A palletizing control method, device and palletizing robot

Abstract

The application discloses a kind of palletizing control method, equipment and palletizing robot, belong to the technical field of automatic palletizing control, its method includes: obtaining the initial pose of the box to be moved, determines the target pose of the box to be moved based on the preset geometric constraint relationship;Based on the feature extraction layer, the feature vector between initial pose and target pose is extracted;The feature vector is input into the trained alignment agent, and the next relative pose of the box to be moved is obtained;Control the box to be moved from initial pose to next relative pose;Next relative pose is regarded as the initial pose, and the alignment is recycled until next relative pose is the target pose.The application is based on target pose, and the relationship characteristics between current pose and target pose are used for alignment control cycle, so that the final placement position of box satisfies the given geometric constraint relationship, to realize accurate and collision-free box alignment palletizing.

Description

Technical Field

[0001] This invention relates to the field of automated palletizing control technology, and in particular to a palletizing control method, equipment and palletizing robot. Background Technology

[0002] With the booming development of e-commerce and the increasing number of parcels, how to use robots to achieve automated palletizing has become one of the cutting-edge technologies that urgently need to be solved in the Logistics 4.0 era in order to realize automated box-type cargo transportation, handling and warehousing.

[0003] In existing technologies, the placement of boxes is analyzed using neural networks, and a robot is used to place the boxes in designated positions to complete a palletizing operation. For example, in the invention patent CN112047113A, a basic AI-based palletizing system is introduced. This system identifies the workpiece size from a QR code on the workpiece, and based on the workpiece size, analyzes the available placement positions in the current palletizing area using a 3D structured light camera and an artificial neural network. The robot's posture is then adjusted to place the workpiece in the designated position to complete the palletizing operation. However, this system does not consider various disturbances such as robot controller accuracy, workpiece deformation, and neural network analysis errors. This can easily lead to problems such as low space utilization, collisions between workpieces causing collapse, and palletizing failure. Existing technologies all rely on neural networks to determine the absolute coordinates of the box placement to complete the palletizing, which cannot achieve accurate and collision-free box placement during the placement process.

[0004] Therefore, there is an urgent need for a palletizing control method, equipment, and palletizing robot to solve the problem that existing technologies cannot achieve accurate and collision-free placement of boxes. Summary of the Invention

[0005] In view of this, it is necessary to provide a palletizing control method, equipment and palletizing robot that can combine 2D image deep learning with 3D point cloud template matching technology to achieve accurate identification of battery pack defects.

[0006] To address the aforementioned technical problems, this invention provides a palletizing control method, comprising:

[0007] S101. Obtain the initial pose of the box to be moved, and determine the target pose of the box to be moved based on the preset geometric constraint relationship.

[0008] S102. Extract the feature vector between the initial pose and the target pose based on the feature extraction layer;

[0009] S103. Input the feature vector into the trained alignment agent to obtain the next relative pose of the box to be moved;

[0010] S104. Control the moving box to move from the initial pose to the next relative pose;

[0011] S105. Use the next relative pose as the initial pose, and return to S102. Execute S102~S105 until the next relative pose is the target pose.

[0012] In one possible implementation, obtaining the initial pose of the box to be moved includes:

[0013] Identify the dimensions and current position of the container to be moved based on the image data of the container to be moved;

[0014] Identify the structural data of the placed boxes based on the image data of the placed boxes;

[0015] Calculate the height offset of the box to be moved based on the size and current position of the box to be moved, the maximum height in the structural data of the already placed boxes, and the preset gap parameters.

[0016] The initial pose of the box to be moved is determined based on the height offset and the current position of the box to be moved.

[0017] In one possible implementation, the height offset is calculated using the following formula:

[0018]

[0019] in, For height offset, This represents the maximum height of the placed boxes. The current position of the container to be moved. The height of the box. These are the preset gap parameters.

[0020] In one possible implementation, a feature vector between the initial pose and the target pose is extracted based on a feature extraction layer, including:

[0021] Based on preset geometric constraints, a simulation environment for box stacking is constructed according to working errors and environmental disturbances. The geometric constraints include various geometric alignment types.

[0022] Based on different types of geometric constraints, multiple initial pose recognition models are constructed with image data as input and target pose as output. The initial pose recognition model includes a backbone feature extraction network layer.

[0023] Using the target pose as the fitting target, multiple initial pose recognition models are trained in the simulation environment to obtain multiple pose recognition models;

[0024] Acquire real-time image data of the box to be moved and the already placed box at the initial position, and determine the geometric alignment type of the box to be moved;

[0025] Based on the geometric alignment type of the box to be moved, a target pose recognition model with corresponding geometric constraint type is determined. The backbone feature extraction network layer of the target pose recognition model is used to extract the feature vector between the initial pose and the target pose in the real-time image data.

