Method for operating a padding installation, padding installation, method for training a function for operating the padding installation
Patent Information
- Application Number
- PCT/EP2026/056262
- Authority / Receiving Office
- WO · WO
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2025-03-10
- Filing Date
- 2026-03-06
- Publication Date
- 2026-09-17
Smart Images
Figure EP2026056262_17092026_PF_FP_ABST
Abstract
Description
[0001] DREISS PATENT ATTORNEYS 06.03.2026
[0002] 35020917WO
[0003] Gardner AG
[0004] Gartenstrasse 20
[0005] 72663 Großbettlingen
[0006] Method for operating a padding system, padding system, method for training a function to operate the padding system
[0007] The invention relates to advances in the field of the arrangement of cushioning materials in shipping containers.
[0008] In distribution centers for e-commerce applications, small to medium-sized cartons, acting as shipping containers, are filled with several different products, which usually already possess a certain robustness due to their primary packaging. The packer then secures the cartons by inserting suitable cushioning materials.
[0009] Protective packaging, such as paper pads or bubble wrap bags, is secured to the extent that it can withstand further transport to the end customer undamaged.
[0010] The problems of the prior art are solved by a method for operating a cushioning system according to claim 1, by a cushioning system according to a dependent claim, and by a method for training a function according to a dependent claim.
[0011] One aspect of the description concerns the following subject matter: A method for operating a cushioning system, the method comprising: providing information about a plurality of available cushioning options, wherein the available cushioning options differ from one another in at least one material, size, or structural property; acquiring geometric data of the interior of a shipping container by means of a sensor device; generating a plurality of planned cushioning actions based on the determined data representation of the interior of the shipping container and based on the geometric data of the interior of the shipping container, wherein each of the planned cushioning actions comprises at least one of the available cushioning options and a pose of the at least one cushioning option in relation to the interior of the shipping container;and providing the majority of planned cushioning actions to introduce multiple cushioning materials into the interior of the shipping container via a packing opening of the shipping container, according to the assigned cushioning option, and to position them in the interior of the shipping container according to the assigned pose.
[0012] Determining the planned cushioning actions creates a basis for the fully or semi-automated placement of cushioning material in the shipping container. Advantageously, generating these planned cushioning actions allows for machine-guided specifications regarding cushioning material usage and placement, enabling the reliable implementation of optimized cushioning strategies for each individual shipping container.
[0013] In particular, the use of padding material can be reduced across multiple padding stations, which has a positive effect on the environmental footprint of shipping.
[0014] Furthermore, the quality of upholstery can also be improved by reducing the use of upholstery materials, through the implementation of advantageous stable arrangements of upholstery materials across multiple upholstery stations.
[0015] Furthermore, the risk of faulty and insufficient shipping cushioning is significantly reduced, as the proposed procedure ensures compliance with cushioning regulations.
[0016] Furthermore, the availability of planned cushioning actions makes the cushioning time more controllable. Automatic or semi-automatic cushioning allows for reduced cycle times, thereby increasing the throughput of optimally cushioned shipping containers.
[0017] In an advantageous example, the method involves determining a data representation of the shipping container's interior, comprising multiple spatial elements, based on the acquired geometric data. This representation then generates multiple planned cushioning actions based on this data representation and the provided information on cushioning material options. The resulting data representation advantageously simplifies the determination of planned cushioning actions. The division into spatial elements enables machine-guided generation of suggestions for appropriately placing the available cushioning material options within the shipping container.
[0018] An advantageous example is characterized by the fact that the majority of planned cushioning actions are provided in a plurality of layers, with each layer running perpendicular to an imaginary packing axis of the shipping container and each comprising a subset of the plurality of spatial elements.
[0019] Layer-based provisioning simplifies the arrangement of the cushions, as it allows for a layer-based arrangement of the cushions in the shipping container.
[0020] At least one of the layers can also be empty.
[0021] An advantageous example is characterized by the fact that the generation of the plurality of planned cushioning actions is carried out separately for each of a plurality of layers, with the layers each running perpendicular to an imaginary packing axis of the shipping container and each comprising a subset of the plurality of spatial elements.
[0022] Layer-based generation reduces the complexity of the calculations, making it easy and safe to achieve the desired cushioning quality.
[0023] An advantageous example is characterized by the fact that the layer-by-layer generation of the majority of planned cushioning actions is additionally based on a preferred direction for the orientation of the cushioning options inside the shipping container, with the preferred directions of adjacent layers differing from each other, in particular by 90°.
[0024] Changing the preferred direction, particularly regarding the length of the respective padding elements, prevents the creation of unstable padding. The preferred direction is, for example, the direction or axis along which the longitudinal axis of the respective padding element is aligned.
[0025] An advantageous example is characterized by the fact that generating the majority of planned padding actions includes: determining a start layer within the data representation as the starting point for layer-wise generation; determining an end layer within the data representation as the endpoint for layer-wise generation; and performing the layer-wise generation of zero or more planned padding actions in the start layer, in the end layer, and in layers between the start layer and the end layer.
[0026] Limiting the layered padding offers the advantage of creating the possibility of using specific packing patterns for the cushioning materials. In particular, this also allows for a targeted reduction in the amount of cushioning material required.
[0027] An advantageous example is characterized by the fact that the determined planned cushioning actions for the majority of layers are provided after performing the layer-by-layer generation in order to position the cushioning materials assigned to the respective cushioning option according to the respective assigned pose in the shipping container.
[0028] The advantage is that the actual padding can then be implemented in one step.
[0029] An advantageous example is characterized by the following additional features: capturing further geometric data of the interior of the shipping container using the sensor device; determining a further data representation of the interior of the shipping container, encompassing a plurality of spatial elements, depending on the captured further geometric data; determining a cushioning quality based on the further data representation of the interior of the shipping container; closing the shipping container if the cushioning quality is assessed as sufficient, or directing the shipping container to manual inspection if the cushioning quality is assessed as insufficient.
[0030] This approach offers the advantage of achieving a high degree of automation while simultaneously providing a manual review option for problematic cases.
[0031] An advantageous example is characterized by the fact that, after determining zero or multiple cushioning actions for one or more layers, the method comprises: acquiring further geometric data of the interior of the shipping container using the sensor device; determining a further data representation of the interior of the shipping container comprising a plurality of spatial elements, depending on the acquired further geometric data, according to which at least some of the assigned spatial elements are occupied and at least some other spatial elements are not occupied; and continuing to generate at least one further planned cushioning action depending on the further data representation and depending on the available cushioning material options.It is therefore advantageous to carry out an interim assessment of the upholstery in order to determine further planned upholstery actions based on this assessment.
[0032] An advantageous example is characterized by the fact that the start layer of the data representation of the interior of the shipping container is located between a layer representing the bottom of the shipping container and the end layer of the data representation of the interior of the shipping container.
[0033] Advantageously, the process fills the shipping container step by step with virtual cushioning material, in the same order in which the cushioning system fills the shipping container based on the provided cushioning actions.
[0034] An advantageous example is characterized by the fact that the procedure includes:
[0035] Specifying a bottom-fill packing pattern based on a shipping container identifier and / or a predefined configuration of the cushioning system; wherein the method according to the bottom-fill packing pattern is carried out over several layers, starting with a start layer associated with the bottom-fill packing pattern and ending with a finish layer associated with the bottom-fill packing pattern, and includes: determining the start layer associated with the bottom-fill packing pattern within the space elements, the start layer associated with the bottom-fill packing pattern comprising the space elements adjacent to the bottom of the shipping container; determining the finish layer associated with the bottom-fill packing pattern within the space elements, the finish layer associated with the bottom-fill packing pattern being determined based on a predefined number of layers and / or based on the shipping container identifier;and performing the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the bottom-fill pack pattern, in the end layer associated with the bottom-fill pack pattern, and in the layers between the start layer and the end layer associated with the bottom-fill pack pattern.
[0036] An advantageous example is characterized by the fact that the procedure includes:
[0037] Determining a top-fill packing pattern based on an identifier of the shipping container and / or based on a predefined configuration of the cushioning system; wherein the method, according to the defined top-fill packing pattern, is carried out over several layers, starting with a start layer associated with the top-fill packing pattern and ending with a finish layer associated with the top-fill packing pattern, and comprises: determining the start layer associated with the top-fill packing pattern within the space elements, wherein within the start layer the assigned space elements of the interior of the shipping container are marked as empty and the start layer adjoins a layer with space elements, at least one of which is marked as occupied;Defining the final layer associated with the top-fill packing pattern within the spatial elements depending on a predefined parameter and / or depending on an identifier of the shipping container and / or depending on an imaginary plane in which the opening of the shipping container runs, wherein the final layer associated with the top-fill packing pattern is located in the area of the opening of the shipping container or adjacent to this area; and carrying out the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the top-fill packing pattern, in the end layer associated with the top-fill packing pattern and in the layers associated with the top-fill packing pattern located between the start layer and the end layer.
[0038] An advantageous example is characterized by the fact that the procedure includes:
[0039] Specifying a block-and-brace packing pattern depending on an identifier of the shipping container and / or depending on a predefined configuration of the cushioning system; wherein the method according to the block-and-brace packing pattern is carried out over several layers, starting with a start layer associated with the block-and-brace packing pattern and ending with a finish layer associated with the block-and-brace packing pattern, and comprises: determining the start layer associated with the block-and-brace packing pattern within the space elements, the start layer associated with the block-and-brace packing pattern comprising at least one space element of the interior of the shipping container marked as empty, wherein the start layer adjoins a layer with space elements marked as occupied or a floor of the shipping container;Determining the end layer associated with the block-and-brace packing pattern within the spatial elements, the end layer associated with the block-and-brace packing pattern comprising at least one empty marked spatial element of the interior of the shipping container and at least one occupied marked spatial element of the interior of the shipping container, wherein the end layer associated with the block-and-brace packing pattern adjoins a layer within the spatial elements that either comprises only space elements marked as free or comprises space elements that indicate a termination of an object arranged in the shipping container facing away from the bottom of the shipping container;and performing the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the block-and-brace packing pattern, in the end layer associated with the block-and-brace packing pattern, and in the layers between the start layer and the end layer associated with the block-and-brace packing pattern.
