Article processing system and method using parameter exploration

Abstract

Disclosed is a control system for an article handling system having a programmable motion device, the control system comprising: a parameter estimator for estimating article handling parameters for at least some articles whose parameters are known; the parameter explorer is used for identifying the article carrying parameters of the known articles without parameters; and a parameter adjuster to determine whether to employ the parameter estimator or the parameter finder to adjust a parameter of the item handling system.

Description

[0001] priority

[0002] This application claims priority to U.S. Provisional Patent Application No. 63 / 358,310, filed on July 5, 2022, the disclosure of which is hereby incorporated by reference in its entirety. Background Art

[0003] The present invention relates generally to article handling systems and, in particular, to article handling systems such as automated storage and retrieval systems, distribution center systems, and sortation systems for handling a wide variety of articles using automated article handling equipment for handling a wide variety of articles.

[0004] Current item handling systems typically involve the handling of large quantities of items, where items are received in organized or unorganized batches and must be delivered to a desired destination according to a manifest or a specific address on the item (eg, in a mailing system).

[0005] For example, an automated storage and retrieval system (AS / RS) typically includes a computer control system for automatically storing (placing) items and retrieving items from defined storage locations. Conventional AS / RS typically employs totes (or boxes), which are the system's smallest loading unit. In these systems, the tote is brought to a person who picks individual items from the tote. When the person has picked the required number of items from the tote, the tote is redirected back to the AS / RS.

[0006] For example, current distribution center sortation systems typically present an inflexible sequence of operations in which a disorganized stream of incoming items is first pieced together into a stream of individual, isolated items that are presented one at a time to a scanner that identifies the items. A guiding element (e.g., a conveyor, tilt tray, or manually moved box) transports the items to a desired destination or further processing station, which may be a box, tilt shelf, chute, bag, or conveyor, etc.

[0007] In a typical package sorting system, human workers or automated systems retrieve packages, usually in the order they arrive, and sort each package or item into a collection bin based on a given set of heuristics. For example, all items of similar types, or all items in a single customer order, or all items shipped to the same consignment destination, etc., may go into one collection bin. Human workers or automated systems are required to receive the items and move each item to its designated collection bin. If the number of different types of input (received) items is large, a large number of collection bins are required.

[0008] Some of the most advanced sorting systems today rely to some degree on human labor. Most solutions rely on workers who perform sorting by scanning items from a guide area (chute, workbench, etc.) and placing them in a staging location, conveyor, or collection bin. When the bin is full or the control software system determines that the bin needs to be emptied, another worker empties the bin into a bag, box, or other container and sends the container to the next processing step. Such systems are limited in terms of throughput (i.e., the speed at which a human worker can sort or empty a bin in this manner) and number of transfers (i.e., for a given bin size, only so many bins can be arranged within the effective reach of a human worker).

[0009] Some automated handling systems rely on known processing parameters about the items to be gripped, for example, by a robotic system, and then moved to another location. These known processing parameters may include weight, size, and packaging materials. Since the purpose of such systems is to properly handle all items within a defined input set (e.g., without dropping items), the automated processing parameters are typically set to be conservative (to ensure adequate gripping of individual items and to set safe movement speeds). Likewise, however, such systems limit throughput and the number of transfers. When scaling provides more processing stations to increase throughput, the number of transfers may need to be expanded with the number of processing stations.

[0010] In addition to these challenges, there are also situations where some items may have information about the item incorrectly entered into the manifest or shipping label. For example, if the manifest in the distribution center includes an incorrect item size or weight (e.g., because the item size or weight was incorrectly entered manually), or if the shipper enters an incorrect size or weight on the shipping label, the processing system may reject the item as unknown. Additionally, and with respect to incorrect information on the shipping label, the shipper may be undercharged due to the incorrect information, for example, if the size or weight was incorrectly entered by the shipper.

[0011] There remains a need for a more efficient and cost-effective article handling system that handles articles of varying sizes and weights into appropriate collection bins or boxes, but is efficient in handling such articles of varying sizes and weights. Summary of the invention

[0012] According to one aspect, the present invention provides a control system for an article handling system having a programmable motion device. The control system includes: a parameter estimator for estimating article handling parameters of at least some articles whose parameters are known; a parameter explorer for identifying article handling parameters of articles whose parameters are not known; and a parameter adjuster for determining whether to use the parameter estimator or the parameter explorer to adjust the parameters of the article handling system.

[0013] According to another aspect, the present invention provides an item handling system comprising: a programmable motion device; and a control system, the control system comprising a parameter adjuster, which uses any of the following to adjust the parameters of the item handling system: a parameter estimator, which is used to estimate item handling parameters of at least some items whose parameters are known; and a parameter explorer, which is used to identify item handling parameters for adjustment through exploration, and to adjust at least one item handling parameter provided by the parameter estimator.

[0014] According to another aspect, the present invention provides a method for handling articles using a programmable motion device. The method includes: providing a parameter estimator, the parameter estimator being used to estimate article handling parameters of at least some articles for which the parameters are known; providing a parameter explorer, the parameter explorer being used to identify article handling parameters of articles for which no article handling parameters are known or for identifying article handling parameters for adjustment through exploration and for adjusting at least one article handling parameter provided by the parameter estimator; and employing a parameter adjuster to determine whether to use the parameter estimator or the parameter explorer to adjust the parameters of the article handling system. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] The following description may be further understood with reference to the accompanying drawings, in which:

[0016] Figure 1 An illustrative diagrammatic view showing an article handling system according to an aspect of the present invention;

[0017] Figure 2 Show Figure 1 An illustrative, schematic, enlarged side view of a portion of an article handling station in an article handling system of FIG.

