions of the robot system in various environments. The reconfiguration without a target configuration is process-driven and generally studies the motion planning and optimization problems among multiple module units under given environmental information. Pickem established a rule set based on graphs and could perform motion path planning for any initial configuration composed of a large number of module units. Kurokawa et al. proposed a decentralized asynchronous control algorithm for distributed modular systems, which could plan the module reorganization order of the M-TRAN robot. The above methods provide useful references for the reconfiguration motion methods of modular robots without target configurations such as moving and obstacle crossing in various scenarios, but all assume that the external environment is known. Therefore, the effectiveness of the above methods in unknown environments still needs to be further verified, and their applications also have great limitations. Therefore, it is urgent to study an intelligent self-reconfigurable robot cooperative configuration optimization method. Summary of the Invention

[0005] Aiming at the detection requirements of unstructured complex terrain on the lunar surface, the present invention provides an intelligent self-reconfiguring robot collaborative configuration optimization method. By using a multi-constraint dynamic graph optimization model to uniformly represent the configuration and motion control, and using a non-dominated sorting genetic algorithm with an elite strategy to generate an optimal configuration scheme online, the configuration optimization efficiency of the self-reconfiguring modular robot is improved.

[0006] In a first aspect, there is provided an intelligent self-reconfiguring robot collaborative configuration optimization method, characterized by comprising:

[0007] Using a rooted tree structure to perform restricted configuration representation on the robot, and the restricted configuration is used to represent the directional observability of the sensing unit, the anti-collision condition, the cost and quality constraints;

[0008] Constructing a sensing-driving integrated neural network for robot motion control representation to describe the robot motion control model;

[0009] Constructing a multi-constraint dynamic graph optimization model R to uniformly represent the restricted configuration and motion control model of the robot, and the restricted configuration is used as the collaborative input and constraint of the control model;

[0010] Constructing a fitness function for reconstruction optimization according to the evaluation function corresponding to the task;

[0011] Using the non-dominated sorting genetic algorithm to perform restricted configuration-motion control optimization solution of the multi-constraint dynamic graph optimization model R.

[0012] Combined with the first aspect, in some implementation manners of the first aspect, the restricted configuration of the modular robot is described as γ(M)=G(V,E), γ is a rooted tree structure, G is an environmental representation function; M is the original configuration modeled into a network by using network technology; V is the set of faces or the set of modular units in the grid M; E is the dual edge set, expressed as E={(f,g)}, f and g are two adjacent faces in the grid M.

[0013] Combined with the first aspect, in some implementation manners of the first aspect, the description of the robot motion control model includes:

[0014] Taking the installation positions and measurement data of the sensors in the sensing system as the input neurons of the neural network, taking the motor positions and output torques of the drivers in the motion system as the output neurons, and selecting the Sigmoid function as the activation function of the hidden neurons and output neurons to construct a complete motion control model based on the sensing-driving integrated network.

[0015] Combined with the first aspect, in some implementation manners of the first aspect, the multi-constraint dynamic graph optimization model is expressed as R=(r,X), r represents the restricted configuration of the robot indicated by the rooted tree, and X is the neural network weight of the motion control model.

[0016] Combined with the first aspect, in some implementations of the first aspect, in the multi-constraint dynamic graph optimization model, the orientation observation characteristic of the sensing unit is expressed as ROOT is the root node module, Child(ROOT) represents the child module of the root node ROOT, Ori represents the dihedral angle between the two, representing the upper and lower limits of the observation range of the sensing unit.

[0017] Combined with the first aspect, in some implementations of the first aspect, the anti-collision condition of the module unit is:

[0018]

[0019] ε represents the collision penalty threshold, n is the number of module units, cp i j is the collision penalty coefficient between module units i and j, satisfying:

[0020]

[0021] SE i is the safety distance ellipse of module unit i, p j is the position of module unit j, sd i is the safety distance of module unit i, d ij is the distance between the mass centers of module units i and j.

[0022] Combined with the first aspect, in some implementations of the first aspect, the cost and quality constraint is

[0023] ω N N(r)+ω M M(r)<δ

[0024] N(r) is the number of actuators in configuration r; M(r) is the total mass of configuration r, ω N 、ω M are the weights corresponding to the actuator cost and quality constraints, and δ is the constraint threshold.