[0040] An advantageous example is characterized by the fact that generating the at least one planned cushioning action comprises: determining at least a part of the plurality of cushioning actions in an output area of a function, providing the representation of the plurality of spatial elements of the interior of the shipping container or a part thereof (in particular the spatial elements) of a respective layer in an input area of the function, and providing the information on the plurality of available cushioning options in the input area of the function.
[0041] An advantageous example is characterized by the fact that in the output area of the function, those planned cushioning actions are provided which are assigned to one of the multiple layers of the spatial elements, while in the input area of the function, the information on the available cushioning options and the representation of the interior of the shipping container are provided which is assigned to one of the multiple layers.
[0042] An advantageous example is characterized by the fact that a spatial element includes at least the following parameters: two-dimensional data and a layer affiliation or three-dimensional coordinates; and a status occupied or free.
[0043] An advantageous example is characterized by the fact that the provisioning includes the output of the at least one planned cushioning action to the automated cushioning system: providing an available cushioning material associated with the cushioning option of the at least one planned cushioning action by means of a cushioning material provisioning device of the cushioning system; placing the provided cushioning material in the interior of the shipping container based on the determined position by means of a handling device of the cushioning system.
[0044] An advantageous example is characterized in that the provision of the cushioning material includes a manipulation step prior to placing the cushioning material inside the shipping container, wherein the manipulation step prior to placing the cushioning material includes reshaping the cushioning material by means of the handling device and / or a separate reshaping device.
[0045] An advantageous example is characterized by the fact that a manipulation step after placing the cushioning material inside the shipping container includes reshaping the placed cushioning material using the handling device and / or a separate reshaping device.
[0046] An advantageous example is characterized in that at least the placement of the at least one provided cushioning material is carried out layer by layer, in that an arrangement of cushioning materials in a subsequent layer is only started when an arrangement of cushioning materials in a preceding layer of the subsequent layer has been completed.
[0047] This allows cushioning materials to be stacked layer by layer, preventing unintentional gaps that are later inaccessible to the handling equipment, as these gaps are closed by a layer of cushioning.
[0048] A second aspect of the description concerns the following item: A non-volatile, computer-readable medium that stores instructions which, when executed by one or more processors, cause the one or more processors to execute the procedure according to the previous aspect.
[0049] A third aspect of the description concerns the following object: A cushioning system comprising at least one cushioning provisioning device for providing information on a plurality of available cushioning options, wherein the available cushioning options differ from one another in at least one material, size, or structural property; a sensor device for capturing geometric data of the interior of a shipping container; a determination unit for determining a data representation of the interior of the shipping container comprising a plurality of spatial elements, depending on the captured geometric data;the investigation unit for generating a plurality of planned cushioning actions based on the determined data representation of the interior of the shipping container and based on the provided information on available cushioning options, wherein the at least one planned cushioning action includes at least one of the available cushioning options and a pose of the at least one cushioning option in relation to the spatial elements of the interior of the shipping container;and a cushioning material delivery device for providing the plurality of planned cushioning actions to position cushioning material in the interior of the shipping container according to the assigned cushioning material option and the assigned pose. A fourth aspect of the description concerns the following: A method for training a function comprising providing a simulation environment for simulating different states of an interior of a shipping container divided into spatial elements, wherein, according to an initial state of the interior, at least some of the assigned spatial elements are unoccupied; providing information about a plurality of available cushioning material options, wherein the available cushioning material options differ from each other in at least one material, size, or structural property;Generating a plurality of cushioning actions by an agent based on at least one state of the simulated interior of the shipping container, in particular based on the initial state, and based on the provided information on cushioning options, wherein each planned cushioning action comprises a cushioning option and a pose in relation to the spatial elements of the interior of the shipping container; applying the plurality of cushioning actions to the simulated interior of the shipping container to generate a subsequent state of the simulated interior of the shipping container, wherein each plurality of cushioning actions comprises a respective arrangement of a simulated cushioning material in at least one or more spatial elements marked as free, according to the associated pose;Determining a cushioning quality as a function of the subsequent state of the simulated interior of the shipping container caused by at least one cushioning action; and adjusting at least one parameter of the function to be trained, which controls the agent, using the determined cushioning quality and the associated planned cushioning actions, wherein the procedure at least comprehensively repeats the steps of generating, applying, determining, and adjusting in order to improve the agent's decision-making behavior.
[0050] An advantageous example is characterized in that generating the plurality of planned cushioning actions includes: identifying a subset of the plurality of spatial elements that are part of a layer, the layer optionally following a plane parallel to an access opening of the shipping container leading into the interior of the shipping container, or following a plane parallel to a floor of the interior of the shipping container, or following a plane parallel to a reference plane of the shipping container or cushioning system otherwise defined; and generating at least a part of the plurality of planned cushioning actions, in particular per layer, based on the identified subset of the plurality of spatial elements and based on the provided information on cushioning options.An advantageous example is characterized by the fact that the cushioning quality is determined in terms of a greater reward for the identified cushioning actions, the lower the number of overlapping edges or distal ends of cushioning materials in adjacent layers running parallel to the bottom of the shipping container.
[0051] Advantageously, the function is trained in such a way that the cushioning actions issued by the trained function result in a more stable cushioning material structure being introduced into the shipping container.
[0052] An advantageous example is characterized by the fact that the function to be trained is implemented as a policy function in the form of a deep neural network, which maps a given state of the simulated interior of the shipping container directly to a probability distribution over cushioning options and poses, and that a policy gradient method is used to optimize this policy function.
[0053] This enables the agent to learn directly which actions in which states deliver good results regarding cushioning quality, without having to explicitly store all state-action pairs as Q values, for example.
[0054] An advantageous example is characterized by the fact that the deep neural network representing the agent's function includes at least one convolutional level to process the spatial elements of the interior space, and that fully connected levels are used in further hidden layers of the deep neural network to derive a suitable padding action from the extracted features.
[0055] Thanks to the CNN structure, spatially relevant patterns in the arrangement of already placed upholstery elements and available free space can be better recognized and utilized, thus improving the quality of the learned upholstery planning. Additionally, an efficiency assessment is possible, as fewer steps are required to achieve high spatial utilization.
[0056] An advantageous example is characterized by the fact that, in addition to spatial utilization and / or the protection of the goods in the shipping container, further parameters are considered when determining the cushioning quality, in particular a cost factor of the cushioning material and / or a stability assessment. Another advantageous example is characterized by the fact that, during the simulation of the interior of the shipping container, a physical collision detection is performed, which checks whether the planned cushioning material can be positioned correctly when applying the respective cushioning action, and in the event of a collision or overlap, the cushioning quality is reduced or a penalty is generated as a negative reward signal for the agent.
[0057] Physical collision detection ensures that the planned arrangement of the padding material in the simulator is also practically feasible. For this purpose, when selecting and applying a padding action in the simulated interior, the system checks whether the padding material overlaps with existing objects or other padding materials. If an overlap occurs, the padding quality is reduced accordingly, thus issuing a negative reward signal intended to discourage the agent from repeating physically implausible actions. This ensures that the learned decision-making behavior is geared towards realistic actions.
[0058] In one example, during training, only permitted actions of the agent are actually evaluated based on collision detection. Illegal actions are not evaluated or are discarded.
[0059] An advantageous example is characterized by the fact that the cushioning quality is additionally determined depending on a regulatory parameter, which, for example, limits the amount of cushioning material used or ensures compliance with specified environmental or shipping guidelines by including a corresponding penalty component or an additional reward component factor as part of the cushioning quality.
[0060] Such a regulatory parameter could include, for example: a) Limiting the amount of cushioning material used: If too much material is used per action, a penalty component increases or the reward decreases; b) Reflecting environmental or transport regulations: Certain cushioning materials are only permitted under certain restrictions, or packaging must meet specific safety standards. Here, too, penalty terms for rule violations can be integrated; c) Considering time or capacity limitations: If laying or placing cushioning materials becomes problematic when a time limit is exceeded, a corresponding negative term can reduce the reward; d) It is also possible to include unstable structures of cushioning materials in the evaluation as a regulatory parameter, in relation to the previous layers or voids.To fine-tune the upholstery strategy being learned, it is advisable, as this example shows, to incorporate further criteria into the cost or reward function. This allows the agent to receive detailed feedback on the quality of their chosen strategy, going beyond the mere criterion of upholstery quality. The agent thus learns to find not only the most effective arrangement of upholstery material, but also one that is resource-efficient and compliant with regulations.
[0061] The drawing shows:
[0062] Fig. 1 shows a schematic block diagram and a schematically depicted cushioning system for inserting cushioning materials into a shipping container;
[0063] Fig. 2 shows a schematic flowchart;
[0064] Fig. 3 shows another schematic flowchart;
[0065] Fig. 4 shows a schematic flowchart and illustration of a bottom-fill pack pattern;
[0066] Fig. 5 shows a schematic flowchart and an illustration of a top-fill pack pattern;
[0067] Fig. 6 shows a schematic flowchart and illustration of a block-and-brace packing pattern;
[0068] Fig. 7 shows a function with an input and output area;
[0069] Fig. 8 is an illustration of the described method;
[0070] Fig. 9 shows the upholstery system in a schematic perspective view;
[0071] Fig. 10 shows a training procedure for the function in schematic form; and
[0072] Fig. 11 shows a schematic representation of a packing station and a schematic flowchart.