[0018] Figure 3 Show Figure 2 A further enlarged view of an illustrative diagram of an end effector of an item handling station dropping an item;

[0019] Figure 4A and Figure 4B Show Figure 2 A portion of an item handling station that grasps a plurality of items ( Figure 4A ) and drops multiple items ( Figure 4B ) is an illustrative diagrammatic view of;

[0020] Figure 5A and Figure 5B Show Figure 2 Part of the item handling station has poor grip on items ( Figure 5A ) and dropping poorly grasped items ( Figure 5B ) is an illustrative diagrammatic view of;

[0021] Figure 6 A method for performing a Figure 1 Functional block diagram of the control system used in the system;

[0022] Fig. 7A and Figure 7B Show Figure 1 The end effector of the article handling station approaches the new vacuum cup ( Fig. 7A ) and attached to the new vacuum cup ( Figure 7B ) is an illustrative diagrammatic view of;

[0023] Fig. 8A and Figure 8B Indicates that the item has been dropped on the conveyor ( Fig. 8A ) and end effectors for clearing objects ( Figure 8B ) is an illustrative diagrammatic view of;

[0024] Fig. 9 shows an illustrative diagrammatic view of an article handling system according to further aspects including a removable access floor;

[0025] Fig.10 Show Fig. 9 An illustrative diagrammatic view of a removable access panel being pulled out;

[0026] Fig.11 shows an illustrative graphical representation of a probabilistic failure model of drops versus robot velocity for use in accordance with an aspect of the present invention;

[0027] Fig.12 shows an illustrative graphical representation of a fault model of shuttle timeout versus shuttle speed for use in accordance with an aspect of the present invention;

[0028] Fig.13 shows an illustrative graphical representation of a drop versus robot transfer speed and vacuum cup size fault model for use according to an aspect of the present invention;

[0029] Fig.14 shows an illustrative graphical representation of a fault model for multi-pick versus robot speed and vacuum cup size for use according to an aspect of the present invention;

[0030] Fig.15 An illustrative diagrammatic view showing a system architecture for an item handling system according to an aspect of the present invention;

[0031] Fig.16 An illustrative diagrammatic flow chart showing the logic flow of a parameter adjuster in a system according to an aspect of the present invention; and

[0032] FIG. 17A to FIG. 17D An illustrative diagrammatic flow chart of an optimizer procedure for use in an item handling system according to an aspect of the present invention is shown.

[0033] The drawings are shown for illustrative purposes only. DETAILED DESCRIPTION

[0034] Automated item handling systems typically include a programmable motion device (e.g., a robot) that receives incoming material and processes the material according to robotic handling parameters. Robotic material handling is subject to many choices (or parameters) when the system operates and / or performs actions. As the robot manipulates its environment (such as picking up items, such as SKUs (stock keeping units) for e-commerce or store delivery), there are outcomes that can be good or bad. The probability of a good outcome depends on the parameters selected and the item being handled. According to one aspect of the invention, the goal is for the system to learn good parameters or bad parameters for a given item from good outcomes and bad outcomes so that the system can optimize its performance over time in a feedback loop.

[0035] During automated item handling, data on the results of processing the item is recorded as being associated with the set of robotic material handling parameters. After the robot has completed the set of actions, data on the results is recorded in association with the material handling parameters used in the item handling. Machine learning tools may be used to develop one or more models that determine optimal values ​​for planning parameters. The one or more machine learning models may preferably be based on sequential hypothesis testing of variable sample sizes, rather than iterative optimization methods. This is accomplished by a parameter estimator that estimates optimal processing parameters based on known performance data. For example, known performance data may be provided in an initial evaluation.

[0036] However, the performance of the parameter estimator is limited by the action set initially evaluated. There may be an optimal parameter, but if the parameter estimator is not fed with performance data around this operating parameter, the parameter estimator will never propose the optimal parameter because the optimal parameter is beyond the scope of its experience or training set. According to various aspects, the system and method of the present invention provide a generator of initial selection in the absence of a large amount of information about what is a good initial selection. Potential optimal parameters are proposed before any such parameters may have been executed, and are therefore outside the training set. The function of the parameter explorer is to propose parameters outside the training set. The decision to extract parameters from the parameter estimator or the parameter explorer must be made automatically and based on various information states. The parameter adjuster makes these determinations through a feedback process and decides whether new information needs to be obtained by using a parameter explorer, or whether there is enough certainty in the parameter to adopt a parameter estimator. At the end of each action, the robot applies the detection result (e.g., successful or some kind of failure such as falling, multiple selection or wrong selection) and stores this information in a database. The action result is sent to the parameter adjuster, and the feedback loop is therefore closed.

[0037] The parameter adjuster may acquire new information in order to achieve the following short-term and long-term information acquisition goals. With respect to the short-term information acquisition goal, the parameter adjuster may determine that there is an insufficient amount of information about, for example, the picking performance of a particular SKU, and may therefore determine that exploration is needed. Information about the size or weight of the SKU may not be available with certainty, and any feasible parameter matches may not be initially known. Therefore, the information must be acquired by examining multiple parameters proposed by the parameter explorer with the aid of a picking action.

[0038] With respect to long-term information acquisition goals, a robotic unit may operate normally across a wide range of parameters and SKUs, but the goal of operation may be to improve performance over time by exploring new parameters that may improve the performance of its system. The parameter adjuster may employ the parameter explorer in those cases where it has decided to increase the speed gradually and conservatively. It may do so only in contexts where the recent performance history is otherwise reliable and acceptable, so that the risk of error can be acceptably absorbed. Conversely, if the robotic unit has recently made many errors, the parameter adjuster may decide to only adopt more conservatively selected parameters from the parameter estimator in order to meet the service level agreement. Operational constraints that may exist include safety, reliability, and speed.

[0039] Thus, according to various aspects, the present invention provides systems and methods for optimizing handling parameters of a robotic material handling system. The systems and methods govern the choice of whether the next task is to rely on a machine learning based performance model or to explore the parameter space to find good parameters. Exploration may be selected, for example, in the absence of information about the properties of an item. The decision to use exploration may be made based on the history of previous handling of a given item or based on other factors such as the history of operational performance. Exploration is selected when there is a lack of history of handled items and / or a lack of information about the item's properties, which, if known, would better inform parameter selection. The system may perform exploration in a hybrid parameter space using mixed continuous and discrete parameters.