[0025] Combined with the first aspect, in some implementations of the first aspect, when the robot task is switched, according to the multi-constraint reconstruction optimization model, with the switched task as the drive, the fitness function of the reconstruction optimization model is constructed as is one or more evaluation models corresponding to a specific task, k 1 ,k 2 ,…,k s are their corresponding weights; the evaluation models include at least one of the following: the optimal reconstruction root node, the shortest motion path, the highest variable configuration optimization efficiency, and the fewest configuration change modules;

[0026] The optimal configuration and motion parameters of the robot are searched for the optimal individual R by maximizing the fitness function * Satisfy:

[0027]

[0028] Ω is the search space of the multi-constraint dynamic graph optimization model, including the restricted configuration space and the motion control space, and Φ is the constraint set.

[0029] Combined with the first aspect, in some implementation manners of the first aspect, the configuration-motion graph model gene encoding is performed according to the non-dominated sorting genetic algorithm, including a configuration genome and a motion control genome; the configuration genome includes module genes and connection genes, which respectively describe module attributes and inter-module connection attributes; the motion control genome includes node genes and connection genes, which respectively describe neuron node attributes and inter-node connection attributes.

[0030] Combined with the first aspect, in some implementation manners of the first aspect, the non-dominated sorting genetic algorithm with an elite strategy is used as a framework for optimization and solution, and the process is as follows:

[0031] Randomly generate a configuration-motion population P = {R 1 , R 2 , …, R m}, where Ri is the encoded representation of the configuration-motion reconstruction scheme, i = 1~m, and m is the population size;

[0032] The population individuals execute the task process, evaluate the pros and cons of the individuals according to the fitness function and record the scores F p = {F 1 , F 2 , …, F m};

[0033] Perform basic operator operations in sequence to iterate out a new population P next , and the basic operator operations include selection, crossover, mutation, and cloning;

[0034] Repeat the above process until the termination condition is reached, and select the optimal individual R * = (r * , X * ) as the optimal configuration-motion reconstruction scheme.

[0035] Compared with the prior art, the present invention takes into account the unstructured lunar landform with unknown structure and shape, the restrictive constraints of robot self-reconfiguration, the environmental terrain adaptability requirements of the robot for multi-tasks, the computing power constraints and engineering practice experience. Especially for complex areas such as lunar rilles and lunar lava tubes on the moon surface, which are unstructured, have less prior information and high scientific exploration value, a collaborative configuration optimization design method for intelligent self-reconfigurable robots is proposed. The solution provided by the present invention at least includes the following beneficial technical effects:

[0036] (1) It has a unique insight into dealing with restrictive reconfiguration optimization problems. It uses a graph model to uniformly represent the configuration-motion model, and explores the organic combination of evolutionary robotics theory and graph optimization method, enabling the robot to self-organize and emerge ideal configurations and motion patterns during the interaction with the environment, so as to efficiently solve the dynamic graph optimization problem under local over-constraint in the topological space;

[0037] (2) It has a unique method for dealing with non-linear optimization problems. The configuration-motion optimization problem is a non-linear optimization problem in mathematical description, and it is difficult to directly solve. The non-dominated sorting genetic algorithm with elitist strategy is used to solve the configuration-motion optimization. This method has the characteristics of swarm intelligence optimization, can be unrestricted by the nature of the problem and achieve global optimization, and has high solution robustness;

[0038] (3) It has strong computing real-time performance and high configuration optimization efficiency. The root node optimization efficiency of the intelligent modular robot self-reconfiguration is better than 90%, which can effectively and efficiently generate on-line configuration reconfiguration schemes, and has extremely strong real-time performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 It is a schematic diagram of the principle of a collaborative configuration optimization method for an intelligent self-reconfigurable robot.

[0040] Figure 2 It is a schematic diagram of the design of the basic module library of the self-reconfigurable robot.

[0041] Figure 3 It is a flow chart of the configuration-motion optimization of the self-reconfigurable robot.

[0042] Figure 4 It is a schematic diagram of the genetic encoding of the configuration-motion graph structure of the intelligent self-reconfigurable robot.

[0043] Figure 5 Schematic diagram of the configuration optimization efficiency of the intelligent self-reconfigurable robot. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0044] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0045] The overall configuration of the self-reconfiguring robot and the system motion are mutually coupled. The robot's motion mode highly depends on the configuration, resulting in an exponential increase in the computational complexity and amount of calculation for the reconfiguration plan, making it difficult to ensure real-time performance. Moreover, the constraints on the observation angle and range of the sensors impose requirements on the observability and controllability of the perception system. As Figure 1 shown, the present invention provides an intelligent collaborative configuration optimization method for self-reconfiguring robots to address the above difficulties, and the specific steps are as follows.

[0046] Step S1: Use a rooted tree structure to represent the restricted configuration of the robot.