[0073] Figure 1 schematically shows a method for operating a cushioning system 1000 for inserting cushioning into a shipping container. The method comprises providing 102 information I about a plurality of available cushioning options, wherein the available cushioning options, i.e., the cushioning materials that can be provided, differ from one another in at least one material, size, or structural property.
[0074] Geometric data G of an interior space 902 of the shipping container 900 is acquired by means of a sensor device 300 directed at a packaging opening 904 of the shipping container 900. Subsequently, a data representation R of the interior space 902 of the shipping container 900, comprising a plurality of spatial elements, is determined 106 as a function of the acquired geometric data G.
[0075] Geometric data (G) refers to data that describes the spatial structure of an interior space, particularly one that already contains an object. Specifically, geometric data includes at least a point cloud captured by a sensor device (e.g., a 3D scanner or depth camera). Each captured unit (point) in the point cloud typically contains positional coordinates that allow for a unique assignment of the points in space.
[0076] The data representation R of the interior space is based on the captured geometric data and consists of a plurality of spatial elements, each with specific properties. A "spatial element" can be a discrete section or a logical unit within the created representation, described, for example, by its location, extent, or material composition. The properties of a spatial element can include, among other things, dimensions, geometric shapes, potential obstacles, or existing objects.
[0077] Within the scope of this description, the spatial elements of the shipping container's interior can be represented as part of the data representation, for example, in the form of so-called voxels. The term "spatial element" here refers to a discrete volume or partial volume element used to model and / or simulate the three-dimensional interior of the shipping container. A voxel, in particular, represents the smallest discrete unit within a three-dimensional grid or lattice structure.
[0078] A spatial element includes at least the following parameters: two-dimensional data and a layer affiliation or three-dimensional coordinates; and a status occupied or free.
[0079] A voxel is the three-dimensional equivalent of a "pixel" in two-dimensional image data. While in a two-dimensional image each pixel represents a small image segment with a fixed position on a plane (x, y coordinates), a voxel describes a three-dimensional volume or spatial element with a fixed position in an x, y, and z coordinate system. In other words, a voxel is a data point in a regular three-dimensional grid, where each voxel has a specific size (edge length), shape (usually a cube or cuboid), and a defined position in three-dimensional space.
[0080] A voxel stores properties. For example, it records for each voxel whether it is free, partially occupied, or fully occupied (e.g., by cargo or cushioning material). Furthermore, other parameters can be stored, such as material properties (e.g., foam, paper, air cushion), color values, expected density, expected compressibility, or other physical quantities.
[0081] In the context of this description, voxel data is used to divide the interior of the shipping container into discrete subvolumes, which allows for the calculation, simulation, and evaluation of various arrangements of cushioning materials (as well as, if applicable,
[0082] (charging objects). A voxel grid makes it possible, in particular, to determine for each individual voxel whether it is currently free or occupied. This information can be used to determine cushioning actions or within the framework of a reinforcement learning approach described below.
[0083] Accordingly, a voxel forms a central structural element for the description, since every upholstery action (e.g., the insertion of a specific upholstery material at a defined position in a particular pose) subsequently results in the filling of those voxel units that are covered by the upholstery material. In this way, the subsequent state of the simulated interior can be unambiguously described and stored in a suitable data model, which in turn enables automated evaluation and incremental improvement of the upholstery strategy.
[0084] Advantages of voxel modeling in the context of the invention include, in particular:
[0085] High-resolution representation: By using a voxel grid, the interior of the shipping container can be represented in many small spatial elements. The finer the resolution, the more precise the simulation.
[0086] Simple collision detection: Voxels are very well suited for collision checks, as the contents of individual voxels quickly reveal whether upholstery material overlaps with other objects. Versatile storage options: Each voxel can carry multiple attributes (e.g.,
[0087] Occupancy level, material type, density), which allows different aspects such as stability, cushioning density or costs to be represented within a uniform data structure.
[0088] Easier calculation of spatial properties: Information such as the total volume of the padding or the fill level can be determined relatively easily and quickly by summing the occupied voxels.
[0089] Support for advanced learning algorithms: Reinforcement learning algorithms can build well on voxel-based representations, as it is clearly visible for each training step how an action affects the voxel allocation and what reward / punishment this results in.
[0090] Variable voxel size: This allows the accuracy to be adjusted depending on the application. This takes into account that different cushioning materials may require different tolerances, which can only be captured with smaller voxel sizes.
[0091] For the reasons mentioned, using a voxel-based approach for the interior of a shipping container is advantageous. It should be emphasized that alternative forms of discrete spatial units—such as polygon-based meshes or segmented spatial elements of irregular shape—could also be used. However, the term "spatial element" is explicitly intended not to be limited to voxels, but to include them as a preferred, though not necessarily the only, option.
[0092] Generating the majority of planned padding actions P is done using a function F. This function F can be a machine-trained function or, in another example, an untrained function for determining the padding actions P. For example, a deterministic function F, which always produces the same result for identical inputs, can be used as an untrained, hard-coded function F. Similarly, a non-deterministic function can be used as an untrained function F, which makes decisions to determine the planned padding actions P partly based on rules and partly based on randomness.
[0093] The available interior space is divided into a three-dimensional voxel grid, with each voxel representing a discrete point in space. The analysis is performed layer by layer in the Z-direction from bottom to top, i.e., from the point in the available interior space closest to the bottom of the shipping container towards the opening of the shipping container.
[0094] For each Z-height, i.e., each layer or layer package, several predefined packing strategies are evaluated. These strategies take into account different geometries of the cushioning material and their placement options, which are loaded from a configuration file.
[0095] Each strategy is evaluated based on a fill rate (fill_quota), which quantifies the ratio between used and available space in the respective shift. The strategy with the highest space utilization is selected.
[0096] The system takes into account already placed objects (contents) and avoids overlaps. The selected dunnage elements are positioned to ensure optimal stabilization of the load, i.e., the shipped products, while simultaneously utilizing the available space efficiently.
[0097] The calculated positions are stored in a sequence and can then be visualized and / or used to control packaging robots. The system also allows for dynamic adjustment of the container height by reducing volume when upper layers remain unused.
[0098] Function F operates based on spatial elements, which in one example can be mapped to a specific volume. Each spatial element has a defined height, depth, and width. Across different carton sizes, the function uses spatial elements of the same dimensions. Those spatial elements that lie outside the interior boundaries of a specific shipping container (size 900) are marked as occupied or used before or during the application of function F.
[0099] The respective volume of the spatial elements is characterized, for example, by a width, height, and depth within a range between and including 0.1 cm and 5 cm, in particular between and including 0.5 cm and 4 cm, and especially between 1 cm and 3 cm. In one example, each spatial element has an identical size.
[0100] In one example, the height of the respective room element is less than the height of an available cushion.
[0101] In another example, the height of the upholstery material is greater than a multiple of the height of the respective room elements. This allows several superimposed layers of room elements to be advantageously combined into a single layer in which upholstery material is placed. At the same time, it is possible to identify partially covered areas of a layer and, if necessary, to cover the partially covered area with upholstery material using compressible material.
[0102] In another example, the height of the respective room element corresponds approximately to the expected height of the available upholstery material(s).
[0103] This makes the process easier if only one level of padding material is used.
[0104] The generation of the plurality of planned cushioning actions P is carried out separately for each of a plurality of layers #, wherein the layers # each run perpendicular to an imaginary packing axis PA of the shipping container 900 and each comprise a subset of the plurality of room elements.
[0105] A cushioning action (P) is defined as an action or sequence of actions related to cushioning an item to be shipped and / or the interior of the shipping container. At a minimum, a cushioning action includes selecting a specific cushioning material option and defining a pose for placing and positioning that material. Furthermore, a cushioning action may include additional steps such as cutting, inflating, reshaping, or activating the cushioning material, as well as coordinating the timing of multiple cushioning steps.
[0106] In this context, the term "pose" describes the specific spatial position and orientation of a piece of upholstery within the interior space. This includes specifying both the coordinates of its position and its orientation (e.g., using rotation angles or quaternions). A pose can also refer to a temporal sequence in relation to other poses, thus allowing for the description of a temporal progression through multiple pose states (e.g., step-by-step placement).
[0107] A plurality of planned cushioning actions P is generated based on the determined data representation R of the interior 902 of the shipping container 900 and based on the provided information I on cushioning options 108, wherein each of the planned cushioning actions P includes at least one of the available cushioning options and a pose of the at least one cushioning option in relation to the spatial elements of the interior 902 of the shipping container 900.
[0108] Subsequently, the majority of planned cushioning actions P are provided to introduce several cushioning materials according to the assigned cushioning option via a packing opening 904 of the shipping container 900 into the interior 902 of the shipping container 900 and to position them in the interior 902 of the shipping container 900 according to the assigned pose.
[0109] In one example, no postal action P may be required in a layer, especially if the layer in question is already assessed as full or occupied by room elements.
[0110] In connection with the representation R of the interior 902, the pose defines an area of the interior of the shipping container 900 in which the cushioning material to be provided, assigned according to the planned cushioning action, is to be positioned, the area being represented at least partially by one or more spatial elements marked as still free.