[0040] The system and method may also perform exploration when the properties (e.g., weight and size) of the item that affect its handling are unknown, and the system and method estimate both these handling properties and good properties of the item during its exploration. In some applications, the system and method determine whether the result is not to carry the item at all. The system can optimize performance metrics (e.g., throughput) over time while keeping other metrics such as uptime within a service level agreement (SLA). The system may start to carry items that may not be in a well-functioning state according to some requirements, but implement strategies to either become well-functioning (meeting thresholds) when completing tasks with the item, or determine that the item cannot be carried in a well-functioning manner and should therefore be marked as not eligible for automation. The system collects data and performs sequential hypothesis testing on the data set (e.g., using Bayesian optimization). The optimization process operates simultaneously with the operation of one or more robotic systems and provides data throughout the grasping operation, moving (trajectory) operation, and placing operation. The system accelerates learning by collecting item properties and good parameters learned by multiple operating systems or multiple sites; the system then shares the learned parameters across multiple units and sites. The system is applicable to multiple robotic material handling tasks.

[0041] The systems and methods of various aspects of the present invention may be used with a variety of item handling systems, such as robotic articulated arm systems and shuttle systems that handle multiple items in a warehouse, including discrete items and boxes of discrete items. Such item handling systems may include, for example, sorting, distribution, store replenishment, and fulfillment of e-commerce orders at an order fulfillment center. For store replenishment, a distribution center holds an inventory of goods that need to be individually distributed to multiple stores. In e-commerce order fulfillment, items are removed from the item inventory and placed into shipping boxes or bags for delivery to consumers by the shipper. In each of these operations, the robot is handling a single item or SKU.

[0042] Figure 1For example, an item handling system 10 is shown in which the systems and methods of various aspects of the present invention may be employed. There are a wide variety of handling tasks to which the systems of the present invention may be applicable, some permutations of which are presented herein. A brief overview includes: multiple tasks performed with task-specific parameters that can be controlled by the robot and that affect performance; multiple potential or unknown properties of the items being handled that can affect performance; failures, including multiple physical / real-world problems that can go wrong in each of the tasks such as dropped items, and these failures are captured by sensors and algorithms that process sensor data to infer physical problems; and multiple goals, ways of measuring performance, such as speed or uptime or a combination of multiple metrics that the system is configured to automatically optimize through a feedback process, provided that the system can meet the SLA.

[0043] like Figure 1 As shown, the article handling system 10 includes an input conveyor system 12; an article handling station 14; and a distribution system 16, which includes, for example, a plurality of shuttle wing systems 18, 20, 22. The input conveyor system 12 includes an input conveyor 24 on which articles (e.g., discrete articles, articles in open totes or boxes, or containers containing articles for processing) are presented to the article handling station 14. The article handling station 14 includes, for example, a programmable motion device 26, such as an articulated arm of a robotic system having an end effector 28, the end effector for grasping articles for transport to any of a plurality of destination locations 30 of the shuttle wing systems 18, 20, 22. The robotic system can be suspended from a support structure 32, which includes a plurality of sensing systems 34 for monitoring the input of articles and the grasping and delivery of articles by the end effector 28 including a vacuum cup 29. Figure 2 As further shown. The operation of the system can be controlled by one or more computer processing systems 100, 101 as further discussed herein. A high flow vacuum can be provided to the end effector 28 from a high flow vacuum source 103, for example, providing a flow rate of at least about 100 cubic feet per minute (e.g., 130-140 cubic feet per minute), and the vacuum pressure provided by the high flow vacuum source can be no more than about 25,000 Pascals below atmospheric pressure in some examples and no more than about 50,000 Pascals below atmospheric pressure in other examples.

[0044] The system can be designed to adapt to failures in item handling, such as poor grip of the end effector, grip of multiple items, altered (failed) grip of items, and dropped items, as the system explores a variety of different item handling parameters. Figure 2An enlarged view of the article handling station 14 is shown (with portions of the support structure 32 removed for clarity) showing articles being moved to one of a plurality of spaced-apart brackets 36 of the shuttle wing systems 18, 20, 22. The input conveyor 24 includes rollers 38 mounted on torque sensors or load cells 40 for monitoring the weight of each box or tote as the articles are removed and for detecting whether an article has fallen onto the rollers 38. Additionally, the system includes a sensing system 42 for detecting whether an article has fallen onto the rollers 38. As shown in FIG. Figure 3 As further shown, a removable access panel 44 (also shown as Fig. 9 and Fig.10 ) to pick up any items that have fallen (accidentally or intentionally) from the end effector (e.g. Figure 3 48 as shown), and a sensing unit 46 is provided in the removable access bottom plate to monitor / confirm that each item is received in the access bottom plate 44.

[0045] Different parameters may be selected and adjusted by the automated system for gripping items, transporting items and placing / packaging items, as well as marking items and box dumping. For example, gripping may involve picking items from a conveyor, a tote or a box with a gripper for transport to another location in the robot's working environment. The gripper may be vacuum-based or may be based on some other type of gripping system. Adjustable parameters for gripping tasks may include the speed of the robot arm during picking and movement, the selection of end effectors or vacuum cups, vacuum pressure and vacuum flow (where such selection is available). For example, vacuum cup changes may be provided based on the probability of falling. Adjustable parameters may also include parameters that affect the perception process, which uses depth and image sensor data to generate gripping candidate locations. Adjustable parameters may also include thresholds and other parameters that affect the behavior during the initial movement of the item, such as a pressure threshold, below which it is considered to be a loss of gripping force. Adjustable parameters may also include controllable parameters of the motion planning system other than speed, such as those that may be changed depending on whether the item tends to sway or is deformable. Adjustable parameters may also include parameters for a grasp selection strategy, such as whether to select one of many grasp strategies, or whether to supply as input additional parameters for interpreting image and depth based data to determine grasp points.