[0026] In one possible implementation, the operating errors include the pose recognition of the box to be moved, the alignment pose recognition of the box to be moved, the hand-eye relationship, the camera observation angle, the depth estimation error, and the placement of the box to be moved; the environmental disturbances include changes in the box size and noise from the robotic arm's movements.

[0027] In one possible implementation, the feature vector is input into a trained agent to obtain the next relative pose of the box to be moved, including:

[0028] Based on preset geometric constraints and action limits, multiple initial alignment agents with different geometric constraint types are constructed, using feature vectors as input and alignment action parameters as output.

[0029] Training datasets with different geometric constraint types are generated based on multiple pose recognition models, and the training datasets include training feature vectors and target pose labels;

[0030] In the simulation environment, multiple initial alignment agents with different geometric constraint types are trained according to the training datasets of different geometric constraint types and the preset value network, resulting in multiple trained alignment agents.

[0031] The target alignment agent with the corresponding geometric constraint type is determined based on the geometric alignment type of the box to be moved, and the next relative pose is obtained based on the target alignment agent and the feature vector.

[0032] In one possible implementation, the alignment agent includes an observation network and a policy network; multiple initial alignment agents of different geometric constraint types are trained based on training datasets of different geometric constraint types and a preset value network to obtain multiple trained alignment agents, including:

[0033] The training dataset is used as the observation data for the observation network;

[0034] Based on the policy network, alignment decision actions are obtained according to the observed data and the preset action limit range;

[0035] Based on a pre-defined value network, the alignment reward is determined according to the box placement posture and geometric constraint relationship when the final alignment decision action is terminated.

[0036] The initial alignment agent is optimized based on the alignment reward to obtain a trained alignment agent.

[0037] In one possible implementation, obtaining the next relative pose based on the target alignment agent and the feature vector includes:

[0038] The feature vector is input into the target alignment agent to obtain alignment action parameters, which include continuous adjustment actions and discrete alignment completion. The continuous adjustment actions include displacement coordinate parameters and Euler angle parameters.

[0039] When the alignment action parameter is a continuous adjustment action, the displacement coordinate parameter and Euler angle parameter are converted into a homogeneous matrix, and the target pose of the box at the next moment is calculated based on the homogeneous matrix and the current pose of the box.

[0040] When the alignment action parameter is "Discrete alignment complete", the alignment control loop ends and the box is moved vertically according to its current pose.

[0041] Secondly, the present invention also provides a palletizing control device, including a memory and a processor, wherein,

[0042] The memory is used to store programs and data;

[0043] The processor, coupled to the memory, is used to execute the program stored in the memory to implement the palletizing control method as described above.

[0044] Thirdly, the present invention also provides a palletizing robot, including a robotic arm, a pneumatic suction cup, a depth camera and an alignment control system;

[0045] The robotic arm is used to move the box to be moved.

[0046] The pneumatic suction cup is used to pick up the box and place it onto the robotic arm;

[0047] The depth camera is used to acquire depth visual observation image data;

[0048] The alignment control system is used to implement the above-mentioned palletizing control method and to perform alignment control on the boxes to be moved.

[0049] The beneficial effects of this invention are as follows: First, the target pose of the box to be moved is determined through preset geometric constraints. Then, cyclic alignment begins, and feature vectors between the current pose and the target pose of the box to be moved are extracted through a feature extraction layer. These feature vectors reflect the characteristics between the current pose of the box to be moved and the geometric constraints. The extracted feature vectors are input into a trained alignment agent to obtain the next relative pose of the box to be moved. After the box to be moved is moved to this relative position, the feature vector extraction and alignment with the agent continue to be cyclically performed until the box moves to the target pose. Based on the target pose, this embodiment analyzes the relative pose relationship between the current pose of the box to be moved and the corresponding target pose in the geometric constraints. Based on these characteristics, the alignment control loop of the box to be moved is cyclically performed, so that the final placement position of the box satisfies the given geometric constraints, thereby achieving accurate and collision-free box alignment and stacking. Attached Figure Description

[0050] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0051] Figure 1 This is a schematic flowchart of an embodiment of the palletizing control method provided by the present invention;

[0052] Figure 2 Schematic diagrams of three geometric constraint relationships provided by the present invention;

[0053] Figure 3 A schematic flowchart of an embodiment of the alignment start pose determination method provided by the present invention;

[0054] Figure 4 Provided by the present invention Figure 1 A schematic flowchart of an embodiment of step S102;

[0055] Figure 5 Provided by the present invention Figure 1 A schematic flowchart of an embodiment of step S103;

[0056] Figure 6 Provided by the present invention Figure 5 A schematic flowchart of an embodiment of step S503;

[0057] Figure 7 This is a schematic diagram of an embodiment of the palletizing control device provided by the present invention. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0059] In the description of the embodiments of the present invention, unless otherwise stated, "multiple" means two or more. "And / or" describes the relationship between related objects, indicating that there can be three relationships. For example, A and / or B can represent three situations: A exists alone, A and B exist simultaneously, and B exists alone.