[0111] The majority of planned cushioning actions P is subdivided into a plurality of layers # provided 110, with each layer # being perpendicular to an imaginary packing axis PA of the shipping container 900 and each comprising a subset of the plurality of space elements.
[0112] The imaginary packing direction or packing axis PA runs perpendicular to a bottom 906 of the shipping container 900 and / or perpendicular to the packing opening 904 of the shipping container.
[0113] The generation of the plurality of planned cushioning actions P thus comprises: identifying a subset of the plurality of spatial elements that form a layer, wherein the layer optionally follows a) an imaginary plane parallel to the packing opening 904 of the shipping container 900, which leads into the interior 902 of the shipping container 900, or b) a plane parallel to the floor 906 of the interior 902 of the shipping container 900, or c) a plane parallel to a reference plane of the shipping container 900 or the cushioning system defined otherwise; and generating at least a part of the plurality of planned cushioning actions P, in particular per layer, based on the identified subset of the plurality of spatial elements and based on the provided information on cushioning options.In other words, the volume of the interior of the shipping container is divided into a plurality of imaginary planes running parallel to the bottom of the shipping container or to the access or packing opening of the shipping container.
[0114] A layer can comprise several stacked sub-layers of spatial elements or a single layer of spatial elements.
[0115] An example is characterized by the fact that the identification of the subset of the plurality of room elements and the generation of at least a part of the plurality of planned upholstery actions in a repeated or parallel execution mode is carried out at least once or only once per layer, with at least one or a plurality of planned upholstery actions being provided per layer.
[0116] Regarding the steps of identifying and generating, in one example a single processing run is carried out per shift in order to have a plan of the cushioning materials to be arranged for each shift.
[0117] For packaging scenarios where the internal arrangement is changed by the introduction of cushioning materials, in one example, further geometric data is recorded after the introduction of a first layer of cushioning materials in order to generate cushioning actions for the second layer following the first layer.
[0118] In another example, it is possible to carry out several processing runs in repeated and / or parallel sequences per shift in order to achieve a refined or optimized planning of the cushioning actions for the shipping container through iterative or simultaneous evaluation steps.
[0119] The layer-by-layer generation 108 of the majority of planned cushioning actions P is additionally based on a preferred direction x, y for the longitudinal orientation of the cushioning options in the interior 902 of the shipping container 900, and wherein the preferred directions x, y of adjacent layers differ from each other, in particular by 90°.
[0120] The provisioning includes the output of the at least one planned cushioning action to the automated cushioning system 1000, which comprises: providing 126 an available cushioning material Q associated with the cushioning option of the at least one planned cushioning action P by means of a cushioning material provisioning device 400 of the cushioning system; placing 128 the provided cushioning material Q according to the determined position in the interior of the shipping container 900 by means of a handling device 500 of the cushioning system 1000.
[0121] In one example, the provision 126 of the cushioning material includes a manipulation step prior to placing the cushioning material Q inside the shipping container, wherein the manipulation step prior to placing the cushioning material Q includes reshaping the cushioning material using the handling device and / or a separate reshaping device.
[0122] In one example, the placement 128 of the cushioning material comprises a manipulation step, in particular after the placement 128 of the cushioning material, in the interior of the shipping container, wherein the placed cushioning material Q is reshaped by means of the handling device and / or a separate forming device.
[0123] For example, the placed cushioning material Q is further compressed by temporarily applying a force towards the bottom of the shipping container.
[0124] A parameter for manipulating the respective cushioning material provided by the cushioning material dispenser includes, for example, compression in at least one spatial dimension, or single or multiple folding of a provided cushioning material. This at least one parameter for manipulation is part of the cushioning material option available for selection. In other words, one of the cushioning material options already includes the at least one parameter for manipulation. Of course, not every cushioning material option needs to be manipulated; rather, it can be picked up by the cushioning material dispenser and placed in the shipping container in its provided form.
[0125] The placement 128 of the at least one provided cushioning material Q is carried out layer by layer, by starting an arrangement of cushioning materials Q in a subsequent layer #2 only when an arrangement of cushioning materials Q in a preceding layer #1 of the subsequent layer #2 is completed.
[0126] A control unit 600 controls the components of the cushioning system 1000. A non-volatile, computer-readable medium 600M is part of the control unit 600, wherein the medium 600M stores instructions which, when executed by one or more processors 600P, cause the one or more processors 600P to execute the method according to one of the preceding claims. The generation of the cushioning actions is performed layer by layer in one example. This means that the spatial elements are grouped into layers, with the layers running, for example, parallel to the bottom of the shipping container. By dividing the system into layers, the cushioning problem is simplified by one spatial dimension, which is why the calculation is simplified and is performed, for example, on the basis of two-dimensional pixels that correspond to the voxels of a layer.
[0127] The elements described in this text as cushioning materials can alternatively be referred to as protective packaging or protective packaging materials. Similarly, other terms derived from the concept of cushioning materials, such as cushioning material configuration or cushioning material parameters, can also be referred to as protective packaging material configuration or protective packaging parameters.
[0128] The term "packing opening 904" refers to the closable opening of the shipping container 900, which serves to insert or remove items or cushioning material into the interior of the container. This opening can be formed, for example, by open cover flaps, a lid, or similar, and allows access to the interior.
[0129] The interior space 902 is the cavity within the shipping container 900 into which items and / or cushioning material can be placed. This interior space may already be partially filled with other items; the described cushioning procedures therefore take into account empty, partially filled, or fully packed containers.
[0130] A shipping container 900 is a receptacle for shipping goods, for example in the form of a carton, box, crate, or similar packaging. The shipping container 900 has at least the interior space 902 and the loading opening 904, through which items are filled and removed. The external shape, size, and material of the shipping container can vary and are typically tailored to the items being shipped and the intended transport conditions.
[0131] Figure 2 shows a schematic flowchart illustrating the determination of planned cushioning actions and the arrangement of the cushions in the shipping container. Generating the plurality of planned cushioning actions P comprises: determining a start layer #1 within the data representation as the starting point for layer-wise generation; determining an end layer #3 within the data representation as the endpoint for layer-wise generation; and performing the layer-wise generation of zero or more planned cushioning actions in the start layer, in the end layer, and in layers #2 located between the start layer #1 and the end layer #3.
[0132] The final layer is reached when the desired fill level of the shipping container is achieved. Therefore, the final layer can be determined based on the desired fill level.
[0133] Step 116 checks whether there is another layer for which planned padding actions should be determined. In other words, if the other layer is the starting layer, the ending layer, or an intermediate layer, the process proceeds to step 118. In step 118, specific padding actions are then determined for the existing layer.
[0134] In one example, it is provided that the determined planned cushioning actions P for the majority of layers # are made available 110 after performing the layer-wise generation 108 in order to position the cushioning materials assigned to the respective cushioning option P according to the respective assigned pose in the shipping container 900.
[0135] Based on the cushioning actions for each layer provided in step 110, the planned cushioning actions are implemented in step 120. Preferably, work begins with the bottom layer, which is closest to the bottom of the shipping container. In step 122, it is checked whether another layer needs to be processed for cushioning. If so, step 124 checks whether there are any further cushioning actions to be carried out. If so, cushioning material is provided in step 126. Providing the cushioning material in step 126 can include, for example, cutting a roll of material such as paper to length, creating an air cushion, or manipulating a paper material, such as crumpling it. Once the cushioning material is provided in step 126, it is placed in the shipping container in step 128 according to the determined position.
[0136] Once all planned padding actions have been carried out on all layers, step 120 is complete.
[0137] In a further development of the example, the procedure includes: acquiring 134 additional geometric data of the interior of the shipping container using the sensor device; determining 136 a further data representation of the interior of the shipping container, encompassing a plurality of spatial elements, depending on the acquired further geometric data; determining 138 a cushioning quality based on the further data representation of the interior of the shipping container 900; closing 140 the shipping container 900 if the cushioning quality is rated as sufficient 144, or directing 142 the shipping container 900 to manual inspection if the cushioning quality is rated as insufficient 144.
[0138] In one example, the starting layer of the data representation of the interior of the shipping container is arranged between a layer representing the bottom of the shipping container and the final layer of the data representation of the interior of the shipping container.
[0139] Figure 3 shows another schematic flowchart. Unlike Figure 2, the process in Figure 3 involves intermediate steps to check whether the arrangement of padding materials up to a certain point in time is sufficient, in order to then plan further padding actions depending on this intermediate state. For further details, please refer to the description in Figure 2.
[0140] In steps 112 and 114, a number of layers is defined, which does not include the total number of layers in the shipping container, nor all the layers to be cushioned, but only a subset. After the cushioning has been arranged according to the planned cushioning actions in step 120, step 150 checks whether another subset of layers still needs cushioning. If so, the process returns to step 112 to re-scan the interior, now with the cushioning arranged there.
[0141] It is intended that, after determining zero or multiple cushioning actions for one or more layers, the procedure includes: acquiring further geometric data of the interior of the shipping container using the sensor device; determining a further data representation of the interior of the shipping container comprising a plurality of spatial elements, depending on the acquired further geometric data, wherein, according to the further data representation of the interior of the shipping container, at least some of the assigned spatial elements are occupied and at least some other spatial elements are unoccupied; and continuing to generate at least one further planned cushioning action, depending on the further data representation and depending on the available cushioning options.If, however, step 150 reveals that all planned cushioning actions have been carried out and all planned shifts have been completed, the final assessment of the overall cushioning of the shipping container is performed.
