A multi-robot collaborative grasping method and system based on meta-heuristic algorithm

Abstract

The present invention provides a multi-robot collaborative grasping method and system based on a metaheuristic algorithm. The method comprises: performing camera calibration on at least two robots to obtain geometric model parameters; performing active visual detection to select the optimal grasping object and obtain visual features; calculating the target posture deviation of the robot manipulator; performing deep reinforcement learning in an offline and online manner to optimize and converge the target posture deviation of the manipulator; inputting the deviation into a PID controller to output the manipulator's torque, then using the target posture deviation as the input signal of a self-organizing neural network and the manipulator's torque as the target parameter of a deep belief network, simultaneously training two serial neural networks to construct a deep learning controller; and finally, guiding the manipulator to a designated position to grasp the target through the negative feedback mechanism of the feedback controller. The present invention achieves effective active visual recognition and metaheuristic optimization control, improving system stability.

Description

Technical Field

[0001] The present invention relates to the technical field of robot intelligent control, and in particular to a multi-robot collaborative grasping method and system based on a meta-heuristic algorithm. Background Art

[0002] In recent years, the development of robotics technology has gradually alleviated my country's problems of intensive manual labor, an aging population, and difficulty recruiting workers. Robotic grasping technology, which involves retrieving target items one by one from a pile of disordered items, is a key link in automated scenarios such as logistics sorting, machine tool loading and unloading, and palletizing. It can reduce workers' workload, improve work efficiency, and operate 24 hours a day. In industrial production, robotic grasping is mainly achieved through pre-programming, lacking autonomy and the ability to adapt to complex environments, making it difficult to meet the needs of modern industrial intelligent manufacturing and flexible manufacturing. Grasping operations, as a basic skill of robots, play a very important role in the road to robot intelligence. If robots are to be used to replace humans to complete various tasks, they must be able to autonomously complete more complex grasping operations.

[0003] Therefore, the study of intelligent robotic grasping technology in unstructured environments has become an important topic. Based on the calculation method of the grasping pose of the target object, robotic grasping methods are divided into two categories: analytical and data-driven. Analytical robotic grasping methods require the acquisition of an accurate geometric model of the object and the dynamic model of the mechanical gripper in advance. Then, a complex target optimization function is established through multiple constraints to finally solve the optimal grasping pose of the target object. However, in real-world robotic grasping applications, these solution conditions are often difficult to meet, which limits the widespread application of this grasping method. Vision is known as the "eye of wisdom" of robots, and the use of visual recognition methods to obtain the grasping pose of the target object is the main idea of ​​current data-driven grasping algorithms.

[0004] Existing analytical and data-driven robotic grasping methods struggle to accurately and quickly identify targets. They also struggle to effectively construct and organize reliable recognition algorithms and implement them successfully. Real-time performance is also a significant challenge. The low image acquisition speed and the long image processing time introduce significant time lags into the system. Furthermore, the introduction of visual information significantly increases the system's computational workload, such as calculating the image Jacobian matrix and estimating depth information. Image processing speed is one of the main bottlenecks affecting the real-time performance of visual systems. Stability is also a primary consideration for all control systems. For visual control systems, whether position-based, image-based, or hybrid visual servoing methods, they all face the following challenges: How to ensure system stability when the initial point is far from the target point, i.e., how to increase the stable region and ensure global convergence; and how to ensure that feature points remain within the field of view to avoid servo failure.

[0005] Metaheuristic algorithms are simple in concept, easy to implement, require no gradient information, and are applicable to multiple disciplines. They are now widely used to solve practical engineering problems. Nature-inspired metaheuristic algorithms are proposed by simulating biological or physical phenomena. First, an initial population is randomly generated, and then the individuals of each generation are updated according to certain iterative rules. There are three main methods for population update: evolution-based methods, physics-based methods, and population-based methods. These methods achieve the goal of optimization by continuously selecting the best individuals to form the next generation of populations. Modern heuristic algorithms have different optimization mechanisms, but they have similar optimization processes. They all start from a set of initial solutions, set key parameters, generate several solutions according to certain rules, update the current state through acceptance criteria, and perform multiple iterative searches until the optimal result is obtained. Metaheuristic algorithms have been widely used to solve practical engineering problems. The existing technology lacks methods and systems that combine metaheuristic algorithms to perform robotic grasping work. Summary of the Invention

[0006] In view of the above-mentioned deficiencies in the prior art, the present invention provides a multi-robot collaborative grasping method and system based on a meta-heuristic algorithm, which effectively solves the problems of the prior art in that it is difficult to accurately and quickly identify targets, and has poor real-time performance and stability.