[0060] The terms "first," "second," etc., used in the embodiments of this invention are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a technical feature defined with "first" or "second" may explicitly or implicitly include at least one of that feature.

[0061] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of the invention. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0062] This invention provides a palletizing control method, equipment, and palletizing robot, which are described below.

[0063] Figure 1 This is a schematic flowchart of an embodiment of the palletizing control method provided by the present invention, as shown below. Figure 1 As shown, the palletizing control method includes:

[0064] S101. Obtain the initial pose of the box to be moved, and determine the target pose of the box to be moved based on the preset geometric constraint relationship.

[0065] S102. Extract the feature vector between the initial pose and the target pose based on the feature extraction layer;

[0066] S103. Input the feature vector into the trained alignment agent to obtain the next relative pose of the box to be moved;

[0067] S104. Control the box to be moved to move from the initial pose to the next relative pose;

[0068] S105. Take the next relative pose as the initial pose and return to S102. Execute S102~S105 until the next relative pose is the target pose.

[0069] It should be noted that this embodiment is based on a robot with a robotic arm and a depth camera. The depth camera on the robotic arm acquires image data of the box to be moved and the image data of the boxes already placed. The initial pose and target pose of the box are identified based on the image data. The method for identifying the initial pose and target pose of the box is not specifically limited here and can be obtained by existing technology. In addition, the target pose of the box to be moved is the position coordinate of the box to be moved when it is about to be placed on the boxes already placed. Here, the multiple boxes already placed are regarded as a whole as a pallet, and the optimal placement position of the box to be moved is determined as the target pose.

[0070] It should be further noted that the placed boxes are considered as the entire pallet, and the preset geometric constraints include three types: planar alignment, one-side alignment, and two-side alignment, as detailed below. Figure 2 As shown, Figure 2 The present invention provides three schematic diagrams of geometric constraint relationships, wherein, Figure 2 In the diagram, 'a' represents the geometric constraint relationship for planar alignment, that is, the two mutually perpendicular sides of the box to be moved and the pallet. The two mutually perpendicular sides of the corresponding pallet The boxes to be moved are respectively aligned relative to each other and aligned in the same direction. The line of intersection between the top and bottom surfaces With pallets The line of intersection between the top surface and the bottom surface By distance Parallelism, this geometric constraint is applicable to scenarios where boxes to be moved are stacked on pallets or on the surface of already stacked boxes; Figure 2 In the diagram, 'b' represents the geometric constraint of side alignment, meaning that the two perpendicular sides of the box to be moved and the box already placed are aligned. The two mutually perpendicular sides of the corresponding placed box. The boxes to be moved are respectively aligned relative to each other and aligned in the same direction. The line of intersection between the face and its bottom surface With the corresponding placed box The line of intersection between the face and its upper base By distance Parallelism, this geometric constraint is applicable to scenarios where one side of the box to be moved coincides with the side of the pallet in the placement pose; Figure 2 In the diagram, 'c' represents the geometric constraint relationship of aligning two sides, that is, for two mutually perpendicular sides of the box to be moved. And the two sides that intersect with the bottom surface. , and the two sides of the corresponding two placed boxes The box whose two sides intersect with the corresponding top bottom surface and have the highest height. middle respectively with coincide, Zhongyu The side with the smallest alternating angle is at a distance of Parallelism, this geometric constraint relationship is applicable to general compact stacking scenarios where planar alignment and edge alignment geometric constraints are not applicable. Based on the above constraints, the optimal placement pose of the box to be moved is determined.

[0071] Specifically, firstly, based on three geometric constraint relationships—planar alignment, one-sided alignment, and two-sided alignment—the geometric constraint type of the box to be moved is determined according to the image data. That is, it determines which alignment type is suitable for achieving ideal alignment between the current structure of the box to be moved and the structure of the already placed box. After determining the type of geometric constraint relationship, the target pose for placing the box to be moved is determined according to the corresponding geometric constraint relationship. Then, the robotic arm transports the box to be moved to the starting pose, and the alignment control loop begins at the starting pose. The alignment control loop is implemented through a trained alignment agent, which includes a feature extraction layer and an alignment control layer. The alignment control layer extracts feature vectors between the initial pose and the target pose through the feature extraction layer. The alignment control layer makes decisions based on the extracted feature vectors. During the alignment control loop, the feature extraction layer extracts feature vectors of the relative position between the current pose and the target pose of the box to be moved through depth image data acquired by a depth camera fixed to the end of the robotic arm. The extracted feature vectors are used as observation data, and decisions are made based on the observation data to control the box to be moved to the next relative position. This process is repeated to gradually control the robotic arm to adjust the relative pose of the box until it moves to the target pose.