[0046] Adjustable parameters for the transport task may include transporting the items from the location where they are received using a shuttle system to the point where they are ejected by tipping over the hopper or by conveying in one direction or the other (because the carrier is a cross-belt conveyor). Such parameters may include the speed of transport in one or more of its axes (1 axis: for example, horizontal only; 2 axis: horizontal and vertical), and the speed of ejection by tipping over or conveying, depending on the carrier ejection mechanism.

[0047] Adjustable parameters associated with the packing task may include packing the items picked by the robotic arm into a shipping box or other container such as a tote. Such parameters may include margins (spacing) to be left around the items to be placed or items that may already be within the container, as well as adjustable parameters of the motion planning system other than speed, such as those that may affect how the items are controllably reoriented for packing.

[0048] Adjustable parameters associated with the identification task may include identifying items for sorting by presenting them to a barcode or RFID scanner in successively different orientations, or orienting and releasing the items to a drop scanner or RFID scanner with a barcode, or using visual or geometric features of the items to identify the items. Such parameters may include parameters that affect the selection of an orientation for presentation or release.

[0049] Adjustable parameters associated with the case dumping task may include manipulating a tote from a cardboard case into a tote using a gripper or other electromechanical system dumping unit, such as for storage in an AS / RS. Such parameters may include parameters of the motion of the case as it dumps, including the amplitude or frequency of the vibratory motion.

[0050] The above is applicable to many different automated material handling tasks where there is complexity in repetitive decision making and actions. As described herein, the parameter set for the selected task is denoted as B.

[0051] The attributes of an item may include a variety of characteristics that affect material handling but may not be immediately apparent, including, for example, weight, size, geometric modeling, internal weight distribution and inertia, characteristics of packaging such as cardboard, bagged, plastic, cardboard with one or more plastic windows, and combinations thereof. Such characteristics may also include mechanical properties, such as whether the item is rigid, whether it is easily deformed (bagged clothing), whether it contains liquid that is pushed, whether it can be rolled, whether it is porous, and the manufacturer and general category of the item (such as small electronic devices, food, toys, etc.). Warehouse management systems (WMS) and database systems may or may not provide these attributes and characteristics to the robotic system. For example, a warehouse database may or may not contain weights and dimensions, and if the warehouse database contains weights and dimensions, these attributes may be incorrect, they may have small but far-reaching errors, or item types may have attributes that vary a lot between different instances of the item type.

[0052] All of the potential attributes listed above can be discovered during the handling of an item with varying degrees of certainty. Weight can be inferred from a scale; dimensions and geometry can be inferred from a 3D scanner; internal weight can be inferred during handling; packaging characteristics can be sensed visually; manufacturing information can be inferred from OCR of text on the item; and category can be sensed using machine vision.

[0053] The impact of a bad choice of parameters depends on the task. The occurrence of failures is a signal that the system uses to train the system to avoid the given parameters. The system must be able to detect these failures in order to automatically perform a feedback process that ultimately improves performance and finds good parameters that are more trouble-free than alternative parameters.

[0054] According to one aspect, the systems and methods of the present invention provide for a variety of faults that occur within a controlled environment that can detect the faults while not being adversely affected by faults associated with the handling of other items. Examples of faults and mechanisms that a robotic system can use to detect faults are presented herein. For example, with respect to grasping tasks, the faults detected may include: 1) dropped items as detected by loss of vacuum as detected by a pressure sensor or failure of a conservation of mass test as detected by a scale, where one or more scale measurements at one or more stations are fused to infer the possibility of an incomplete transfer, 2) multiple picks - or picking more than one item when only one item was intended to be picked (detected by using a scale and detecting a difference greater than the expected eight, which implies that more than one item was lifted from the picking area, 3) torque faults reported by the robotic system (e.g., due to collisions in the robotic environment as reported by the robotic system), 4) conveyor failures as detected by a conveyor subsystem, and 5) failures in the conveyor system to move items. A conveyor jam fault, where the conveyor subsystem reports inconsistent tracking, which implies a conveyor jam, 5) a source or destination container overfill fault, as detected by a container scanner, which infers that the item is over or overhanging beyond the maximum height based on the depth image, 6) a motion planning failure fault, as detected by an application, which reports that it is unable to find a plan for the robotic system to achieve the operation goal, 7) a failure to see the item, as detected by a grip selection system reporting that the grip was not found, despite a WMS requesting an order from the container; and 8) a failure where excessive force is applied to the item, which may cause damage, as detected by a scale reporting an excessive load during a pick or place operation. Figure 4A For example, the system 10 is shown where the vacuum cup 29' of the end effector 28 grasps a multi-pick as shown at 49. When the system detects a multi-pick, the system can release two items 49 into the removable access plate 44, as shown in FIG. Figure 4B shown. Figure 5AFor example, the system 10 is shown in which the vacuum cup 29 of the end effector 28 has a poor grip on the item 52. When the system detects a poor grip, the system can release the item 52 into the removable access base 44, such as Figure 5B shown.

[0055] Figure 6 A functional block diagram of one or more computer control systems 100 for controlling the operation of the control system is shown. Specifically, the one or more systems 100 include a central processor system 200 that communicates with a vacuum controller 202 that controls, for example, vacuum pressure and vacuum flow, and a robotic system controller 204. The robotic control system 204 controls the vacuum cup determination system, the articulated arm movement speed system, the grasping planning system, the motion planning system, the weighing duration system, and the reorientation controller. The central processor 200 is also coupled to a storage database 206 and a parameter regulator 208. The parameter regulator 208 controls inputs from a parameter estimator 210 and a parameter explorer 212 that are also coupled to the central processor 200. The central processor 200 is also coupled to an identification task controller 214 and a case dumping task controller 216 and a transport task controller 218. The central controller is further coupled to a shuttle transport controller 220 that controls the transport speed and ejection speed and a packaging task controller 222 that controls the item-to-wall margin and the item-to-item margin.