[0007] In a first aspect, the present invention provides a multi-robot collaborative grasping method based on a meta-heuristic algorithm, comprising the following steps:

[0008] Calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters;

[0009] Based on the visual 3D-2D projection ratio transformation, at least two sets of high-quality images corresponding to the grasping target are provided for active visual detection, the optimal grasping object is selected, and visual features are obtained;

[0010] Calculating the positions of the manipulators of the first robot and the second robot relative to the cooperation target in real time based on the geometric model parameters and the visual features to obtain target position deviations of the manipulators;

[0011] Using offline + online robot-assisted guidance control technology, deep reinforcement learning is performed on the target posture deviation of the manipulator to optimize and converge the target posture deviation of the manipulator;

[0012] The target posture deviation of the manipulator is input into the PID controller to output the torque of the manipulator. The target posture deviation serves as the input signal of the self-organizing neural network, and the torque of the manipulator serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller.

[0013] The target posture deviation of the manipulator is used as the input of the deep learning controller, and the torque generated by the deep belief network is used as the output of the deep learning controller. Through the negative feedback mechanism of the feedback controller, the manipulator is guided to reach the specified position for target grasping;

[0014] The first robot and the second robot collaborate to grasp the target. Based on the characteristics of offline training, the first robot can perform active visual detection in real time. Through the technical revision of offline + online robot assisted guidance control, the target posture deviation of the manipulator of the second robot is optimized to grasp the target more accurately.

[0015] Preferably, the calibrating two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters includes:

[0016] The active vision camera calibration method is used to determine the geometric model parameters, including camera intrinsic parameters and camera extrinsic parameters. The camera intrinsic parameters are used to determine the projection relationship of the camera from three-dimensional space to two-dimensional image, and the camera extrinsic parameters are used to determine the relative position relationship between the camera coordinates and the world coordinate system.

[0017] Preferably, the method provides at least two sets of high-quality images corresponding to the grasping targets based on the visual 3D-2D projection ratio transformation, performs active visual detection, selects the optimal grasping object, and obtains visual features, including the following steps:

[0018] Active vision predicts areas where there may be targets on the feature map, and limits the features to a fixed size through the region of interest pooling method. It recognizes the features of the region of interest through convolution operations, detects potential target areas with high reliability, and excludes background areas where there are no targets, performing multi-class recognition, reducing the computational complexity of the recognition process, and ensuring robust pose estimation.

[0019] Preferably, during the training process of the active visual detection, focal loss is used to perform stochastic gradient descent optimization, so as to balance the ratio of positive and negative samples and optimize the training process.

[0020] Preferably, the offline + online robot assisted guidance control technology is used to perform deep reinforcement learning on the target posture deviation of the manipulator to optimize and converge the target posture deviation of the manipulator, including:

[0021] The results of offline learning are used to guide deep reinforcement learning. The guided policy search uses differential dynamic programming to generate guidance samples, and uses regularized important sample strategy optimization to merge the guidance samples into the policy search to avoid deep reinforcement learning falling into local optimality in the case of multiple parameters.

[0022] Preferably, the guidance samples are generated by at least one of offline planning and manual demonstration.

[0023] Preferably, the target posture deviation of the manipulator is input into a PID controller to output the torque of the manipulator, the target posture deviation is used as the input signal of the self-organizing neural network, the torque of the manipulator is used as the target parameter of the deep belief network, and two serial neural networks are trained simultaneously to construct a deep learning controller, comprising the following steps:

[0024] Inputting the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator;

[0025] Inputting the target posture deviation of the manipulator into a self-organizing neural network for training, and dividing the output into different types of regions according to the data clustering characteristics of the network;

[0026] The torque of the manipulator is used as the target parameter of the deep belief network. The data vector is back-propagated using the gradient descent method to modify the weight vectors between the hidden layers in the network layer by layer until the control torque as the bottom state is generated, completing the network training.

[0027] A deep learning controller is constructed by connecting a self-organizing neural network and a deep belief network in series.

[0028] In a second aspect, the present invention provides a multi-robot collaborative grasping system based on a meta-heuristic algorithm, comprising:

[0029] A camera calibration module is used to calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters;

[0030] Active visual detection module, which is used to provide at least two sets of high-quality images corresponding to the grasping target based on the visual 3D-2D projection ratio transformation, perform active visual detection, select the optimal grasping object, and obtain visual features;

[0031] A posture calculation module is used to calculate the postures of the manipulators of the first robot and the second robot relative to the cooperation target in real time based on the geometric model parameters and the visual features, and obtain the target posture deviations of the manipulators;

[0032] A deep reinforcement learning module is used to perform deep reinforcement learning on the target posture deviation of the manipulator using offline + online robot assisted guidance control technology to optimize and converge the target posture deviation of the manipulator;

[0033] A deep learning controller construction module is used to input the target posture deviation of the manipulator into a PID controller to output the manipulator's torque. The target posture deviation serves as the input signal of the self-organizing neural network, and the manipulator's torque serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller.