[0072] This embodiment determines the target pose of the box to be moved through preset geometric constraints. During the alignment process, a feature extraction layer extracts feature vectors between the current pose and the target pose of the box to be moved. These feature vectors reflect the characteristics between the current pose and the geometric constraints. The extracted feature vectors are input into a trained alignment agent to obtain the next relative pose of the box to be moved. After the box is moved to this relative position, the feature vector extraction and alignment with the agent continue to be looped until the box moves to the target pose. Based on the target pose, this embodiment analyzes the relative pose relationship between the current pose of the box to be moved and the corresponding target pose in the geometric constraints. Based on these characteristics, the alignment control loop of the box to be moved is looped, so that the final placement position of the box satisfies the given geometric constraints, thereby achieving accurate and collision-free box alignment and stacking.

[0073] In some embodiments of the present invention, such as Figure 3 As shown, Figure 3 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S101 includes:

[0074] S301. Identify the dimensions and current position of the container to be moved based on the image data of the container to be moved;

[0075] S302. Identify the structural data of the placed box based on the image data of the placed box;

[0076] It should be noted that the image data of the box to be moved and the image data of the box already placed are obtained by the depth camera on the robotic arm. Based on these image data, the three-dimensional dimensions of the box to be moved, its current position, and the structure of the box already placed are identified. The identification method is not limited here; it can be image recognition technology or other neural network models.

[0077] S303. Calculate the height offset of the box to be moved based on the size and current position of the box to be moved, the maximum height in the structural data of the already placed boxes, and the preset gap parameters.

[0078] The formula for calculating the height offset is:

[0079] ,

[0080] in, For height offset, This represents the maximum height of the placed boxes. The current position of the container to be moved. The height of the box. These are the preset gap parameters.

[0081] It should be noted that when the box is in the ideal alignment pose and moves from top to bottom to the placement pose, it can be placed without collision and while satisfying geometric constraints. Therefore, at the start of the alignment control cycle, the box to be moved needs to be moved a certain height before the alignment control cycle begins. This initial height is related to the maximum height of the already placed box, the size of the box to be moved, and the preset gap parameters.

[0082] It should be further noted that in geometric constraints, the alignment of faces and edges is allowed to have a certain range of error gaps. However, when the gap is too large, it will cause the box to deviate significantly from the placement area. When the gap is too small, the box may collide, causing the alignment to fail. Therefore, by setting appropriate gap parameters through training, a certain range of error is allowed during alignment.

[0083] S304. Determine the initial pose of the box to be moved based on the height offset and the current position of the box to be moved.

[0084] This embodiment calculates the height offset of the box to be moved by using the dimensions and current position of the box to be moved, the maximum height in the structural data of the placed boxes, and the preset gap parameters. Based on the height offset, the alignment starting pose is determined so that the alignment starting point of the box to be moved is higher than the maximum height of the pallet within the allowable range, thus avoiding collisions during the alignment process.

[0085] In some embodiments of the present invention, such as Figure 4 As shown, Figure 4 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S102 includes:

[0086] S401. Based on preset geometric constraints, a simulation environment for box stacking is constructed according to working errors and environmental disturbances. The geometric constraints include various geometric alignment types.

[0087] It should be noted that during the operation of a palletizing robot, operational errors and environmental disturbances are unavoidable issues. These issues directly affect the robot's operational accuracy and efficiency. Therefore, during the training of the neural network and alignment agent, it is necessary to incorporate the operational errors of each module in the robotic arm and environmental disturbances into the simulation environment. This allows the simulation environment to realistically simulate the errors of each module in the system and the environmental disturbances during the actual palletizing process of boxes, simulate the placed box scenarios corresponding to the three geometric constraint relationships, and determine whether the box to be inserted is aligned within a given allowable range.

[0088] S402. Based on different types of geometric constraint relationships, multiple initial pose recognition models are constructed with image data as input and target pose as output. The initial pose recognition model includes a backbone feature extraction network layer.

[0089] S403. Using the target pose as the fitting target, train multiple initial pose recognition models in a simulation environment to obtain multiple pose recognition models.

[0090] S404. Obtain real-time image data of the box to be moved and the box already placed at the initial position, and determine the geometric alignment type of the box to be moved.

[0091] S405. Determine the target pose recognition model with the corresponding geometric constraint type based on the geometric alignment type of the box to be moved, and use the backbone feature extraction network layer of the target pose recognition model to extract the feature vector between the initial pose and the target pose in the real-time image data.

[0092] It should be noted that, based on the three types of geometric constraints, this embodiment constructs three different pose recognition models for different types of geometric constraints in sequence. The acquired depth image data is cropped into multiple rectangular regions of the same size. These rectangular regions contain the observations of most of the geometric features involved in the alignment constraints. Then, these three rectangular regions are stacked along the channel. During the model training process, Gaussian noise and random holes are used for image enhancement processing. Then, median filtering is used to normalize the image according to a given depth threshold before it is used as the input of the model. The structure of the pose recognition model includes two parts: a backbone feature extraction network layer and a fully connected output head.