[0142] Figure 4 shows an embodiment of the described method for carrying out a bottom-fill packing pattern. According to this packing pattern, the bottom 906 of the shipping container 900 is packed with a number of cushioning materials Q. Typically, no item is placed in the shipping container 900 for this packing pattern.
[0143] The bottom-fill packing pattern is specified based on an identifier of the shipping container and / or a predefined configuration of the cushioning system. The process according to the bottom-fill packing pattern is carried out layer by layer, beginning with a start layer associated with the bottom-fill packing pattern and ending with a finish layer associated with the bottom-fill packing pattern. The process comprises the following steps:
[0144] Determine 112 of the start layer associated with the bottom-fill pack pattern within the space elements, wherein this start layer includes the space elements that are directly adjacent to the bottom of the shipping container.
[0145] Determine 114 of the final layer associated with the bottom-fill packing pattern within the spatial elements, whereby this final layer is determined depending on a predetermined number of layers and / or based on the identifier of the shipping container.
[0146] Performing 108 of the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the bottom-fill pack pattern, the end layer and the intermediate layers associated with the bottom-fill pack pattern.
[0147] Figure 5 illustrates the execution of a described method for determining a top-fill packing pattern. According to this packing pattern, one or more layers of cushioning material Q#1-4 are arranged over items G#1, G#2 already in the shipping container.
[0148] The procedure involves determining a top-fill packing pattern based on a shipping container identifier and / or a predefined configuration of the cushioning system. The packing is carried out layer by layer according to the defined top-fill packing pattern, beginning with a start layer associated with the pattern and ending with a finish layer also associated with the pattern. The procedure comprises the following steps:
[0149] Determine 112 of the starting layer associated with the top-fill packing pattern within the space elements, where the assigned space elements of the interior of the shipping container within this starting layer are marked as empty. This starting layer borders a layer with space elements, at least one of which is marked as occupied.
[0150] Define 114 the end layer associated with the top-fill packing pattern within the spatial elements depending on a predefined parameter and / or an identifier of the shipping container and / or an imaginary plane in which the opening of the shipping container is located. The end layer associated with the top-fill packing pattern is located in the area of the opening of the shipping container or adjacent to this area.
[0151] Performing 108 of the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the top-fill pack pattern, the end layer and the intermediate layers associated with the top-fill pack pattern.
[0152] In the example shown, the inserted cushions Q#1 and Q#3 fell onto the object G#1 on the left, but were placed in the same layers as cushions Q#4 and Q#2 according to the packing pattern.
[0153] Figure 6 illustrates the execution of a method for determining a block-and-brace packing pattern. According to this packing pattern, cushioning material Q#l-8 is placed both between the items G#l, G#2 already arranged in the shipping container 900 and between these items and the side walls of the shipping container 900.
[0154] The procedure comprises specifying a block-and-brace packing pattern based on an identifier of the shipping container 900 and / or a predefined configuration of the cushioning system. The procedure is carried out layer by layer according to the block-and-brace packing pattern, beginning with a start layer #1 associated with the block-and-brace packing pattern and ending with a final layer #3 associated with the block-and-brace packing pattern. The procedure comprises the following steps: Determining the start layer #1 associated with the block-and-brace packing pattern within the space elements, wherein this start layer includes at least one space element of the interior of the shipping container 900 marked as empty. The start layer #1 is either adjacent to a layer of space elements marked as occupied or directly adjacent to the bottom of the shipping container.
[0155] 114 determine the end layer #3 associated with the block-and-brace packing pattern within the space elements, wherein this end layer comprises at least one space element marked as empty and at least one space element marked as occupied within the interior of the shipping container. End layer #3 borders a layer within the space elements that either comprises exclusively space elements marked as empty or elements indicating an end facing away from the bottom of the shipping container of an item placed in the shipping container.
[0156] Performing 108 of the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the block-and-brace packing pattern, the end layer and the intermediate layers associated with the block-and-brace packing pattern.
[0157] In step 106 of Figure 1, the data representation RI is determined in an initial state. In this initial state, only items G#1 and G#2 are present in shipping container 900. Within the data representation RI, which divides the interior of shipping container 900 into layers #1-5, items G#1 and G#2 are marked as occupied in their respective layers. The occupied status is represented in Figure 6 by black filled boxes. Accordingly, layers #4 and #5 are occupied only by space elements marked as free.
[0158] The data representation RI serves as the basis for defining the start layer #1 and the end layer #3 in step 156. In step 156, the layers relevant to the planned block-and-brace packing pattern are identified. Layers #1-#3 are identified as relevant, encompassing both free and occupied space elements. These layers are suitable for filling the free space elements with cushioning material, thereby securing and cushioning the items G#1 and G#2 located in shipping container 900. During generation 108, the spaces marked as free are assigned planned cushioning actions P#1-7 to place the cushioning material Q#1-8 in the respective free areas after generation.
[0159] For example, the cushioning material Q#7 can be seen to extend into layer #4 because object G#1 protrudes from layer #2 into layer #3. Due to the compressible nature of the cushioning materials, such inaccuracies are tolerable.
[0160] The packing patterns described above can of course be applied individually or in combination to a shipping container.
[0161] For example, the block-and-brace packing pattern can be used in the first section of the layers, followed by the top-fill packing pattern in a subsequent section. In this case, one of several maximum fill heights is predetermined, so that after the cushioning material is arranged according to the planned cushioning actions, the shipping container can be closed at a specified height.
[0162] Figure 7 schematically shows the function F for determining the padding actions P# per layer #. Generating 108 of the at least one planned padding action P comprises: determining at least a part of the plurality of padding actions P in an output area F_O of a function F, wherein the representation R of the plurality of spatial elements of the interior of the shipping container or a part thereof, in particular the spatial elements R# of a respective layer #, is provided in an input area F_I of the function F, and wherein the information I on the plurality of available padding options is provided in the input area F_I of the function F.
[0163] The machine-trained function F operates based on layers # of the interior space. This means that for each determination of upholstery actions P # per layer #, the spatial elements of this layer are provided in the input area F_I in the form of the representation R#. Additionally, the information I about the available upholstery options is provided in the input area F_I. At least one or more hidden layers F_H of the network were previously trained using a machine learning method and are set to a corresponding state based on the state of the input area. Based on the hidden layers F_H, the result is provided in the output area F_O in the form of one or more planned upholstery actions P# per layer # of the interior space.An example is characterized by the fact that in the output area F_0 of the function F, those planned cushioning actions P are provided which are assigned to one of the plurality of layers of the room elements, wherein in the input area F_I of the function F, the information I on the available cushioning options and the representation R of the interior of the shipping container are provided which is assigned to one of the plurality of layers # of the interior.
[0164] Figure 8 illustrates the application of function F from Figure 7. For example, the information I regarding the available cushioning options is already available with a pre-selected orientation relative to the interior of the shipping container. An object G is positioned in the dark-marked area of the interior layer, as represented by R#. The task of function F is therefore to determine a suggested arrangement of cushioning materials in the free areas, i.e., a planned cushioning action P for each cushioning material. The planned cushioning actions P are thus provided in the output area of function F. In the example shown, function F suggests the arrangement of cushioning materials available according to information I in the areas labeled 1, 2, and 3.
[0165] Figure 9 shows an example of the cushioning system 1000 from Figure 1. A conveying device 200 includes, for example, a roller conveyor and other elements that control the flow of the shipping containers 900.
[0166] In an entrance section 210, a light barrier 212 registers the arrival of a shipping container 900. If the cushioning system 1000 is ready to receive another shipping container 900, the gate 214, schematically depicted as a barrier, is opened.
[0167] The sensor device 300 includes, for example, a camera for capturing a digital image. The camera is directed towards the opening of the shipping container 900, which provides a view of the interior of the shipping container 900, so that the generated digital image contains information in the form of geometric data G from Figure 1, including the arrangement of shipping products 910, the arrangement of the inner walls of the shipping container 900, and the arrangement of the empty spaces between the inner wall and one of the shipping products 910 or between several shipping products 910. A lighting device 310 is associated with the sensor device 300, which emits light towards the opening of the shipping container 900. Of course, other sensor devices, such as ultrasonic sensors or similar devices, can also be used to determine the geometric data G.
[0168] In the example shown, the 400a-b upholstery dispensers provide Q#ab upholstery material in the form of paper pads of varying widths. The 400c-d upholstery dispensers provide Q#cd upholstery material in the form of air cushions of varying widths.
[0169] A supply of raw material for the production of a paper pad as cushioning material comprises, as in the case of the cushioning material dispensers 400a-b, for example, sheet-shaped material which is provided in a compact form, such as by means of a roll of sheet-shaped material or by means of a rectangular stack of zigzag-folded sheet-shaped material.
[0170] In another example, according to the cushioning material dispensers 400c-d, the starting material is plastic film and is supplied in a compact form, e.g., on a roll. The plastic film is converted into bags, inflated with compressed air, and sealed.
[0171] Other examples of upholstery dispensers not shown include, for example, dispensers for foam upholstery, inflatable upholstery, or pourable individual upholstery.
[0172] In one example, a sensor is provided in each upholstery dispenser 400 to detect when the raw material for producing the respective upholstery material Q is depleted or when a particular upholstery material type is no longer available in that specific dispenser. This information is then transmitted to the control unit 600 as information about the available upholstery options.
[0173] In the example shown, the handling device 500 comprises a robot arm 510 fixed to a ceiling above the conveyor 200 and a robot hand 520 as a gripping device. Alternatively or additionally, a vacuum or negative pressure gripper can be used instead of the robot hand.