[0047] Figure 2 It is a schematic diagram of the basic module library of the self-reconfiguring robot. The self-reconfiguring robot is restricted by various constraints during the reconfiguration process, such as: the directional observability of the sensing unit, anti-collision conditions, cost and quality constraints, etc. In terms of robot configuration representation, a rooted tree structure (Rooted Tree) is used to represent the configuration of the modular robot. There is a parent-child relationship between the nodes of the rooted tree, and the root node has no parent node. Since the root node module in the rooted tree is relatively fixed during the reconfiguration motion and the modules adjacent to it cannot be reconfigured by connecting and separating, the sensing unit with directional observability is set as the root node module, thereby constructing a complete restricted configuration description model. The restricted configuration of the modular robot can be described in the form of a graph (Graph) mathematically as

[0048] γ(M) = G(V, E) (1)

[0049] where γ is the rooted tree structure, G is the environmental representation function; M is the original configuration modeled into a network using network technology; V is the set of faces or module units in the grid M; E is the dual edge set, which can be expressed as E = {(f, g)}, and f, g are two adjacent faces in the grid M.

[0050] Step S2: Construct a sensing-driving integrated neural network for robot motion control representation.

[0051] In terms of motion model representation, a sensing-driving integrated artificial neural network (Artificial Neural Networks, ANNs) is constructed to describe the robot motion control model. Specifically as follows: The sensors in the perception system (including installation location, measurement data, etc.) are used as the input neurons of the neural network, and the drivers of the motion system (including motor position, output torque, etc.) are used as the output neurons. The Sigmoid function is selected as the activation function for the hidden neurons and output neurons, thereby constructing a complete motion control model based on the sensing-driving integrated network.

[0052] Step S3: Construct a multi-constraint dynamic graph optimization model to uniformly represent the restrictive configuration and motion control model of the robot.

[0053] For configuration representation, a rooted tree is used to represent the restrictive configuration of the modular robot, and for motion representation model, a sensing-driving integrated artificial neural network is used. These two representation methods can be separately represented by graph structures. Therefore, establishing a multi-constraint graph optimization model can represent its restrictive configuration and motion control model within a unified framework. The restrictive configuration is used as the collaborative input and constraint of motion representation for unified graph structure representation, as Figure 3 shown.

[0054] The present invention uses a graph structure to uniformly represent the restrictive configuration and motion control model. The restrictive configuration - motion control unified model of the modular robot can be expressed as R = (r, X), where R is the graph model of the restrictive configuration - motion control unified model, r represents the robot configuration represented by a rooted tree, and X is the parameter of the motion controller, which represents the neural network weights here. Thus, the reconstruction optimization problem can be described as a dynamic graph optimization problem. Based on the configuration - motion graph representation model, the multi-constraint conditions of the modular robot reconstruction optimization problem can be described as follows.

[0055] The orientation observation characteristic of the sensing unit is

[0056]

[0057] where ROOT represents the root node module, Child(ROOT) represents the child module of the root node ROOT, Ori represents the dihedral angle between the two, represents the upper and lower limits of the observation range of the sensing unit.

[0058] The anti-collision condition of the module unit is

[0059]

[0060] where ε represents the collision penalty threshold, n is the number of module units, and cp ij is the collision penalty coefficient between module units i and j, which can be expressed as

[0061]

[0062] where SE i is the safety distance ellipse of module unit i, p j is the position of module unit j, sd i is the safety distance of module unit i, and d ij is the distance between the mass centers of module units i and j.

[0063] The cost and quality constraint is

[0064] ω N N(r) + ω M M(r) < δ (5)

[0065] Among them, N(r) is the number of drivers in configuration r. Due to the high cost of drivers, it needs to be restricted; M(r) is the total mass of configuration r, ω N 、ω M are the weights corresponding to the cost and mass constraints of the driver, and δ is the constraint threshold.

[0066] Step S4: Construct a fitness function for reconstruction optimization according to the evaluation function corresponding to the task.

[0067] When the modular robot task needs to be switched, the configuration - motion scheme needs to be re - optimized. According to the above - mentioned graph model, a complete multi - constraint reconstruction optimization model is constructed. Driven by the switched task, the configuration - motion scheme is optimized. The fitness function for constructing the reconstruction optimization model according to the evaluation function corresponding to the task is

[0068]

[0069] Among them, is one or more evaluation models corresponding to a specific task, k 1 , k 2 , …, k s are their corresponding weights. The evaluation model can be selected according to the task requirements. Under the satisfaction of collision constraints, motion constraints, and sensor orientation constraints, options such as the optimal reconstruction root node, the shortest motion path, the highest efficiency of variable - configuration optimization, and the fewest configuration - changing modules can be selected as the evaluation model. The evaluation model can include multiple items for simultaneous optimization, which is specifically determined according to the environmental terrain and task attributes. The evaluation model is the optimization criterion for the optimization objective function and also the optimization constraint function for solving the optimization problem constructed in step S3.