[0093] Specifically, in the simulation environment, the displacement of the training box relative to the target pose is... ,in, To train the displacement of the box's current pose relative to the target pose. To train the current pose of the box, To train the target pose of the box, this relative displacement is converted into a six-tuple representation of three-dimensional coordinates and Euler angles. As the fitting target of the initial pose recognition model, the original visual observations on the training box in the simulation environment are used as sample attributes. The deviation between the current pose and the target pose is used as a label to construct a regression task for relative pose deviation estimation. A training sample can be obtained by randomly initializing the training simulation environment each time. The optimized model is trained using these cyclic samples from the simulation environment. This makes the optimized model have a backbone feature extraction network that extracts the relative positional relationship features of the box to be placed relative to the ideal alignment position in the visual observation. At the same time, different pose recognition models are constructed for different types of geometric constraints. The pose recognition models for these different geometric constraints are cyclically applied to obtain a target pose recognition model that can adapt to different geometric constraints.

[0094] Furthermore, the backbone feature extraction network layer of the trained pose recognition model can accurately extract the feature vectors between the box to be moved and the target pose. The fully connected output head maps the extracted feature vectors to a six-tuple representation of the three-dimensional coordinates and Euler angles of the current pose relative to the target pose. During the alignment control loop, only the backbone network in the target pose recognition model is used to extract the feature vectors of the relative positional relationship between the box to be moved and the target pose. These feature vectors can accurately represent the relative positional relationship between the box to be moved and the target pose, providing accurate data for the next step of the alignment control loop.

[0095] In some embodiments of the present invention, the working errors include the pose recognition of the box to be moved, the alignment pose recognition of the box to be moved, the hand-eye relationship, the camera observation angle, the depth estimation error, and the placement method of the box to be moved; the environmental disturbances include changes in the size of the box and the noise of the robotic arm's movements.

[0096] Specifically, the changes in the size of the boxes to be moved include workpiece size deviations caused by deformation or manufacturing tolerances of the boxes to be moved in general palletizing tasks, and different box sizes in mixed-box palletizing tasks.

[0097] The workpiece pose recognition is derived from the error in the algorithm for recognizing the pose of the moving box and estimating the box size in a typical palletizing system. This error can be expressed as the relationship between the actual pose of the box and the recognized pose, and its value range is determined based on the accuracy of the box recognition module.

[0098] The alignment pose recognition of the box to be moved originates from the error in the palletizing planner's planning of the box placement pose in a typical palletizing system. This error can be expressed as the relationship between the ideal placement pose and the planned placement pose of the box, and its value range is determined based on the planning accuracy of the palletizing planning module.

[0099] The hand-eye relationship arises from camera hand-eye calibration errors, disturbances caused by external mechanical factors such as vibration during use, and errors in the box size estimation algorithm. This error can be expressed as the relationship between the ideal camera pose and the actual camera pose, and its value range is determined based on the hand-eye relationship calibration and the box size recognition accuracy.

[0100] The camera observation angle is a small error between the intrinsic parameters and the nominal value caused by manufacturing defects between cameras of the same model. It is represented by the error of the observation angle relative to the nominal value, and its range is determined based on the performance parameters of the depth camera used.

[0101] Depth estimation error mainly manifests as holes and depth estimation errors caused by reflection, occlusion, etc. during the depth estimation process of the depth camera. The value range is determined according to the performance parameters of the depth camera used.

[0102] Robotic arm motion noise originates from errors caused by the robot arm's parameter identification and planning algorithms, as well as control, during path planning and decision-making by the intelligent agent. This manifests as an error between the actual pose reached by the robot arm and the target pose. This error can represent the relationship between the pose of the box attached to the end effector of the robot arm in the ideal motion and the actual pose after the motion. The value range is determined based on the robot arm's motion accuracy.

[0103] When moving boxes with the same placement method, in a real palletizing environment, there are multiple possibilities for the placement of boxes that are subject to the three geometric constraints.

[0104] In this embodiment, the box palletizing errors caused by the above-mentioned multiple working errors and environmental disturbances are taken into account in the simulation environment, so that the feature extraction layer and alignment agent obtained by simulation training can eliminate the error influence of these working errors and environmental disturbances on box palletizing, making the palletizing control more accurate.

[0105] In some embodiments of the present invention, such as Figure 5 As shown, Figure 5 Provided by the present invention Figure 1 A flowchart illustrating an embodiment of step S103 includes:

[0106] S501. Based on preset geometric constraints and action limits, using feature vectors as input and alignment action parameters as output, construct multiple initial alignment agents with different geometric constraint types.

[0107] S502. Generate training datasets with different geometric constraint types based on multiple pose recognition models. The training datasets include training feature vectors and target pose labels.