[0174] The conveying device 200 comprises a fixing system 222, which fixes the shipping container 900 in the cushioning section 220 of the conveying device 200 at least during the period in which the at least one cushioning material 800 is introduced into the associated space 920. In the example shown, the fixing system 222 comprises a gate 224, schematically depicted as a barrier, which stops the shipping container 900 being conveyed by the conveying device 202 in a conveying direction 204. A slide 226 presses the stopped shipping container 900 against a wall 228 opposite the slide 226, thereby achieving the cushioning position of the shipping container 900 in the cushioning section 220.
[0175] In another example, the handling device 500 includes a hold-down device 700 for holding down the products defining the empty space, in order to prevent unintentional displacement of the products during the insertion of the cushioning material Q. The hold-down device 700 secures the at least one shipping product 910 in the shipping container 900, at least during the arrangement of the at least one cushioning material 800. The schematically depicted hold-down device 700 is fixed horizontally above the cushioning section 220 and can also be designed as a robot arm with a distal fixing end.
[0176] A slide gate 242 is provided which, in the event of a fault detected with respect to the shipping container 900, pushes at least one shipping container 900 into the isolation section 240. The slide gate 242 is activated by the control unit 600 to move the shipping container 900 into the isolation section 240 if the cushioning quality is deemed insufficient. From the isolation section 240, the shipping containers 900 are manually removed, inspected, manually re-cushioned if necessary, and then returned to the input section 210. Then, automatic re-cushioning may occur, or the control unit 600 determines that sufficient cushioning is present, and the shipping container 900 is made available in the output section 230 for forwarding to a sealing station for sealing.
[0177] The reward system evaluates the placement of dunnage elements based on several factors: It considers the number of cells covered (coverage_reward), a bonus for contact points with other elements or edges (contact_bonus), and future potential (future_potential). The latter is comprised of move retention, diversity of available element types (diversity_retention), and space efficiency (space_efficiency). At the end of the game, an additional completion bonus is awarded based on the overall coverage ratio. Invalid moves are penalized with a negative reward of -1. The various components are weighted and combined, with move retention, element diversity, and space efficiency contributing to the overall score with different factors.
[0178] Figure 10 shows a computer-implemented method for training the function F. The training includes providing a simulation environment S to simulate various states of the interior space 902 of the shipping container 900, which is divided into spatial elements. According to an initial state of the interior space 902, at least some of the assigned spatial elements are unoccupied. Information I about a plurality of available cushioning options is provided, wherein the available cushioning options differ from one another in at least one material, size, or structural property.
[0179] Subsequently, in 2008, an agent A generates a plurality of cushioning actions P based on at least one state of the simulated interior 902 of the shipping container 900, in particular based on the initial state, and based on the provided information I on cushioning options, wherein each planned cushioning action P comprises a cushioning option and a pose in relation to the spatial elements of the interior 902 of the shipping container 900.
[0180] By applying the plurality of cushioning actions P to the simulated interior 902 of the shipping container 900, a subsequent state Z of the simulated interior 902 of the shipping container 900 is generated, wherein plurality of cushioning actions P comprises a respective arrangement of a simulated cushioning material in at least one or more spatial elements marked as free, according to the associated pose.
[0181] In 2012, a cushioning quality RW is determined as a function of the subsequent state Z of the simulated interior 902 of the shipping container 900 caused by at least one cushioning action P.
[0182] Subsequently, at least one parameter of the function F to be trained, which controls the agent, is adjusted in 2014 using the determined cushioning quality RW and the associated planned cushioning actions P, whereby the procedure is repeated at least comprehensively the steps of generating, applying, determining and adjusting in order to improve the decision-making behavior of agent A.
[0183] The generation of the majority of planned upholstery actions P in 2008 includes:
[0184] Identify in 2016 a subset of the plurality of spatial elements which are part of a layer, wherein the layer optionally follows a) a plane parallel to an access opening of the shipping container leading into the interior of the shipping container, or b) a plane parallel to a floor of the interior of the shipping container, or c) a plane parallel to a reference plane of the shipping container or the cushioning system defined otherwise; and generate in 2018 at least a part of the plurality of planned cushioning actions, in particular per layer, based on the identified subset of the plurality of spatial elements and based on the information provided on cushioning options.
[0185] During training, the function also operates in layers, with each detected cushioning action creating a new state of the shipping container's interior. Refer to the preceding description for using the function.
[0186] In one example, the cushioning quality RW is determined in terms of a greater reward for the identified cushioning actions P, the lower the number of overlapping edges or distal ends of cushioning materials in adjacent layers running parallel to the bottom of the shipping container #.
[0187] In one example, the function F to be trained is implemented as a policy function in the form of a deep neural network, which maps a given state of the simulated interior 902 of the shipping container 900 directly to a probability distribution over cushioning options and poses, and a policy gradient method is used to optimize this policy function.
[0188] In this example, instead of a value-function-based method such as Q-Learning, a direct policy approach is chosen. Here, the function to be trained is represented as a neural network with parameters. This neural network receives as input a representation of the current state of the shipping container, e.g., the current padding situation, available free space, and previously placed padding materials.
[0189] The network output consists of a distribution of possible padding actions. Each possible action comprises the selection of a padding material and its corresponding pose, e.g., position, orientation. The policy thus indicates the probability with which a specific action will be executed in the current state.
[0190] The optimization of this policy function is performed using a so-called policy gradient method. For this purpose, a suitable objective function is defined, maximizing the expected cumulative reward (in this case, the cushioning quality). This is typically done using gradient descent methods, where the gradient of the expected reward is estimated with respect to the network parameters. An update is then performed. In another example, the deep neural network representing the agent's function F includes at least one convolutional layer to process the spatial elements of the interior space. Further hidden layers of the deep neural network employ fully connected layers to derive a suitable cushioning action from the extracted features.
[0191] Upholstery planning requires considering the spatial information of the interior space. This space can be represented as a 2D grid or as a 3D volumetric representation. A convolutional neural network (CNN) is well-suited for effectively extracting local features such as free spaces, occupied areas, and the positions of objects.
[0192] One possible implementation looks like this: a Input layer: A multidimensional input signal, e.g., a 2D or 3D matrix representing the interior. In one example, the input layer is fed with a layer parallel to the bottom of the shipping container, or its spatial elements. b Convolutional layers: Several successive convolutional layers that apply specific filters to the input data to detect relevant features. c Pooling layers (optional): Reduce the spatial dimension to decrease computational effort and improve generalization. d Fully-connected layers: One or more layers of neural networks that function as feature combination and classification / action selection. d Output layer: Outputs either a probability distribution over different cushioning options and poses (policy network) or an evaluation Q or value function.
[0193] One example is characterized by the fact that, in determining the cushioning quality RW, in addition to spatial utilization and / or the protection of the goods in the shipping container, further parameters are taken into account, in particular a cost factor of the cushioning materials and / or a stability assessment.
[0194] To enable realistic optimization, cushioning quality is not defined solely by the fill level or space utilization, but also incorporates additional factors. For example, the cost factor can be considered, aiming to minimize the consumption of cushioning material. In this case, using a large amount of potentially expensive cushioning material reduces the reward. Similarly, a weight parameter can be taken into account if weight optimization is important for shipping. Furthermore, a stability assessment can be included to determine whether the applied cushioning material adequately secures the goods being protected. Additionally or alternatively, a shock absorption assessment is possible, examining how well the cushioning materials are able to absorb vibrations or impacts during transport.All this information is combined in the reward function "cushioning quality" to obtain a holistic evaluation of the cushioning actions performed.
[0195] In another example, it is provided that during the simulation of the interior of the shipping container, a physical collision detection is carried out, which checks whether the planned cushioning material can be placed correctly when applying the respective cushioning action, and in the event of a collision or...
[0196] Overlap reduces cushioning quality or creates a penalty as a negative reward signal for the agent.
[0197] In another example, it is proposed that the cushioning quality is additionally determined depending on a regulatory parameter, which, for example, limits the amount of cushioning material used or ensures compliance with specified environmental or shipping guidelines by including a corresponding penalty component or an additional reward component factor as part of the cushioning quality.
[0198] Such a regulatory parameter could include, for example: a) Limiting the amount of cushioning material used: If too much material is used per action, a penalty component increases or the reward decreases; b) Reflecting environmental or transport regulations: Certain cushioning materials are only permitted under certain restrictions, or packaging must meet specific safety standards. Penalty terms for rule violations can also be integrated here; c) Considering time or capacity limitations: If laying or placing cushioning material becomes problematic when a time limit is exceeded, a corresponding negative term can reduce the reward.
[0199] An example is characterized by the fact that the function to be trained, which controls the agent, is implemented as a Q-function approach using a Q-learning method, whereby for each possible combination of the state of the simulated interior and a cushioning action, a Q-value is determined that represents the expected cushioning quality, and that the updating of the Q-values is based on the receipt of a reward signal, which is based on the cushioning quality of the resulting subsequent state.
[0200] In this embodiment, a so-called Q-learning algorithm is used in each training step. The agent is provided with a state space that includes all relevant parameters of the simulated interior of the shipping container. These can be, for example, the positions of already placed cushioning materials, the remaining free space, or a metric for the overall cushioning quality achieved so far. Furthermore, an action space is defined in which each possible cushioning action (consisting of selecting a cushioning material option and its position within the interior) is represented as an individual action.
[0201] In Q-learning, Q-values (quality values) are recorded for each possible combination (state, action). A Q-value indicates the expected cumulative reward, in terms of the cushioning quality (RW), when performing a specific action in a given state and subsequently following the optimal strategy. In this case, the reward corresponds to the cushioning quality (or a value derived from it) that the agent receives for its actions.