[0056] The system can change the vacuum cup at the cup changing station 50, such as Fig. 7A and Figure 7B The cup changing station 50 provides a plurality of types of vacuum cups of different sizes on a frame 56 as shown at 58. One or more magnets 60 on the end effector 28 engage the cups from a vertical direction (e.g. Fig. 7A ), and can be moved in the horizontal direction away from the frame 56 together with the engaged cup 54 (as shown Figure 7B Placing the vacuum cup back onto the gantry 56 involves moving the end effector in the opposite direction (first horizontally, and then vertically, to disengage the cup from the one or more magnets 60 on the end effector 28).

[0057] With respect to transport faults, detected faults may include: 1) item loss faults such as detected by loss of vacuum as detected by a pressure sensor or failure of a mass conservation test on a scale, 2) jamming or failure to move along an axis as desired due to an item being blocked from moving after or during transport, and 3) timeout waiting faults while waiting for a shuttle to be available to receive an item.

[0058] With respect to packaging failures, detected failures may include: 1) dropped items as detected by loss of vacuum as detected by a pressure sensor or failure of a scale's conservation of mass check, 2) multiple picking - or picking multiple items when only one item was intended to be picked (as detected by a failure of a scale's conservation of mass check, where the scale detects that more than one item has been lifted from the picking area, 3) torque failures reported by the robotic system (e.g., due to collisions in the robotic environment as reported by the robotic system), 4) destination container overfill failures as detected by a container scanner where the container scanner infers that the item exceeds or overhangs beyond a maximum height based on a depth image, and 5) failures due to a container being expected to be overfilled after placement and therefore unable to be packaged as detected by a container scanner where the container scanner infers that the item exceeds or overhangs beyond a maximum height as predicted prior to placement.

[0059] With respect to identification failures, detected failures may include an inability to scan an item, as detected by failure to successfully scan a barcode or failure to identify from geometric or visual features. With respect to dumping failures, detected failures may include a case contents jam, as detected by failure to dispense the contents of the case, such as due to the case being packed too tightly, as detected by scanning the case contents or load sensing, and a tote overfull failure, as detected by scanning the tote contents.

[0060] Some faults may impede the operation of the robot in such a way that a person must approach the robot unit and intervene, for example to clean up a dropped item. This generates an intervention request, where the system visually and / or audibly indicates on the UI or via an API that the system needs to be serviced. Other faults may not impede the operation of the robot, but instead may attach to the WMS via an API that a given container needs to be serviced. These kinds of faults, called anomalies, do not stop the robot unit because they postpone the handling of the anomaly by encapsulating it in a tote or container that leaves the robot unit and travels away from the robot unit.

[0061] For example, according to one aspect, an item may fall into the removable access floor 44, which may be removed from beneath the programmable motion device so that the item therein may be collected and reintroduced into the input system or manually processed. According to other aspects, when the system determines that an item 60 has fallen onto the input conveyor 26 (e.g., Fig. 8A ), the programmable motion device may be instructed to grasp (or sufficiently grasp) the item so that the programmable motion device can bring the item over the bottom plate receiving basin to drop into the bottom plate receiving basin for processing as detailed above (such as Figure 8B). If the item cannot be so gripped for transfer to the floor receiving basin, the system will visually and / or audibly signal that human intervention is required, as discussed above.

[0062] Fig. 9 The removable access base plate 44 is shown being removed with any items that have been accidentally dropped or intentionally ejected from the end effector, and Fig.10 Sensing units 26 are shown located on opposite sides of the removable access panel 44 for monitoring and / or confirming that each item expected to be received within the access panel 44 has been received within the access panel 44 .

[0063] During any one instance of a mission, there are yet additional potential or unmodeled environmental conditions that may contribute to failures, such as how the items are arranged in a tote or how the items are packed in a box. Modeling or detecting these potential conditions may be difficult. Averaging these unmodeled potential conditions, the propensity for any of these failures will depend on (i) the controlled parameters; and (ii) the properties of the items being handled. This can be described as the probability of failure as a function of the controlled parameters and the modeled properties.

[0064] Finally, there may be events that occur in the unit that should be classified as faults, but are not detected. The system can be designed to minimize these false negatives by providing the system with a set of appropriate mechanisms to detect such events. For example, if the robotic unit has not detected the event, there is no feedback loop and therefore these events cannot be automatically minimized over time. In addition, the system may not detect some types of damage to the item, in which case the system will not be able to reduce those types of damage over time.

[0065] Operational performance objectives and probabilistic failure models can be used to evaluate parameter sets and item combinations over time. The parameters and physical properties of the items being handled will together determine the performance of the task on the item. Some parameter and item combinations may produce poor performance or a high rate of a certain set of failures. Conversely, some parameter and item combinations may produce low failures and good performance.

[0066] The operational performance of the robot at a given task is measured by a set of key performance indicators (KPIs) that measure the average performance of all tasks that the robot has performed. These are post hoc inference metrics averaged over a certain time window or number of tasks performed. Examples of KPIs include: Throughput: number of tasks completed per hour; Intervention Rate: number of interventions per hour; Exception Rate: number of exceptions per hour and other combinations or measures of related failure rates.

[0067] These indicators are metrics against which services are measured in a customer / provider relationship in order to achieve operational performance objectives. A Service Level Agreement (SLA) is a contractual agreement on operational performance, typically expressed as minimum / maximum target levels for KPIs. Failure to reach the target levels and therefore failure to achieve the SLA may result in penalties for the robotic cell supplier; conversely, exceeding the target levels may be rewarded. Some elements of these KPIs may not depend on parameters or items, but may be caused by other uncontrolled / unmodeled factors. The function B(B) can generally be written as a relevant measure representing past transmissions encoded in a sequence of task results T. Such measures are assumed to have finite memory, windowed by time or number of task results.