[0108] S503. In the simulation environment, multiple initial alignment agents with different geometric constraint types are trained according to training datasets of different geometric constraint types and preset value networks to obtain multiple trained alignment agents.

[0109] S504. Determine the target alignment agent with the corresponding geometric constraint type based on the geometric alignment type of the box to be moved, and obtain the next relative pose based on the target alignment agent and the feature vector.

[0110] It should be noted that the implementation method for obtaining continuous alignment action parameters from feature vectors is not specifically limited. In this embodiment, it is implemented through an intelligent agent, which includes a state space and an action space. The backbone feature extraction network layer of the target pose recognition model is added to the intelligent agent as a feature extractor, extracting feature vectors containing the target pose satisfying geometric constraints and the current pose of the box to be moved, as the state space set. In the action space, a low-level parameterized action space is used, that is, the intelligent agent can only choose one of two mutually exclusive actions (adjusting pose and completing alignment) to execute at each time t. Adjusting pose also includes alignment action parameters, which are six-tuple motion parameters containing three-dimensional coordinate displacement and Euler angle rotation parameters to define the specific adjustment behavior. Completing alignment has no alignment action parameters. Therefore, the action space set is... .

[0111] It should be further noted that during the alignment cycle, in order to prevent the robotic arm from making excessive movements that could lead to collisions between the boxes, the robotic arm's range of motion must be within a limited range. Therefore, a preset action limit range is set to restrict the alignment action parameters, ensuring that the parameters used by the robotic arm to adjust the pose of the box to be moved do not exceed this limit range. In this embodiment, the limit range for coordinate displacement is [-5mm, 5mm], and the limit range for Euler angle adjustment is... .

[0112] In some embodiments of the present invention, the alignment agent includes an observation network and a policy network; such as Figure 6 As shown, Figure 6 Provided by the present invention Figure 5 A flowchart illustrating an embodiment of step S503 includes:

[0113] S601. Use the training dataset as the observation data for the observation network;

[0114] S602. Based on the policy network, the alignment decision action is obtained according to the observation data and the preset action limit range;

[0115] S603. Based on the preset value network, determine the alignment reward according to the box placement posture and geometric constraint relationship when the final alignment decision action is terminated.

[0116] S604. Optimize the initial alignment agent based on the alignment reward to obtain a trained alignment agent.

[0117] Specifically, during the training process of the alignment agent, if the alignment agent outputs that alignment is complete, an alignment check is performed to determine whether the training box meets the preset geometric constraints, i.e., whether alignment is satisfied and whether the gap is within the preset gap range, and whether there is a collision. The agent is rewarded and penalized in turn. If the alignment agent outputs the adjustment pose and alignment action parameters, the current alignment step size is checked to see if it exceeds the preset maximum step size to prevent continuous interaction from increasing training time. The reward is calculated and then a perturbation based on the sampling of the robot arm control error is applied to the pose of the robot arm. The purpose of calculating the reward before applying the perturbation is to ensure that the reward function can correctly judge the adjustment effect of the agent without being affected by environmental perturbations, so as to facilitate the fitting of the value network.

[0118] In some embodiments of the present invention, obtaining the next relative pose based on the target alignment agent and the feature vector includes:

[0119] The feature vector is input into the target alignment agent to obtain the alignment action parameters, which include continuous adjustment actions and discrete alignment completion. The continuous adjustment actions include displacement coordinate parameters and Euler angle parameters.

[0120] It should be noted that the alignment decision action includes continuous alignment action parameters and discrete alignment completion. Continuous alignment action parameters include three-dimensional coordinate displacement parameters and Euler angle rotation parameters. Discrete alignment completion indicates the end of the alignment control loop.

[0121] When the alignment action parameters are continuously adjusted, the displacement coordinate parameters and Euler angle parameters are converted into homogeneous matrices, and the target pose of the box at the next moment is calculated based on the homogeneous matrix and the current pose of the box.

[0122] Specifically, since palletizing problems typically require robotic arms to move within a small spatial range in different postures, the joint space of the robotic arm is not used as the motion parameter. Instead, the six-tuple parameters represented by coordinates and Euler angles are converted into a homogeneous matrix of the displacement of the box to be moved. The conversion formula for the homogeneous matrix is ​​as follows:

[0123] ,

[0124] ,

[0125] ,

[0126] Where c is the abbreviation for cos and s is the abbreviation for sin. For three-dimensional coordinate displacement parameters, These are the Euler angle rotation parameters. Let t be the pose of the box to be moved at the current time t. To determine the ideal placement position for the container to be moved. Let be the homogeneous matrix of the displacement of the box to be moved.

[0127] Furthermore, the robotic arm motion planning algorithm uses the homogeneous matrix as the planning target at time t, transforms the box displacement into a usable motion planning target, and finally determines the target pose that the robotic arm needs to move to based on the sampling position of the suction cup on the box.