[0202] In another example, the function to be trained is implemented as an actor-critic architecture, in which an actor network generates the padding actions depending on the current state of the simulated interior space, and a critical network provides a value function (value function or Q function) to evaluate the padding quality and update the parameters of the actor network.
[0203] This variant combines the advantages of value-based and policy-based methods. Actor and Critic are separate but closely interconnected neural networks:
[0204] The actor (policy network) proposes a specific cushioning action based on the current state. This state is then processed via input layers (e.g., fully networked layers, CNN, RNN), so that the actor network ultimately outputs a probability distribution of possible cushioning actions or a deterministic action (in the case of a deterministic policy).
[0205] The Critic (value function network) evaluates the selected action in terms of its expected cushioning quality. For this purpose, the Critic can, for example, calculate the Q-value or the Value-value for the state or state / action pair. Training proceeds as follows: The Critic provides a learning signal (TD error or temporal difference) which is used to optimize the parameters of the Critic network and subsequently the policy parameters of the Actor network. This ensures that the Actor makes only small, guided updates based on the respective feedback from the Critic, which usually results in a more stable learning curve than purely value- or purely policy-based methods.
[0206] In another example, it is proposed that the agent be subjected to domain randomization to avoid overfitting and to increase generalization capability, by stochastically varying different parameters of the simulated environment during training, in particular the dimensions of the shipping container, the type and number of goods, and the available cushioning options.
[0207] In machine learning, especially in reinforcement learning, there is a risk that the agent will be overfitted to a specific configuration of the problem and therefore only be usable to a limited extent in real application scenarios.
[0208] The domain randomization described in this example addresses this problem by regularly making random variations to the environment parameters during the training process. For example, the interior dimensions (width, height, depth), the number and arrangement of goods, or the available padding materials are changed.
[0209] This exposes the agent to a wide range of possible scenarios, forcing it to learn general strategies rather than a rigid solution tailored to a specific configuration. As a result, the generality and robustness of the learned decision-making behavior are increased.
[0210] In another example, it is provided that when adjusting at least one parameter of the function to be trained, a Stochastic Gradient Descent (SGD) or an Adaptive Optimization Method (e.g., Adam, RMSProp) is used to iteratively update the weights of the neural network based on the determined padding quality.
[0211] The training process for the neural network representing the agent typically takes place in several iterations or episodes. In each step, based on the determined cushioning quality, an error term (loss) or a reward difference is calculated to improve the network. Stochastic gradient descent (SGD) evaluates small batches of experiences (transitions of state, action, reward, and subsequent state) in each step. From these, a stochastic gradient of the error term is calculated, which is then used to update the network weights.
[0212] In adaptive optimization methods such as Adam or RMSProp, the learning process is further refined by integrating statistical properties such as first and second moments (mean and variance) of the gradients into the update process. This can accelerate convergence and lead to a more stable learning curve.
[0213] In this way, the network gradually adjusts its parameters to select better actions in relation to the previously defined cushioning quality (reward).
[0214] Figure 11 shows a packing station specifically designed for preparing shipping cartons 900 and enabling an optimized cushioning process. In contrast to the fully automated cushioning process shown in Figure 9, in the example of Figure 11 the packer 1010 is guided to arrange the provided cushioning material.
[0215] Advantageously, the guided upholstery method described below can achieve a consistently desired quality of upholstery while using upholstery material sparingly.
[0216] The cushioning system 1000 is part of the packing station and performs the cushioning functions. The packer 1010 works at a table 1020, on which the shipping container 900 to be processed is positioned. The raw shipping cartons 990a-b, in various sizes, are folded and ready on a shelf 1030, so that the packer 1010 can access a wide variety of outer packaging options at any time.
[0217] The process begins with packer 1010 removing one of the folded shipping cartons 990a-b, opening it, and arranging the shipping products in shipping container 900. This process, which corresponds to order picking, is completed by confirmation of the completed picking in step 2002 on a screen 610. Subsequently, in step 2004, the instruction is given to place the already loaded shipping container 900 under sensor 300, which is a prerequisite for the subsequent data acquisition.
[0218] In the next step 104, the sensor 300 captures the geometric data G of the positioned shipping container 900, while in step 106 the digital representation R of the interior of the shipping container 900 is created. Parallel to this, or before or after, the information I on the available cushioning options is transmitted in step 102.
[0219] In the example shown, the cushioning material dispenser or cushioning material supply device delivers 400 cushioning materials in paper form with a fixed width, fixed height and variable length, which allows the cushioning actions P to be generated later to be optimally adapted to the conditions.
[0220] Of course, additional cushioning material supply facilities can be provided to increase the variety of cushioning materials available at the packaging station.
[0221] In step 108, the control unit 600 evaluates both the recorded geometric data G and the provided information I to generate a plurality of upholstery actions P that precisely determine the further course of the upholstery.
[0222] The results of these calculations, i.e., the planned cushioning actions P, are provided to packer 1010 in step 110, specifically displayed on screen 610. Screen 610 provides packer 1010 with step-by-step instructions (provide 128) and also displays the target position of the cushioning material Q (provide 126) supplied by cushioning material dispenser 400, thus indicating the position of the planned cushioning action P in relation to shipping container 900.
[0223] The step-by-step instructions for the packer are divided into layers, starting with the bottom layer of the interior of shipping container 900, which is to be filled with at least one of each type of cushioning material Q. Of all the layers to be processed, the bottom layer is closest to the bottom of the shipping container.
[0224] Additionally, a sequence for arranging the cushioning materials Q can be specified for each shift. For example, each shift begins with the largest cushioning material Q in terms of surface area, and then the packer 1010 is offered the next smallest cushioning material Q for arrangement in the shipping container 900.
[0225] An exemplary screen image 612 visualizes the shipping container 900 from above, showing the opening 904 and a shipping product G#1 positioned inside the shipping container 900. Based on this representation, the next cushioning action P is visually conveyed to the packer 1010, whereby the planned cushioning action P consists of placing the cushioning material Q as precisely as possible inside the shipping container 900 according to the specified and displayed arrangement position, i.e., the pose.
[0226] After the planned upholstery action P has been carried out, the packer 1010 confirms its completion in step 2006, for example, by tapping the touchscreen screen 602 or via a separate actuation device. Alternatively, multiple planned upholstery actions can be displayed and acknowledged on screen 602. In particular, the planned upholstery actions can be displayed layer by layer. Successful completion can also be automatically verified by further acquiring geometric data of the interior, for example, based on a 3D image – after each step of the packer 1010.
[0227] In step 2008, the system checks whether any further cushioning actions P still need to be performed. If there are any outstanding actions, the process continues accordingly. Otherwise, step 2010 signals to packer 1010 via screen 610 that the entire cushioning process is complete. The shipping container 900 can then be closed and prepared for shipment.
[0228] Optionally, an intermediate check is performed using the sensor device 300 to verify whether the packer 1020 has positioned the cushions Q in the correct location. Refer to the description for Figure 3 for further details.
Claims
Patent claims 1. A method for operating a cushioning plant (1000), the method comprising: Providing (102) information (I) about a plurality of available cushioning options, wherein the available cushioning options differ from each other in at least one size property, in particular in at least one material, size or structure property; Acquisition (104) of geometric data (G) of an interior (902) of a shipping container (900) by means of a sensor device (300); Generating (108) a plurality of planned cushioning actions (P) based on the geometric data (G) of the interior (902) of the shipping container (900) and based on the provided information (I) on cushioning options, wherein each of the planned cushioning actions (P) comprises at least one of the available cushioning options and a pose of the at least one cushioning option in relation to the interior (902) of the shipping container (900); and Providing (110) the majority of planned cushioning actions (P) to introduce multiple cushioning materials according to the assigned cushioning option through a packing opening (904) of the shipping container (900) into the interior (902) of the shipping container (900) and to position them in the interior (902) of the shipping container (900) according to the assigned pose.
2. The method according to claim 1, comprising: Determine (106) a data representation (R) of the interior (902) of the shipping container (900) comprising a plurality of spatial elements depending on the recorded geometric data (G); wherein the generation (108) of the majority of planned cushioning actions (P) is based on the determined data representation (R) of the interior (902) of the shipping container (900) and based on the provided information (I) on cushioning options, wherein each of the planned padding actions (P) includes at least one of the available padding options and the position of the at least one padding option in relation to the spatial elements of the interior (902) of the shipping container (900).
3. The method according to claim 1 or 2, wherein the generation of the plurality of planned cushioning actions (P) is carried out separately for each of a plurality of layers (#), wherein the layers (#) each extend perpendicular to an imaginary packing axis (PA) of the shipping container (900) and each comprise a subset of the plurality of space elements.
4. The method according to one of claims 2 to 3, wherein generating (108) the plurality of planned padding actions (P) comprises: Determine (112) a start layer (#1) within the data representation (R) as the starting point for layer-wise generation; Determine (114) an end layer (#3) within the data representation (R) as the endpoint for layer-wise generation; and Performing (116, 118) the layer-wise generation (108) of zero or more planned padding actions in the start layer, in the end layer and in layers (#2) between the start layer (#1) and the end layer (#3).
5. The method according to claim 4, wherein the determined planned cushioning actions (P) for the plurality of layers (#) are provided (110) after performing the layer-wise generation (108) in order to position the cushioning materials assigned to the respective cushioning option (P) according to the respective assigned pose in the shipping container (900).