[0068] Note that there may be trade-offs between elements of performance: slow transport may result in fewer drops but may reduce overall throughput; fast transport may increase throughput but may increase drops. The system may express a metric as a combination of others in order to balance two or more metrics: B1(B)=BB2(B)+BB3(B). The system may also specify constraints such as B4(B)<B to achieve the SLA for a given KPI. The goal of the system of the present invention is to learn the parameters for each item that enable the robot to achieve its SLA, which the designer may express as some combination of objectives (e.g., maximum B1(B)) under an optional constraint (e.g., B2(B) does not exceed B2). The system addresses these issues by optimizing surrogates for the KPI.

[0069] A set of surrogates can be based on a probabilistic failure model (PFM). For any failure type B, there is a probability that a failure event will occur given a task and item B and parameters B∈B, where B is also a parameter space. Consider the relation B B，B，B = 1, meaning the fault event in question has occurred, or B B，B，B = 0 means that the fault event has not occurred. The probability fault model B of the fault type is:

[0070] B(B B,B，B )=B(B B,B (B)

[0071] Where B(·) represents the Bernoulli distribution, which takes the value of 0 or 1. That is, B(B B，B，B ) is a discrete Bernoulli distribution with a probability failure model function B indexed by failure type B and item B B，B (B) and a function of the parameters B. For any task and fault type, this PFM exists, but will be unknown. The system can only estimate it based on past performance. The estimate is used to predict the performance of the next task performed on the item. Specifically, the probabilistic model can be used to predict changes in the KPI as a function of parameter selection.

[0072] KPIs measure operational performance and have limited memory by design. They measure how the system performed over the past hour or day and will vary based on the distribution of items encountered. The system uses a probabilistic failure model to select the best parameters based on its predictions. The PFM should be the best estimate of the true PFM and will use as much actual data as possible.

[0073] Fig.11 The probabilistic failure model of drops versus robot transport speed for a fixed SKU is shown. Speeds above 3 units (e.g., feet / second) begin to result in increasingly higher drop rates. Fig.12 A failure model showing the probability of a shuttle timeout or item loss during a shuttle versus shuttle speed. Running too slow will result in too many timeouts, and running too fast will result in too many items being lost. Fig.13 The probabilistic failure model of SKU dropping versus robot transfer speed and vacuum cup size is shown. Larger vacuum cups have better stability and can hold items at higher speeds. Fig.14 The probabilistic failure model for multi-pick of SKUs versus robot speed and vacuum cup size is shown. Multi-pick, or when the vacuum cup picks more than one item, depends only on the vacuum cup and not the conveyor speed. However, larger vacuum cups are more likely to result in multipacks than small vacuum cups.

[0074] refer to Fig.15 , the architecture for the system includes a parameter adjuster 300 in communication with a parameter exploration module 302 and a parameter estimator 304 . Fig.15 Summarizing the architecture of the parameter explorer, parameter estimator, and parameter generator. The parameter exploration module 302 includes a model update module 308 in communication with a simplification application 316, and an optimizer 310. The parameter estimator 304 includes a scoring module 312 in communication with the parameter adjuster 300, and a retrainer 314 in communication with the database 306.

[0075] Machine learning is preferably based on sequential hypothesis testing methods (rather than iterative or gradient descent optimization methods) with variable sample sizes and more manageable parameter spaces.

[0076] The Task Application (Task App) performs the material handling tasks and plays a supervisory role in coordinating many subsystems (such as motion planners, perception systems, device drivers, etc.), all in order to perform the task at hand. The application implements detection mechanisms for failures; interfaces with the WMS; and otherwise controls the robot to repeatedly complete the task. At the end of each transfer attempt, the Task App stores the results of the executed task (success or failure and any other relevant data) in a database.

[0077] The Task App interfaces with mechanisms that help it detect faults marked in the graphical fault detection mechanism, as well as devices and software for estimating the properties of the item property estimation mechanism. Each can represent a variety of mechanisms that provide corresponding detection or estimation functions in various aspects. The parameter adjuster serves as an interface between the parameter estimator, parameter explorer, and the Task App. The parameter adjuster receives product data such as SKU attributes from the Task App and decides whether the parameters from the parameter explorer or from the parameter estimator will be used to transport the item. The parameter adjuster can decide not to transport the item, effectively deeming the item incompatible or ineligible for automation.

[0078] The Parameter Explorer consists of two parts: the Item Model Updater and the Optimizer. At the end of each transfer attempt, the Task App calls the Item Model Updater. The Item Model Updater receives the last transfer result, retrieves the Item Parameter Model (IPM) from the database corresponding to the given item, and updates the model with the last transferred result and stores it back to the database. The Optimizer receives the product data from the Parameter Adjuster, retrieves the exploration model from the database product, selects the next parameters and sends them back to the Parameter Adjuster. If the product is picked for the first time, the database does not have its model and the Optimizer will return empty parameters.

[0079] The parameter estimator consists of three parts: the software modules scorer and retrainer, and the data product which is a global parameter model (GPM). The scorer receives the product data from the parameter adjuster, runs the product through the GPM and returns the parameters selected by the model. Since the GPM is a version-controlled entity that changes over time, the scorer is decoupled from the GPM. The retrainer retrieves the new transfer results from the last training, combines them with the old transfer results, and retrains the model, resulting in a new GPM. The retrainer runs periodically, for example, once a day.

[0080] The retrainer can be local or remote from the system. The above architecture can be run individually to control one robot to perform a task in one warehouse. In some cases, the retrainer can operate outside the warehouse and can collect and combine transfer results from one or more warehouses and robot units performing the same handling task. The retrainer derives a global parameter model that can undergo quality assurance or regression testing, and then uploads the new global parameter model to the local database of the robot unit.