[0070] The optimal configuration and motion parameters of the robot will be searched by maximizing the fitness function

[0071]

[0072] Among them, R is the reconstruction scheme graph model defined previously, Ω is the reconstruction scheme search space, including the restricted configuration space and the motion control space, Φ is the constraint set, and Φ is the constraint set.

[0073] Step S5: Use the non - dominated sorting genetic algorithm for configuration - motion optimization solution.

[0074] Such as Figure 4As shown, the non-dominated sorting genetic algorithm needs to perform gene coding for the configuration-motion graph model, mainly including the configuration genome and the motion control genome, which are the basic components of the gene coding graph. The configuration genome includes module genes and connection genes, which describe module attributes and inter-module connection attributes respectively; the motion control genome includes node genes and connection genes, which describe neuron node attributes and inter-node connection attributes respectively. Gene attributes will include the necessary information required for the optimization model. For example, module attributes will include information such as module numbers and module types (sensing modules, driving or driven modules); node connection attributes will include information such as node numbers, the numbers of the nodes connected to them, weights, and whether they are activated.

[0075] In the configuration genes, the module genes mainly include the motion root node, each driver, each joint connector, etc., and the connection genes mainly include the numbers of the parent node and the child node, the parent-child connection surface, the parent-child connection method, etc., which are the constituent factors for the optimization of the configuration gene graph. In the motion genes, the node genes mainly include node numbers, each sensor, the neuron nodes to which they belong, etc., and the connection genes include node numbers, weight coefficients, activation factors, input and output nodes, etc., which are the constituent factors for the optimization of the motion gene graph. The configuration genes and the motion genes are added to a unified neural network structure for unified optimization and solution.

[0076] To solve the above configuration-motion optimization problem, which is a non-linear optimization (NP) problem in mathematical description and is difficult to solve directly. To address this problem, the non-dominated sorting genetic algorithm with elitist strategy (NSGA-II) is used as the framework to solve the configuration-motion optimization model. NSGA-II has the characteristics of swarm intelligence optimization, can be unrestricted by the nature of the problem and achieve global optimality, and has high solution robustness. This method uses the survival rule of "survival of the fittest" to solve, and it is necessary to screen and initialize the population, select superior individuals and generate superior offspring, randomly exchange gene segments of the population to produce new gene combinations, and then randomly change individual genes according to the mutation rate to maintain the diversity of the population and prevent falling into local optimality or premature convergence.

[0077] According to the "survival of the fittest" principle of biological evolution, basic operators such as selection, crossover, mutation, and cloning are defined to respectively achieve: selecting superior individuals and generating superior offspring in the configuration-motion population; randomly exchanging gene segments of the population to produce new gene combinations; randomly changing individual genes according to the mutation rate to maintain the diversity of the population and prevent premature convergence; cloning the genes of some individuals. The process of population initialization is as follows.

[0078] Step S51, randomly generate a configuration-motion population P = {R 1 , R 2 , …, R m}, where R = (r, X) is the coding representation of the configuration-motion reconstruction scheme, and m is the population size;

[0079] In step S52, the individuals of the population will execute a series of task processes, evaluate the performance of the individuals according to the fitness function, and record the score F p ={F 1 , F 2 , …, F m};

[0080] In step S53, the basic operator operations are performed in sequence to iteratively generate a new population P next ;

[0081] In step S54, the above process is repeated until the task is completed or a certain termination condition is reached, and the optimal individual R is selected from the final population * =(r * , X * ), which is the optimal configuration - motion reconstruction scheme.

[0082] The collaborative configuration optimization method in this aspect uses a multi - constraint dynamic graph optimization model to uniformly represent the configuration structure and motion control, and then uses a non - dominated sorting genetic algorithm with an elite strategy for optimization and solution to obtain the optimal configuration optimization scheme, improving the configuration optimization efficiency of the self - reconstructing modular robot, as Figure 5 shown.

[0083] Although the present invention is disclosed above with preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications without departing from the spirit and scope of the present invention. Therefore, the protection scope of the present invention should be determined by the scope defined by the claims of the present invention.