[0128] ,

[0129] ,

[0130] in, This represents the current pose of the container to be moved. This is the position where the moving container is about to arrive. Let be the homogeneous matrix of the displacement of the box to be moved. The target pose for the robotic arm to move in. Invert the matrix for the suction cup coordinate system.

[0131] When the alignment action parameter is "Discrete alignment complete", the alignment control loop ends and the box is moved vertically according to its current pose.

[0132] It should be noted that if the alignment action parameters are not discrete alignment completion, but the alignment exceeds the set maximum alignment time, the alignment control loop ends and the box is placed according to the current pose.

[0133] To better implement the palletizing control method in the embodiments of the present invention, based on the palletizing control method, correspondingly, as follows: Figure 7 As shown, the present invention also provides a palletizing control device 700. The palletizing control device 700 includes a processor 701, a memory 702, and a display 703. Figure 7 Only some components of the palletizing control device 700 are shown, but it should be understood that it is not required to implement all of the components shown, and more or fewer components may be implemented instead.

[0134] In some embodiments, processor 701 may be a central processing unit (CPU), a microprocessor, or other data processing chip, used to run program code stored in memory 702 or process data, such as the palletizing control method of the present invention.

[0135] In some embodiments of the present invention, processor 701 may be a single server or a group of servers. The server group may be centralized or distributed. In some embodiments, processor 701 may be local or remote. In some embodiments, processor 701 may be implemented on a cloud platform. In one embodiment, the cloud platform may include a private cloud, public cloud, hybrid cloud, community cloud, distributed cloud, internal cloud, multi-cloud, etc., or any combination thereof.

[0136] In some embodiments, memory 702 may be an internal storage unit of the palletizing control device 700, such as a hard disk or memory of the palletizing control device 700. In other embodiments, memory 702 may also be an external storage device of the palletizing control device 700, such as a pluggable hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the palletizing control device 700.

[0137] Furthermore, the memory 702 may include both internal storage units of the palletizing control device 700 and external storage devices. The memory 702 is used to store application software and various types of data for which the palletizing control device 700 is installed.

[0138] In some embodiments, display 703 may be an LED display, a liquid crystal display, a touch-sensitive liquid crystal display, or an OLED (Organic Light-Emitting Diode) touchscreen. Display 703 is used to display information from the palletizing control device 700 and to display a visual user interface. Components 701-703 of the palletizing control device 700 communicate with each other via a system bus.

[0139] In some embodiments of the present invention, when the processor 701 executes the palletizing control program in the memory 702, the following steps can be implemented:

[0140] S101. Obtain the initial pose of the box to be moved, and determine the target pose of the box to be moved based on the preset geometric constraint relationship.

[0141] S102. Extract the feature vector between the initial pose and the target pose based on the feature extraction layer;

[0142] S103. Input the feature vector into the trained alignment agent to obtain the next relative pose of the box to be moved;

[0143] S104. Control the box to be moved to move from the initial pose to the next relative pose;

[0144] S105. Take the next relative pose as the initial pose and return to S102. Execute S102~S105 until the next relative pose is the target pose.

[0145] It should be understood that when the processor 701 executes the palletizing control program in the memory 702, in addition to the functions mentioned above, it can also perform other functions, as can be found in the description of the corresponding method embodiments above.

[0146] Furthermore, this embodiment of the invention does not specifically limit the type of palletizing control device 700 mentioned. The palletizing control device 700 can be a portable palletizing control device such as a mobile phone, tablet computer, personal digital assistant (PDA), wearable device, or laptop computer. Exemplary embodiments of portable palletizing control devices include, but are not limited to, portable palletizing control devices running iOS, Android, Microsoft, or other operating systems. The aforementioned portable palletizing control device can also be other portable palletizing control devices. It should also be understood that in some other embodiments of the invention, the palletizing control device 700 may not be a portable palletizing control device, but rather a desktop computer with a touch-sensitive surface (e.g., a touch panel).

[0147] Accordingly, embodiments of the present invention also provide a palletizing robot, including a robotic arm, a pneumatic suction cup, a depth camera, and an alignment control system;

[0148] The robotic arm is used to move the container.

[0149] Pneumatic suction cups are used to pick up the box and place it onto the robotic arm;

[0150] Depth cameras are used to acquire image data from depth vision observations;

[0151] The alignment control system is used to implement the steps or functions in the palletizing control methods provided in the above-described method embodiments.