[0081] Parameter exploration learns item properties and good item parameters. During exploration, two things are learned: (1) the properties of the item; and (2) good parameters for the item. Information about (1) may not be available, or may be degraded in quality or accuracy. If the global parameter model predicts the same parameters that have been learned through exploration, it is because the item properties learned through exploration are able to predict the same good parameters. There may be a mismatch, which may indicate that the global parameter model has poor predictive ability to predict good parameters for a given item. There are several reasons for the mismatch between the parameters discovered by parameter exploration and the parameters from the global parameter model (GPM).

[0082] A GPM may be immature in the sense that there is little training data or support in the area of ​​the learned attribute of an item. For example, poor predictions for 1 kg items would be expected when all prior task results are for items <500 g. Another possibility is that multiple parameters may be equally good. For example, a PFM may be flat for a range of parameters. In this instance, changes in the parameters are not important. Many parameters may be equally good. Thus, training a GPM may produce one good parameter, while a small perturbation of the attribute or task result may produce another good parameter when training an IPM.

[0083] Finally, a third possibility is that there are some latent or hidden item attributes that are not supplied by the WMS and cannot be estimated from the sensor data, but if those attributes were provided, different good parameters would be induced. In machine learning terms, regression learning is inseparable given the available features, but will be separable if some additional features become available. For example, two items may have the same weight and size, but one item may be clothing in a plastic bag and the other item may be an electronic product in a box. If the input to the GPM is only weight and size, the GPM will output the same parameters for each item. However, bagged items may require slower movement to achieve the same failure rate as, for example, boxed items.

[0084] Regardless, the GPM may provide parameters that are non-performing for an item due to lack of support or inseparability. If the IPM is to continue updating for the item through exploration, better parameters will be learned and there will be an IPM / GPM divergence. If the divergence is due to lack of support, the GPM should be updated with new training data. However, if the divergence is due to inseparability, the system should continue to use the IPM because the GPM will never exploit features that distinguish the items.

[0085] The characteristics of the global and item attribute models may also vary. Both the parameter estimator and the parameter explorer generate parameters based on predictions from the machine learning model, the global parameter model, and the item parameter model, respectively. GPM is a model that takes the attributes of an item as input and outputs predictions of good parameters for the item. GPM cannot be used to drive the exploration process. Using the GPM model requires that the item attributes are known, but many warehouses do not maintain a database of item attributes. If the product attributes are known and the system obtains a first pick and product attribute estimate, if the transmission causes a failure, the parameters may not be good. The response may be to use this information to retrain the model. There are two problems with doing this: (1) It may be computationally expensive to retrain the model with all the prior transmission results and this new information; and (2) there may be a lot of inertia or training support around a given input parameter, and a given good parameter may be valid for a large number of other items. GPM may have to perform a large number of failed transmissions before it can generate new parameters for new items.

[0086] As discussed above, the quality of the parameters generated by the model depends heavily on the training set. For example, if the training set only has product delivery at slow speeds, the model will also choose slow speeds to deliver future products. When exploring new items, the parameter explorer may not make assumptions about what good parameters to use for new items. The parameter explorer is learning both item attributes and good parameters. Therefore, the parameter explorer employs an IPM model for each item, which is isolated from the GPM.

[0087] The following table summarizes the differences between GPM and IPM:

[0088]

[0089] With this design, if an item has unknown properties, the system is able to predict and learn the parameters while making minimal assumptions about the item.

[0090] Therefore, the system can provide a robust estimate of the object's properties. Fig.15 , the item attribute estimation mechanism may involve several layers (although shown as boxes), including a sensor driver, a potential independent software module for inference, and potential multiple processing layers, inference layers, and integration layers in the task app.

[0091] Two item attributes are highly predictive of performance on handling tasks: weight and size. The system measures weight using a scale that can verify the number of items picked when the mass of the item is known. When the mass of the item is unknown, the system cannot tell from a single measurement whether the system has grasped 1 item or more than 1 item. Therefore, item weight cannot be inferred from a single pick. However, over the course of multiple picks, it can be inferred from weight history by discretizing the weight to a hopper that is larger than the noise level and using the smallest mode as the weight estimate. Size can be measured by a 3D scanner while the item is being grasped or after the item is set down.

[0092] The parameter adjuster design is an agent that determines whether the task app will use parameters from the parameter estimator or the parameter explorer to move the item. For example, the parameter estimator is used when there is enough information about the item to use a global model similar to regression. In one example, one or more weights and one or more dimensions can be input, and the system outputs parameters such as vacuum, cup size, and speed. For example, the parameter explorer is used when there is not enough information about the characteristics of the item to use regression. For example, there are no weights or dimensions to enter. The parameter explorer can also be used in cases where the performance of the parameters output from the regression clearly leads to insufficient performance, and therefore other parameters must be sought through exploration. Fig.16 The flowchart in shows the logical flow of the parameter adjuster. Fig.16 The system is shown first determining whether the item has known attributes (step 500), if so, the system obtains one or more parameters from the parameter estimator (step 502), and if not, the system forwards the parameters from the parameter explorer to the task app (step 504). After the system obtains the parameters from the parameter estimator (step 502), the system determines whether the parameters have been used less than K times to move the item (step 506). If yes, the system forwards the parameters from the parameter estimator to the task app (step 508). If no, the system determines whether there were many bad events in the last L attempts to move the item (step 510). If yes, the system forwards the parameters from the parameter explorer to the task app (step 512). If no, the system forwards the parameters from the parameter explorer to the task app (step 514).