6. The method according to claim 5, further comprising: Acquisition (134) of further geometric data of the interior of the shipping container using the sensor device; Determine (136) a further data representation of the interior of the shipping container comprising a plurality of spatial elements, depending on the additional geometric data recorded; Determine (138) a cushioning quality based on the further data representation of the interior of the shipping container (900); and close (140) the shipping container (900) if the cushioning quality is assessed as sufficient (144), or Guide (142) the shipping container (900) for manual inspection if the cushioning quality is deemed insufficient (144).
7. The method according to any one of the preceding claims, wherein the method comprises: Specifying (152) a bottom-fill packing pattern depending on an identifier of the shipping container and / or depending on a predefined configuration of the cushioning system; wherein the procedure according to the bottom-fill packing pattern is carried out over several layers, starting with a start layer associated with the bottom-fill packing pattern and ending with a finish layer associated with the bottom-fill packing pattern, and comprises: Determine (112) the start layer associated with the bottom-fill packing pattern within the space elements, wherein the start layer associated with the bottom-fill packing pattern comprises the space elements which are adjacent to the bottom of the shipping container; Determine (114) the end layer associated with the bottom-fill packing pattern within the spatial elements, wherein the end layer associated with the bottom-fill packing pattern is determined depending on a predetermined number of layers and / or depending on the identifier of the shipping container; and Performing (108) the layer-wise generation of one or more planned cushioning actions in the start layer associated with the bottom-fill pack pattern, in the end layer associated with the bottom-fill pack pattern and in the layers associated with the bottom-fill pack pattern located between the start layer and the end layer.
8. The method according to any one of the preceding claims, wherein the method comprises: Determining (154) a top-fill packing pattern depending on an identifier of the shipping container and / or depending on a predefined configuration of the cushioning system; wherein the procedure is carried out according to the defined top-fill packing pattern over several layers, starting with a start layer associated with the top-fill packing pattern and ending with a finish layer associated with the top-fill packing pattern, and comprises: determining (112) the start layer associated with the top-fill packing pattern within the space elements, wherein within the start layer the associated space elements of the interior of the shipping container are marked as empty and the start layer adjoins a layer with space elements, at least one of which is marked as occupied; Defining (114) the end layer associated with the top-fill packing pattern within the spatial elements as a function of a predefined parameter and / or as a function of an identifier of the shipping container and / or as a function of an imaginary plane in which the opening of the shipping container is located, wherein the end layer associated with the top-fill packing pattern is located in or adjacent to the opening of the shipping container; and Performing (108) the layer-by-layer generation of one or more planned cushioning actions in the start layer associated with the top-fill pack pattern, in the end layer associated with the top-fill pack pattern and in the layers associated with the top-fill pack pattern located between the start layer and the end layer.
9. The method according to any one of the preceding claims, wherein the method comprises: Specifying (156) a block-and-brace packing pattern depending on an identifier of the shipping container (900) and / or depending on a predefined configuration of the cushioning system; wherein the procedure according to the block-and-brace packing pattern is carried out over several layers, starting with a start layer (#1) associated with the block-and-brace packing pattern and ending with a finish layer (#3) associated with the block-and-brace packing pattern and comprises: Determine (112) the start layer (#1) associated with the block-and-brace packing pattern within the space elements, wherein the start layer associated with the block-and-brace packing pattern comprises at least one space element of the interior of the shipping container (900) marked as empty, and wherein the start layer (#1) adjoins a layer of space elements marked as occupied or a bottom of the shipping container; Determine (114) the end layer (#3) associated with the block-and-brace packing pattern within the space elements, wherein the end layer (#3) associated with the block-and-brace packing pattern comprises at least one space element of the interior of the shipping container marked as empty and at least one space element of the interior of the shipping container marked as occupied, and wherein the end layer (#3) associated with the block-and-brace packing pattern adjoins a layer within the space elements that either comprises only space elements marked as empty or comprises space elements that indicate a termination of an item arranged in the shipping container facing away from the bottom of the shipping container; and Performing (108) the layer-wise generation of one or more planned cushioning actions in the start layer associated with the block-and-brace packing pattern, in the end layer associated with the block-and-brace packing pattern and in the layers between the start layer and the end layer associated with the block-and-brace packing pattern.
10. The method according to one of the preceding claims, wherein generating (108) the at least one planned padding action (P) comprises: Determining at least a part of the plurality of cushioning actions (P) in an output area (F_O) of a function (F), wherein providing the representation (R) of the plurality of spatial elements of the interior of the shipping container or a part thereof, in particular the spatial elements (R(#)) of a respective layer (#), in an input area (F_I) of the function (F), and wherein providing the information (I) on the plurality of available cushioning options in the input area (F_I) of the function (F).
11. The method according to the preceding claim, wherein in the output area (F_O) of the particularly machine-trained function (F) those planned cushioning actions (P) are provided which are assigned to one of the plurality of layers of the spatial elements, and wherein in the input area (F_I) of the function (F) the information (I) on the available cushioning options and the representation (R) of the interior of the shipping container are provided which is assigned to one of the plurality of layers (#).
12. The method according to one of the preceding claims, wherein the provisioning (110) comprises: providing (126) an available upholstery material (Q) associated with the upholstery option of the at least one planned upholstery action (P) by means of an upholstery material provisioning device (400) of the upholstery system (1000); Placing (128) the provided cushioning material (Q) in the interior of the shipping container (900) according to the determined position using a handling device (500) of the cushioning system (1000).
13. The method according to any one of claims 1 to 12, wherein the provision (110) comprises: Providing (126) an available upholstery material (Q) associated with the upholstery option of the at least one planned upholstery action (P) by means of an upholstery material provisioning device (400) of the upholstery plant (1000); Providing (128) the majority of planned upholstery actions (P) via a screen (610) of the upholstery system (1000).
14. A non-volatile, computer-readable medium (600M) that stores instructions which, when executed by one or more processors (600P), cause the one or more processors (600P) to execute the method according to any of the preceding claims.
15. One upholstery plant comprising: at least one cushion provision device (400) for providing information (I) on a plurality of available cushioning options, wherein the available cushioning options differ from each other in at least one size property, in particular in at least one material, size or structure property; a sensor device (300) for capturing geometric data (G) of an interior (902) of a shipping container (900); a control unit (600) for generating a plurality of planned cushioning actions (P) based on the geometric data (G) of the interior (902) of the shipping container (900) and based on the provided information (I) on available cushioning options, wherein the at least one planned cushioning action (P) comprises at least one of the available cushioning options and a pose of at least one cushioning option in relation to the interior (902) of the shipping container (900); and a provisioning device (110) for providing the majority of planned padding actions (P) to position padding material (Q) according to the assigned padding material option and according to the assigned pose in the interior (902) of the shipping container (900).
16. The upholstery system according to the preceding claim, which is designed to carry out the method according to one of claims 2 to 14.
17. A procedure for training a function comprising: Providing (2002) a simulation environment (S) for simulating different states of an interior space (902) of a shipping container (900) divided into spatial elements, wherein according to an initial state of the interior space at least some of the assigned spatial elements are unoccupied; Providing (2004) information (I) about a plurality of available upholstery options, wherein the available upholstery options differ from each other in at least one material, size or structural property; Generating (2008) a plurality of cushioning actions (P) by an agent (A) based on at least one state of the simulated interior (902) of the shipping container (900), in particular based on the initial state, and based on the provided information (I) on cushioning options, wherein each planned plurality of cushioning actions (P) includes a cushioning option and a pose in relation to the spatial elements of the interior (902) of the shipping container (900); Applying (2010) the plurality of cushioning actions (P) to the simulated interior (902) of the shipping container (900) to generate a subsequent state (Z) of the simulated interior (902) of the shipping container (900), wherein the plurality of cushioning actions (P) comprises a respective arrangement of a simulated cushioning element in at least one or more spatial elements marked as free, according to the associated pose; Determining (2012) a cushioning quality (RW) as a function of the subsequent state (Z) of the simulated interior (902) of the shipping container (900) caused by at least one cushioning action (P); and Adapting (2014) at least one parameter of the function (F) to be trained, which controls the agent, using the determined cushioning quality (RW) and the associated planned cushioning actions (P), wherein the procedure at least comprehensively repeats the steps of generating, applying, determining and adapting to improve a decision behavior of the agent (A).
18. The method according to claim 17, wherein generating (2008) the plurality of planned padding actions (P) comprises: Identify (2016) a subset of the plurality of spatial elements which are part of a layer, wherein the layer optionally follows (a) a plane parallel to an access opening of the shipping container leading into the interior of the shipping container, or (b) a plane parallel to a floor of the interior of the shipping container, or (c) a plane parallel to a reference plane of the shipping container or cushioning system otherwise defined; and Generate (2018) at least a part of the majority of planned upholstery actions, in particular per shift, based on the identified subset of the majority of room elements and based on the information provided on upholstery options.
19. The method according to one of claims 17 to 18, wherein the function (F) to be trained is implemented as a policy function in the form of a deep neural network which maps a given state of the simulated interior (902) of the shipping container (900) directly to a probability distribution over cushioning options and poses, and wherein a policy gradient method is used to optimize this policy function.
20. The method according to any one of claims 17 to 19, wherein, in determining the cushioning quality (CQ), in addition to spatial utilization and / or the protection of the goods contained in the shipping container, further parameters are taken into account, in particular a cost factor of the cushioning material and / or a stability assessment.
21. The method according to any one of claims 17 to 20, wherein the cushioning quality is additionally determined depending on a regulatory parameter, which, for example, limits the amount of cushioning material used or ensures compliance with specified environmental or shipping guidelines by providing a corresponding penalty component or an additional reward component factor as part of the cushioning quality.