[0152] The palletizing control method, equipment, and palletizing robot provided by the present invention have been described in detail above. Specific examples have been used to illustrate the principles and implementation methods of the present invention. The description of the above embodiments is only for the purpose of helping to understand the method and core ideas of the present invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas of the present invention. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A palletizing control method characterized by, include: S101. Obtain the initial pose of the box to be moved, and determine the target pose of the box to be moved based on preset geometric constraints; wherein, obtaining the initial pose of the box to be moved includes: identifying the size and current position of the box to be moved based on the image data of the box to be moved; identifying the structural data of the placed box based on the image data of the placed box; calculating the height offset of the box to be moved based on the size and current position of the box to be moved, the maximum height in the structural data of the placed box, and a preset gap parameter; determining the initial pose of the box to be moved based on the height offset and the current position of the box to be moved; S102. Extracting feature vectors between the initial pose and the target pose based on the feature extraction layer includes: constructing a simulation environment for box palletizing based on preset geometric constraints and according to working errors and environmental disturbances, wherein the geometric constraints include multiple geometric alignment types; constructing multiple initial pose recognition models based on different types of geometric constraints, using image data as input and the target pose as output, wherein the initial pose recognition model includes a backbone feature extraction network layer; training the multiple initial pose recognition models in the simulation environment with the target pose as the fitting target to obtain multiple pose recognition models; acquiring real-time image data of the box to be moved and the already placed boxes at the initial position, and determining the geometric alignment type of the box to be moved; determining the target pose recognition model with the corresponding geometric constraint type based on the geometric alignment type of the box to be moved, and using the backbone feature extraction network layer of the target pose recognition model to extract the feature vectors between the initial pose and the target pose in the real-time image data; S103. Input the feature vector into the trained alignment agent to obtain the next relative pose of the box to be moved; S104. Control the moving box to move from the initial pose to the next relative pose; S105. Use the next relative pose as the initial pose, and return to S102. Execute S102~S105 until the next relative pose is the target pose.

2. The palletizing control method according to claim 1, characterized by, The formula for calculating the height offset is: in, For height offset, This represents the maximum height of the placed boxes. The current position of the container to be moved. The height of the box. These are the preset gap parameters.

3. The palletizing control method according to claim 1, characterized by, The operating errors include the recognition of the pose of the box to be moved, the recognition of the alignment pose of the box to be moved, the hand-eye relationship, the camera observation angle, the depth estimation error, and the placement method of the box to be moved; the environmental disturbances include changes in the size of the box and the noise of the robotic arm's movements.

4. The palletizing control method according to claim 1, characterized by, The feature vector is input into the trained agent to obtain the next relative pose of the box to be moved, including: Based on preset geometric constraints and action limits, multiple initial alignment agents with different geometric constraint types are constructed, using feature vectors as input and alignment action parameters as output. Training datasets with different geometric constraint types are generated based on multiple pose recognition models, and the training datasets include training feature vectors and target pose labels; In the simulation environment, multiple initial alignment agents with different geometric constraint types are trained according to the training datasets of different geometric constraint types and the preset value network, resulting in multiple trained alignment agents. The target alignment agent with the corresponding geometric constraint type is determined based on the geometric alignment type of the box to be moved, and the next relative pose is obtained based on the target alignment agent and the feature vector.

5. The palletizing control method according to claim 4, characterized in that, The alignment agent includes an observation network and a policy network; multiple initial alignment agents with different geometric constraint types are trained based on training datasets of different geometric constraint types and a preset value network to obtain multiple trained alignment agents, including: The training dataset is used as the observation data for the observation network; Based on the policy network, alignment decision actions are obtained according to the observed data and the preset action limit range; Based on a pre-defined value network, the alignment reward is determined according to the box placement posture and geometric constraint relationship when the final alignment decision action is terminated. The initial alignment agent is optimized based on the alignment reward to obtain a trained alignment agent.

6. The palletizing control method according to claim 4, characterized in that, The next relative pose is obtained based on the target alignment agent and the feature vector, including: The feature vector is input into the target alignment agent to obtain alignment action parameters, which include continuous adjustment actions and discrete alignment completion. The continuous adjustment actions include displacement coordinate parameters and Euler angle parameters. When the alignment action parameter is a continuous adjustment action, the displacement coordinate parameter and Euler angle parameter are converted into a homogeneous matrix, and the target pose of the box at the next moment is calculated based on the homogeneous matrix and the current pose of the box. When the alignment action parameter is "Discrete alignment complete", the alignment control loop ends and the box is moved vertically according to its current pose.

7. A palletizing control apparatus characterized by comprising: Including memory and processor, among which, The memory is used to store programs; The processor, coupled to the memory, is used to execute the program stored in the memory to implement the steps in the palletizing control method of any one of claims 1 to 6.

8. A palletizing robot, characterized in that Includes robotic arms, pneumatic suction cups, depth cameras, and alignment control systems; The robotic arm is used to move the box to be moved. The pneumatic suction cup is used to pick up the box and place it onto the robotic arm; The depth camera is used to acquire depth visual observation image data; The alignment control system is used to implement the palletizing control method according to any one of claims 1 to 6, and to perform alignment control on the boxes to be moved.

Images