[0034] The target grasping module is used to use the target posture deviation of the manipulator as the input of the deep learning controller and the torque generated by the deep belief network as the output of the deep learning controller, and guide the manipulator to the specified position for target grasping through the negative feedback mechanism of the feedback controller.

[0035] Preferably, the deep learning controller building module includes:

[0036] A PID controller unit, configured to input the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator;

[0037] A self-organizing neural network training unit is used to input the target posture deviation of the manipulator into the self-organizing neural network for training, and divide the output into different types of areas according to the data clustering characteristics of the network;

[0038] A deep belief network training unit is used to use the torque of the manipulator as the target parameter of the deep belief network, adopt the gradient descent method to backpropagate the data vector, and modify the weight vectors between each hidden layer in the network layer by layer until the control torque as the bottom state is generated, thereby completing the network training;

[0039] A deep learning controller building unit is used to build a deep learning controller using a series of self-organizing neural networks and deep belief networks.

[0040] In a third aspect, the present invention provides a multi-robot collaborative grasping platform based on a meta-heuristic algorithm, the platform comprising a processor, a memory, and a multi-robot collaborative grasping platform control program based on a meta-heuristic algorithm;

[0041] Wherein, the processor executes the multi-robot collaborative grasping platform control program based on the metaheuristic algorithm, the multi-robot collaborative grasping platform control program based on the metaheuristic algorithm is stored in the memory, and the multi-robot collaborative grasping platform control program based on the metaheuristic algorithm implements the multi-robot collaborative grasping method based on the metaheuristic algorithm as described in any one of the first aspects of the present invention.

[0042] The present invention provides a multi-robot collaborative grasping method and system based on a meta-heuristic algorithm. Through the selection of image features, it may sometimes be necessary to switch from one set of features to another set of features for a task, and the global features are combined with the local features to establish a robot-assisted visual system. The technology of offline + online robot-assisted guidance control is adopted. The offline can ensure the real-time and rapid guidance of the manipulator, while the online adapts to personalized production needs. In some more rigid production situations, the offline trained controller can also be used alone, but in the "offline + online" framework, the online learning control is provided with priors and restrictions are added, which can greatly accelerate the convergence of the online process to meet the real-time requirements. The present invention realizes effective active visual recognition and meta-heuristic optimization control, and improves the stability of the overall system. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following is a brief introduction to the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0044] Figure 1 A schematic diagram of the process flow of a multi-robot collaborative grasping method based on a meta-heuristic algorithm provided by an embodiment of the present invention;

[0045] Figure 2 A schematic diagram of a process flow for constructing a deep learning controller according to an embodiment of the present invention;

[0046] Figure 3 A schematic diagram of the structure of a multi-robot collaborative grasping system based on a meta-heuristic algorithm provided by an embodiment of the present invention;

[0047] Figure 4 A schematic diagram of the deep learning controller building block structure provided by an embodiment of the present invention;

[0048] Figure 5 Schematic diagram of the architecture of a multi-robot collaborative grasping platform based on a meta-heuristic algorithm provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0049] To make the purpose, technical solutions and advantages of the present invention clearer, the technical solutions of the present invention will be further clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. It should be noted that the embodiments described are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0050] In the embodiments of the present application, the term "and / or" is merely a description of the association relationship between associated objects, indicating that three relationships may exist. For example, A and / or B may represent three situations: A exists alone, A and B exist at the same time, and B exists alone.

[0051] The terms "first" and "second" in the embodiments of the present application are only used for descriptive purposes and are not to be understood as indicating or implying relative importance or implicitly indicating the number of the indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include at least one of the features. In the description of the present application, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusions. For example, a system, product or device comprising a series of components or units is not limited to the listed components or units, but may optionally also include components or units that are not listed, or may optionally also include other components or units that are inherent to these products or devices. In the description of the present application, the meaning of "plurality" is at least two, such as two, three, etc., unless otherwise clearly and specifically defined.

[0052] Existing analytical and data-driven robotic grasping methods struggle to accurately and quickly identify targets. They also struggle to effectively construct and organize reliable recognition algorithms and implement them successfully. Real-time performance is also a significant challenge. The low image acquisition speed and the long image processing time introduce significant time lags into the system. Furthermore, the introduction of visual information significantly increases the system's computational workload, such as calculating the image Jacobian matrix and estimating depth information. Image processing speed is one of the main bottlenecks affecting the real-time performance of visual systems. Stability is also a primary consideration for all control systems. For visual control systems, whether position-based, image-based, or hybrid visual servoing methods, they all face the following challenges: How to ensure system stability when the initial point is far from the target point, i.e., how to increase the stable region and ensure global convergence; and how to ensure that feature points remain within the field of view to avoid servo failure.