Claims

1. A collaborative configuration optimization method for an intelligent self-reconfigurable robot, characterized in that: include: The rooted tree structure is used to characterize the restrictive configuration of the robot. The restrictive configuration is used to characterize the directional observability of the perception unit, anti-collision conditions, cost and quality constraints. Construct a sensor-actuator integrated neural network to characterize robot motion control and describe the robot motion control model; Construct a multi-constraint dynamic graph optimization model R to unify the representation of the robot's restrictive configuration and motion control model. The restrictive configuration serves as the collaborative input and constraint of the control model. Construct a fitness function for reconstruction optimization based on the evaluation function corresponding to the task; The non-dominated sorting genetic algorithm is used to solve the restrictive configuration-motion control optimization of the multi-constrained dynamic graph optimization model R.

2. The method according to claim 1, characterized in that The restrictive configuration of the modular robot is described as γ(M)=G(V,E), where γ is a rooted tree structure and G is the environment characterization function; M is the original configuration modeled into a network using networking technology; V is the set of faces or module units in the grid M; E is the dual edge set, expressed as E={(f,g)}, and f and g are two adjacent faces in the grid M.

3. The method according to claim 1, characterized in that The robot motion control model includes: The installation position and measurement data of the sensors in the perception system are used as the input neurons of the neural network. The motor position and output torque of the driver of the motion system are used as the output neurons. The Sigmoid function is selected as the activation function of the hidden neurons and the output neurons to construct a complete motion control model based on the sensor-drive integrated network.

4. The method according to claim 1, characterized in that: The multi-constraint dynamic graph optimization model is represented as R = (r, X), where r represents the robot's restrictive configuration indicated by a rooted tree, and X is the neural network weight of the motion control model.

5. The method according to claim 4, characterized in that In the multi-constrained dynamic graph optimization model, the directional observation characteristics of the perception unit are expressed as θ1<Ori<ROOT,Child(ROOT)> <θ2, ROOT is the root node module, Child(ROOT) represents the child module of the root node ROOT, Ori represents the dihedral angle between the two, θ1 and θ2 represent the upper and lower limits of the observation range of the perception unit.

6. The method according to claim 4, characterized in that The module unit anti-collision conditions are: ε represents the collision penalty threshold, n is the number of module units, cp ij is the collision penalty coefficient between module units i and j, satisfying: SE i is the safety distance ellipse of module unit i, p j is the position of module unit j, sd i is the safety distance of module unit i, d ij is the distance between the centroids of module units i and j.

7. The method according to claim 4, characterized in that The cost and quality constraints are oh N N(r)+ω M M(r)<δ N(r) is the number of actuators in configuration r; M(r) is the total mass of configuration r, ω N ,ω M is the weight corresponding to the driver cost and quality constraint, and δ is the constraint threshold.

8. The method according to claim 1, characterized in that When the robot switches tasks, the optimization model is reconstructed according to the multi-constraints, and the fitness function of the reconstructed optimization model is constructed based on the switched tasks. is one or more evaluation models corresponding to a specific task, k1, k2, …, k s is the corresponding weight; the evaluation model includes at least one of the following: optimal reconstruction root node, shortest motion path, highest configuration optimization efficiency, and least configuration change modules; The optimal configuration and motion parameters of the robot are searched for the optimal individual R by maximizing the fitness function * satisfy: St Φ Ω is the search space of the multi-constraint dynamic graph optimization model, which includes the restrictive configuration space and the motion control space, and Φ is the constraint set.

9. The method according to claim 1, characterized in that: The configuration-motion graph model gene encoding is executed according to the non-dominated sorting genetic algorithm, including the configuration genome and the motion control genome; the configuration genome includes module genes and connection genes, which respectively describe the module attributes and the connection attributes between modules; the motion control genome includes node genes and connection genes, which respectively describe the neuron node attributes and the connection attributes between nodes.

10. The method according to claim 9, characterized in that The non-dominated sorting genetic algorithm with elite strategy is used as the framework for optimization and solution. The process is as follows: Randomly generate configuration-movement population P = {R1, R2, ..., R m }, where R i is the encoding representation of the configuration-motion reconstruction scheme, i = 1~m, m is the population size; The population individuals execute the task process, evaluate the performance of the individuals according to the fitness function and record the score F p ={F1,F2,…,F m }; Perform basic operator operations in sequence to iterate the new population P next ,Basic operator operations include selection, crossover, mutation, and cloning; Repeat the above process until the termination condition is reached, and select the best individual R from the final population * =(r * ,X * ) as the optimal configuration-motion reconstruction scheme.

Images