[0053] Metaheuristic algorithms are simple in concept, easy to implement, do not require gradient information, and are applicable to multiple disciplines. They are now widely used to solve practical engineering problems. However, the existing technology lacks methods and systems that combine metaheuristic algorithms to perform robotic grasping.

[0054] In an embodiment of the present invention, a multi-robot collaborative grasping method and system based on a meta-heuristic algorithm are provided to solve the above technical problems.

[0055] Figure 1 The schematic flow chart of the multi-robot collaborative grasping method based on the meta-heuristic algorithm provided in the embodiment of the present invention mainly includes:

[0056] Step S1: calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters;

[0057] In image measurement and machine vision applications, a geometric model of camera imaging must be established to determine the relationship between the 3D geometric position of a point on a surface and its corresponding point in the image. These geometric model parameters are known as camera parameters. In most cases, these parameters (intrinsic parameters, extrinsic parameters, and distortion parameters) must be determined through experimentation and calculation. This process of determining these parameters is called camera calibration.

[0058] The camera motion trajectory information is known to calibrate the camera. No calibration object is required, but the camera needs to be controlled to perform certain special movements. The special characteristics of this movement can be used to calculate the camera's internal parameters.

[0059] This embodiment of the present invention uses an active vision camera calibration method. This method does not require a calibration object, but rather requires controlling the camera to perform certain specific motions. The specificity of these motions can be exploited to calculate the camera's internal parameters. The advantages of active vision-based camera calibration are its simple algorithm and the ability to often obtain a linear solution, resulting in high robustness.

[0060] The geometric model parameters to be determined during camera calibration are divided into two types: intrinsic parameters and extrinsic parameters. The function of camera intrinsic parameters is to determine the projection relationship of the camera from 3D space to 2D image, while the camera extrinsic parameters are used to determine the relative position relationship between camera coordinates and the world coordinate system. Among them, there are 6 camera intrinsic parameters: f, k, Sx, Sy, Cx, Cy, where: f is the focal length; k represents the magnitude of radial distortion. If k is negative, the distortion becomes barrel distortion, and if it is positive, the distortion becomes pincushion distortion; Sx and Sy are scaling factors. Generally, the camera imaging unit is not strictly rectangular, and its size in the horizontal and vertical directions is inconsistent. This results in different scaling factors in the X and Y directions, so two scaling factors need to be defined separately. For pinhole cameras, they represent the distance between adjacent pixels in the horizontal and vertical directions on the image sensor; Cx and Cy are the principal points of the image, that is, the intersection of the imaging plane and the image plane perpendicular to the lens axis. For pinhole cameras, this point is the vertical projection of the projection center on the imaging plane and is also the center of radial distortion.

[0061] There are 6 parameters of camera extrinsics: α, β, γ, Tx, Ty, Tz. The relationship between camera coordinates and world coordinates can be expressed as: P c =RP w +T, where P w is the world coordinate, P c are the camera coordinates; where T = (Tx, Ty, Tz) is the translation vector, and R = R(α, β, γ) is the rotation matrix, which are the rotation angles γ, β, and α around the z-axis, β, and x-axis of the camera coordinate system, respectively; these six parameters constitute the camera extrinsic parameters.

[0062] Step S2: Based on the visual 3D-2D projection ratio transformation, provide at least two sets of high-quality images corresponding to the grasping target, perform active visual detection, select the optimal grasping object, and obtain visual features;

[0063] Based on the visual 3D-2D projection scale transformation, the collaborative robot vision system is provided with at least two sets of high-quality images corresponding to the auxiliary grasping targets (reference objects), enabling the robot to use meta-heuristic algorithms to select the optimal grasping object from discrete manufacturing targets.

[0064] First, active vision predicts areas on the feature map where there may be targets and limits the features to a fixed size through the region of interest pooling method; then, the recognition process identifies the features of the region of interest through convolution operations. Through this two-level screening strategy of category prediction, the algorithm retains potential target areas with high reliability and excludes most background areas in the image where there are no targets. Finally, multi-category recognition is performed, which can significantly reduce the computational complexity of the recognition process and ensure robust pose estimation.

[0065] During the training process, focal loss is used for stochastic gradient descent optimization to balance the ratio of positive and negative samples and optimize the training process.

[0066] Step S3, calculating the positions of the manipulators of the first robot and the second robot relative to the cooperation target in real time based on the geometric model parameters and the visual features, and obtaining the target position deviation of the manipulators;

[0067] Step S4: Using offline + online robot-assisted guidance control technology, deep reinforcement learning is performed on the target posture deviation of the manipulator to optimize and converge the target posture deviation of the manipulator;

[0068] Deep reinforcement learning methods have achieved tremendous success in tasks such as video games and board games, simulated control tasks, and robotics based on raw visual input. A key approach behind this success is online policy deep reinforcement learning methods. These methods select actions based on the current policy to interact with the environment and optimize the policy using the online policy data generated by the agent's interaction with the environment, thereby solving the corresponding sequential decision-making problems. Online policy deep reinforcement learning methods can stably optimize the policy using the online policy data generated by the current policy, which makes them stable during the interaction between the agent and the environment, enabling them to solve complex control problems. However, online policy deep reinforcement learning methods suffer from inefficient data utilization because they can only use the data generated by the current policy once.

[0069] Off-policy deep reinforcement learning methods address the challenges faced by online policy methods by optimizing policies using offline policy data. By leveraging offline policy data, these methods reduce the online policy interactions between the agent and the environment, resulting in improved performance in real-world applications. These off-policy deep reinforcement learning methods can be divided into two categories based on how they optimize policies: value-based and policy-based. Policy-based off-policy deep reinforcement learning methods directly optimize the policy function using offline policy data generated by the behavioral policy. Because these methods directly learn the policy function over the entire action space, they are suitable for sequential decision-making problems in continuous action spaces.

[0070] In an embodiment of the present invention, the results of offline learning are used to guide meta-heuristic algorithm optimization of deep reinforcement learning. The guided policy search uses differential dynamic programming to generate suitable guidance samples. The guidance samples can also be generated by offline planning or manual demonstration. A novel regularization term is used to regularize the important sample policy optimization to merge these guidance samples into the policy search, so as to avoid deep reinforcement learning from falling into local optimality in the case of multiple parameters (hundreds).

[0071] This embodiment of the present invention utilizes a combination of offline and online robot-assisted guidance control. Offline control ensures real-time and rapid robot guidance, while online control accommodates personalized production needs. While offline training can be used independently in certain rigid production scenarios, the "offline + online" framework provides priors and additional constraints for online learning control, significantly accelerating the convergence of the online process to meet real-time requirements.

[0072] Step S5: Input the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator. The target posture deviation serves as the input signal of the self-organizing neural network, and the torque of the manipulator serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller.

[0073] The multi-manipulator collaborative deep learning controller will mainly be composed of meta-heuristic algorithm optimization, and its training process will be mainly based on the existing PID controller.

[0074] Figure 2 A flowchart of a method for constructing a deep learning controller according to an embodiment of the present invention is shown in FIG. Figure 2 The specific steps are as follows:

[0075] Step S51: input the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator;

[0076] Step S52: input the target posture deviation of the manipulator into a self-organizing neural network for training, and divide the output into different types of regions according to the data clustering characteristics of the network;

[0077] The manipulator's target pose deviation is fed into a self-organizing neural network. Based on the network's inherent data clustering properties, the output is then divided into different types of regions. Using the self-organizing neural network, the three-dimensional information obtained through initial visual perception can be mined and clustered in depth, making the desired manipulator torque more robust and robust. This is something that conventional two-dimensional perception information cannot achieve within a two-dimensional plane.

[0078] Step S53: Using the torque of the manipulator as the target parameter of the deep belief network, backpropagating the data vector using the gradient descent method, and modifying the weight vectors between the hidden layers in the network layer by layer until the control torque as the bottom state is generated, thus completing the network training;

[0079] During the deep belief network training process, the torque of the robot (i.e., the output signal of the PID controller) is used as its target parameter. The gradient descent method is used to backpropagate the data vector, and the weight vectors between the hidden layers in the network are modified layer by layer until the control torque as the underlying state is generated, and the network training is completed accordingly.

[0080] A deep learning controller is constructed by connecting a self-organizing neural network and a deep belief network in series.

[0081] After the two networks are trained and stabilized, a deep learning controller can be constructed by connecting the self-organizing neural network and the deep belief network in series to replace the traditional PID controller.

[0082] S6. Use the target posture deviation of the manipulator as the input of the deep learning controller, and the torque generated by the deep belief network as the output of the deep learning controller. Through the negative feedback mechanism of the feedback controller, guide the manipulator to reach the specified position for target grasping.

[0083] The target posture deviations of all manipulators are input into the offline deep learning controller to obtain the initial manipulator torque. This is then adjusted using an online deep reinforcement learning controller based on the target posture deviations. Combined with the correction of the torque feedback from the torque observer, the actual torque of all manipulators is finally output to assist in guiding the operation.

[0084] The first and second robots collaborate to grasp the target. Based on offline training features, the first robot performs real-time active visual detection and optimizes the target pose deviation of the second robot's manipulator through offline and online robot-assisted guidance and control technology to complete the target grasp. Simultaneously, the second robot also performs real-time active visual detection to guide the first robot to accurately grasp the target.

[0085] Based on the multi-robot collaborative grasping method based on the meta-heuristic algorithm in the above embodiment, a second aspect of the embodiment of the present invention further provides a multi-robot collaborative grasping system based on the meta-heuristic algorithm.

[0086] Figure 3 The schematic diagram of the structure of a multi-robot collaborative grasping system based on a meta-heuristic algorithm provided by an embodiment of the present invention mainly includes:

[0087] A camera calibration module 310 is used to calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters;

[0088] Active visual detection module 320, used to provide at least two sets of high-quality images corresponding to the grasping target based on the visual 3D-2D projection ratio transformation, perform active visual detection, select the optimal grasping object, and obtain visual features;

[0089] A posture calculation module 330 is used to calculate the postures of the manipulators of the first robot and the second robot relative to the cooperation target in real time based on the geometric model parameters and the visual features, and obtain the target posture deviations of the manipulators;

[0090] A deep reinforcement learning module 340 is used to perform deep reinforcement learning on the target posture deviation of the manipulator using offline + online robot-assisted guidance control technology to optimize and converge the target posture deviation of the manipulator;

[0091] A deep learning controller construction module 350 is used to input the target posture deviation of the manipulator into a PID controller to output the manipulator's torque. The target posture deviation serves as the input signal of the self-organizing neural network, and the manipulator's torque serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller.

[0092] The target grasping module 360 ​​is used to use the target posture deviation of the manipulator as the input of the deep learning controller and the torque generated by the deep belief network as the output of the deep learning controller, and guide the manipulator to reach the specified position for target grasping through the negative feedback mechanism of the feedback controller.

[0093] Figure 4 This is a schematic diagram of the deep learning controller building block structure provided by an embodiment of the present invention. The deep learning controller building block includes:

[0094] The PID controller unit 351 is used to input the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator;

[0095] A self-organizing neural network training unit 352 is used to input the target posture deviation of the manipulator into the self-organizing neural network for training, and divide the output into different types of regions according to the data clustering characteristics of the network;

[0096] A deep belief network training unit 353 is used to use the torque of the manipulator as the target parameter of the deep belief network, adopt the gradient descent method to backpropagate the data vector, and modify the weight vectors between each hidden layer in the network layer by layer until the control torque as the bottom state is generated, thereby completing the network training;

[0097] The deep learning controller construction unit 354 is used to construct a deep learning controller using a self-organizing neural network and a deep belief network connected in series.

[0098] Based on the same concept, an embodiment of the present invention also provides a multi-robot collaborative grasping platform based on a meta-heuristic algorithm. Figure 5 Schematic diagram of the architecture of a multi-robot collaborative grasping platform based on a metaheuristic algorithm provided by an embodiment of the present invention. The multi-robot collaborative grasping platform based on a metaheuristic algorithm has a built-in processor, which can be composed of an integrated circuit, for example, a single packaged integrated circuit, or a plurality of integrated circuits with the same or different functions, including one or more central processing units (CPUs), microprocessors, digital processing chips, graphics processors, and a combination of various control chips. The processor uses various interfaces and lines to connect various components, and executes or executes programs or units stored in the memory, and calls data stored in the memory to execute the steps of the multi-robot collaborative grasping method based on the metaheuristic algorithm, for example, including:

[0099] Calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters;

[0100] Based on the visual 3D-2D projection ratio transformation, at least two sets of high-quality images corresponding to the grasping target are provided for active visual detection, the optimal grasping object is selected, and visual features are obtained;

[0101] Calculating the positions of the manipulators of the first robot and the second robot relative to the cooperation target in real time based on the geometric model parameters and the visual features to obtain target position deviations of the manipulators;

[0102] Using offline + online robot-assisted guidance control technology, deep reinforcement learning is performed on the target posture deviation of the manipulator to optimize and converge the target posture deviation of the manipulator;

[0103] The target posture deviation of the manipulator is input into the PID controller to output the torque of the manipulator. The target posture deviation serves as the input signal of the self-organizing neural network, and the torque of the manipulator serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller.

[0104] The target posture deviation of the manipulator is used as the input of the deep learning controller, and the torque generated by the deep belief network is used as the output of the deep learning controller. Through the negative feedback mechanism of the feedback controller, the manipulator is guided to reach the specified position for target grasping.

[0105] The memory is used to store program codes and various data. It is installed in a multi-robot collaborative grasping platform based on meta-heuristic algorithms and can automatically access programs or data at high speed during operation.

[0106] The memory includes read-only memory (ROM), random access memory (RAM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), one-time programmable read-only memory (OTPROM), electronically erasable programmable read-only memory (EEPROM), compact disc read-only memory (CD-ROM) or other optical disc storage, magnetic disk storage, magnetic tape storage, or any other computer-readable medium that can be used to carry or store data.

[0107] To achieve the above-mentioned purpose, the present invention also provides a computer-readable storage medium, which stores a multi-robot collaborative grasping platform control program based on a metaheuristic algorithm. The multi-robot collaborative grasping platform control program based on a metaheuristic algorithm implements the steps of the multi-robot collaborative grasping method based on a metaheuristic algorithm.

[0108] In summary, the present invention provides a multi-robot collaborative grasping method and system based on a meta-heuristic algorithm. Through the selection of image features, it may sometimes be necessary to switch from one set of features to another for a task, and the global features are combined with the local features to establish a robot-assisted visual system. The technology of offline + online robot-assisted guidance control is adopted. The offline method can ensure the real-time and rapid guidance of the manipulator, while the online method adapts to the personalized production needs. In some relatively rigid production situations, the offline trained controller can also be used alone, but under the "offline + online" framework, the online learning control is provided with priors and restrictions are added, which can greatly accelerate the convergence of the online process to meet the real-time requirements. The present invention realizes effective active visual recognition and meta-heuristic optimization control, and improves the stability of the overall system.

[0109] References herein to "embodiments" mean that a particular feature, structure, or characteristic described in connection with the embodiments may be included in at least one embodiment of the present application. The appearance of this phrase in various places in the specification does not necessarily refer to the same embodiment, nor does it constitute an independent or alternative embodiment that is mutually exclusive of other embodiments. It is understood, both explicitly and implicitly, by those skilled in the art that the embodiments described herein may be combined with other embodiments.

[0110] The above-described embodiments merely illustrate several implementations of the present invention, and while their descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the present invention. It should be noted that a person skilled in the art would be able to make numerous variations and improvements without departing from the spirit of the present invention, all of which fall within the scope of protection of the present invention. Therefore, the scope of protection of the present invention shall be determined by the appended claims.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A multi-robot collaborative grasping method based on meta-heuristic algorithm, characterized in that: The steps include: Calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters; Based on the visual 3D-2D projection ratio transformation, at least two sets of high-quality images corresponding to the grasping target are provided for active visual detection, the optimal grasping object is selected, and visual features are obtained; Calculating, in real time, the poses of the manipulators of the first robot and the second robot relative to a cooperative target based on the geometric model parameters and the visual features, and obtaining target pose deviations of the manipulators; Using offline + online robot-assisted guidance control technology, deep reinforcement learning is performed on the target posture deviation of the manipulator to optimize and converge the target posture deviation of the manipulator; The target posture deviation of the manipulator is input into the PID controller to output the torque of the manipulator. The target posture deviation serves as the input signal of the self-organizing neural network, and the torque of the manipulator serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller. The target posture deviation of the manipulator is used as the input of the deep learning controller, and the torque generated by the deep belief network is used as the output of the deep learning controller. Through the negative feedback mechanism of the feedback controller, the manipulator is guided to reach the specified position for target grasping; The first robot and the second robot collaborate to grasp the target. Based on the characteristics of offline training, the first robot performs active visual detection in real time, and optimizes the target posture deviation of the second robot's manipulator through offline + online robot assisted guidance control technology revision to complete target grasping.

2. The multi-robot collaborative grasping method based on meta-heuristic algorithm according to claim 1 is characterized in that: The calibrating of two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters includes: The active vision camera calibration method is used to determine the geometric model parameters, including camera intrinsic parameters and camera extrinsic parameters. The camera intrinsic parameters are used to determine the projection relationship of the camera from three-dimensional space to two-dimensional image, and the camera extrinsic parameters are used to determine the relative position relationship between the camera coordinates and the world coordinate system.

3. The multi-robot collaborative grasping method based on meta-heuristic algorithm according to claim 1 is characterized in that: The method of providing at least two sets of high-quality images corresponding to grasped targets based on visual 3D-2D projection ratio transformation, performing active visual detection, selecting the optimal grasped object, and obtaining visual features includes the following steps: Active vision predicts areas where there may be targets on the feature map, and limits the features to a fixed size through the region of interest pooling method. It recognizes the features of the region of interest through convolution operations, detects potential target areas with high reliability, and excludes background areas where there are no targets, performing multi-class recognition, reducing the computational complexity of the recognition process, and ensuring robust pose estimation.

4. The multi-robot collaborative grasping method based on meta-heuristic algorithm according to claim 3 is characterized in that: During the training process of the active visual detection, focal loss is used for stochastic gradient descent optimization to balance the ratio of positive and negative samples and optimize the training process.

5. The multi-robot collaborative grasping method based on meta-heuristic algorithm according to claim 1 is characterized in that: The offline + online robot assisted guidance control technology is used to perform deep reinforcement learning on the target posture deviation of the manipulator to optimize and converge the target posture deviation of the manipulator, including: The results of offline learning are used to guide deep reinforcement learning. The guided policy search uses differential dynamic programming to generate guidance samples, and uses regularized important sample strategy optimization to merge the guidance samples into the policy search to avoid deep reinforcement learning falling into local optimality in the case of multiple parameters.

6. The multi-robot collaborative grasping method based on meta-heuristic algorithm according to claim 5 is characterized in that: The guidance samples are generated by at least one of offline planning and manual demonstration.

7. The multi-robot collaborative grasping method based on meta-heuristic algorithm according to claim 1 is characterized in that: The target posture deviation of the manipulator is input into a PID controller to output the torque of the manipulator. The target posture deviation is used as an input signal of a self-organizing neural network, and the torque of the manipulator is used as a target parameter of a deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller, including the following steps: Inputting the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator; Inputting the target posture deviation of the manipulator into the self-organizing neural network for training, and dividing the output into different types of regions according to the data clustering characteristics of the network; The torque of the manipulator is used as the target parameter of the deep belief network. The data vector is back-propagated using the gradient descent method to modify the weight vectors between the hidden layers in the network layer by layer until the control torque as the bottom state is generated, thereby completing the network training. A deep learning controller is constructed by connecting the self-organizing neural network and the deep belief network in series.

8. A multi-robot collaborative grasping system based on meta-heuristic algorithm, characterized by: include: A camera calibration module is used to calibrate the two cameras corresponding to the first robot and the second robot of the grasping system to obtain geometric model parameters; Active visual detection module, which is used to provide at least two sets of high-quality images corresponding to the grasping target based on the visual 3D-2D projection ratio transformation, perform active visual detection, select the optimal grasping object, and obtain visual features; A posture calculation module is used to calculate the postures of the manipulators of the first robot and the second robot relative to the cooperation target in real time based on the geometric model parameters and the visual features, and obtain the target posture deviations of the manipulators; A deep reinforcement learning module is used to perform deep reinforcement learning on the target posture deviation of the manipulator using offline + online robot assisted guidance control technology to optimize and converge the target posture deviation of the manipulator; A deep learning controller construction module is used to input the target posture deviation of the manipulator into a PID controller to output the manipulator's torque. The target posture deviation serves as the input signal of the self-organizing neural network, and the manipulator's torque serves as the target parameter of the deep belief network. Two serial neural networks are trained simultaneously to construct a deep learning controller. The target grasping module is used to use the target posture deviation of the manipulator as the input of the deep learning controller and the torque generated by the deep belief network as the output of the deep learning controller, and guide the manipulator to the specified position for target grasping through the negative feedback mechanism of the feedback controller.

9. The multi-robot collaborative grasping system based on meta-heuristic algorithm according to claim 8, characterized in that: The deep learning controller building blocks include: A PID controller unit, configured to input the target posture deviation of the manipulator into the PID controller to output the torque of the manipulator; A self-organizing neural network training unit is used to input the target posture deviation of the manipulator into the self-organizing neural network for training, and divide the output into different types of areas according to the data clustering characteristics of the network; a deep belief network training unit, configured to use the torque of the manipulator as a target parameter of the deep belief network, adopt a gradient descent method to perform backpropagation on the data vector, and modify the weight vectors between each hidden layer in the network layer by layer until a control torque is generated as the bottom state, thereby completing the network training; A deep learning controller construction unit is used to construct a deep learning controller using the self-organizing neural network and the deep belief network connected in series.

10. A multi-robot collaborative grasping platform based on meta-heuristic algorithm, characterized by: The platform includes a processor, a memory, and a multi-robot collaborative grasping platform control program based on a meta-heuristic algorithm; Wherein, the processor executes the multi-robot collaborative grasping platform control program based on the metaheuristic algorithm, the multi-robot collaborative grasping platform control program based on the metaheuristic algorithm is stored in the memory, and the multi-robot collaborative grasping platform control program based on the metaheuristic algorithm implements the multi-robot collaborative grasping method based on the metaheuristic algorithm as described in any one of claims 1 to 7.

Images