[0093] Therefore, the system provides that GPM requires item attributes as input, so if the item attributes are unknown, the system cannot use GPM, so exploration must be used instead. If the attributes of the item are known, the large amount of experience and knowledge history built into GPM is used, so the system should use GPM to generate parameters. If the item attributes have been estimated, but it has been found through experience that the performance from GPM is poor, the system should use exploration. If the poor performance is due to lack of support, the retrainer will start collecting data and support in this situation, and over time, the GPM parameter proposals will converge to the IPM proposals over time. If the poor performance is due to inseparability, the GPM output and the IPM output will always be divergent, and the IPM will continue to be selected.

[0094] The parameter explorer design involves multiple operating strategies. There are several different ways to implement the parameter explorer, including optimizers and item model updaters. The system includes two approaches that have been attempted in the context of a picking task. The measured failures that are sought to be reduced are drops and multiple picks, which depend on the choice of suction cups on the gripper, the speed of the robot arm, and optionally the vacuum pressure. In short, the approach described in detail below is based on a fixed heuristic strategy: a manually created set of rules for setting parameter values. For example, if the SKU weight is less than 0.2 kg, a small cup is used, otherwise a medium cup is used. Such a rule set can be created using domain expertise. An ML-based approach that periodically retrains the model has two advantages over manually created rule sets: (1) it can find new correlations between item characteristics and the delivered results that are not captured by the rules; (2) it can customize parameters for packaging types, product categories, and brands that are unknown in advance and will appear in the warehouse in the future.

[0095] In a picking task, the system may encounter many different items (aka SKUs). As mentioned before, there will be a different item parameter model for each item. While picking, the robot may encounter items in a scattered manner (not all at once). Therefore, it is implicitly understood in the following description that the parameter explorer may operate on one instance of a certain item A, and then the task app may switch to another mode (e.g., estimate) for the next item B it encounters. Once it encounters A again, it may switch back to exploration for that item, depending on the decision of the parameter adjuster.

[0096] The method provides a fixed heuristic strategy, which involves an optimizer program with a heuristic strategy implemented as an if-then rule set. For example, and with reference to FIG. 17A to FIG. 17D, the program starts (step 1000), if the product is currently selected using a small cup (step 1002) and there are more than M1 drops in the last N single picks (step 1004), the next pick will use a medium cup (step 1006). If the product is currently selected using a medium cup (step 1008) and there are more than M2 drops in the last N single picks (step 1010), the next pick will use a large cup (step 1012). If the product is currently selected using a medium cup (step 1008) and there are more than M3 multiple picks in the last N picks and there are no single picks with drops in these N picks (step 1014), the next pick will use a small cup (step 1016). If the product is currently selected using a large cup (step 1018) and there are more than M4 multiple picks in the last N picks and there are no single picks with drops in these N picks (step 1020), the next pick will use a medium cup (step 1022). If there are no drops in the last N single picks (step 1024), the weighing duration of the product is reduced by 0.1 (step 1026). If there are more than M5 drops in the last N single picks using the small cup (step 1028), and there are more than M6 multiple picks in the last N picks using the medium cup (step 1030), the small cup is used and the weighing duration is increased by 0.3 (step 1032). If there are more than M7 drops in the last N single picks using the medium cup (step 1034), and there are more than M8 multiple picks in the last N picks using the large cup (step 1036), the medium cup is used and the weighing duration is increased by 0.3 (step 1038). The routine can end (step 1040) or repeat.

[0097] The above if-then rule set represents the strategy implemented in the optimizer. The corresponding item parameter model encodes the following: (i) statistics of parameter choices made in the past; and (ii) the current state or parameter choices. The item model updater updates the given parameter choices output from the above optimizer and updates the statistics as a result of the transfer.

[0098] Note that each of the above steps can be shortened to an acceptance-rejection test so that not all N choices need to be evaluated. For example, if N=20, but two drops have occurred in the first 3 choices for a fixed parameter choice, then the parameter choice can be rejected immediately. This speeds up the exploration process by reducing the number of passes evaluated before changing to the next candidate.

[0099] Fixed strategies as described above have several limitations: (1) they depend on constants M and N, whose optimal values ​​may be difficult to find, (2) they have fixed thresholds for switching between parameters and do not support logic for switching back and forth between different parameters. For example, suppose M=2, N=20, and there were 3 drops in the last 20 picks using a small cup, then there were 3 drops in 3 picks using both a medium cup and a large cup. The if-else rule does not specify which cup to choose. So in this case, the robot will need to use a heuristic to choose the next cup, and this heuristic may fail, and (3) when generalized to many different failures, a large number of manually determined rules will be required to cover the different combinations of bad events and the required parameter changes. This last point motivates the more general approach described below that does not require manually tuned strategies.

[0100] According to one aspect of the invention, an exploration process based on Bayesian optimization can be used. Bayesian optimization is a machine learning approach for optimizing unknown functions. In this approach, parameters are selected by mathematical models rather than by manually tuned heuristics, allowing us to overcome the limitations of using if-then rules for parameter exploration.

[0101] Bayesian optimization recursively updates the estimate of one or more functions being optimized. Here the function is PFM, B given an item. B，B (B), which was defined previously. This algorithm and other recursive estimation algorithms (such as the Kalman filter) balance past information with new information. The operating state in these approaches includes the mean and (co)variance. Low variance in the operating state implies confidence and resists high measurement noise. Conversely, high variance in the operating state implies low confidence and tells the algorithm to attach great importance to new evidence. When a continuous function is being estimated and only a portion of the function is observed, the Gaussian process acts as a function of the mean and covariance.

[0102] A B-dimensional Gaussian variable B~B(B,B) has a probability distribution with mean B and covariance matrix B. It describes a Gaussian-shaped probability block in a multidimensional space centered around the mean and shaped according to the covariance matrix.

[0103] The Gaussian process generalizes the concept of multivariate Gaussian distribution to infinite-dimensional space: instead of defining a finite-dimensional vector B with B∈{1,...,B} B The Gaussian process defines the distribution of infinite-dimensional functions B(B), where B is a real number in some domain B. If B has a GP distribution, we write B~BB(B, B), where B is the mean function and B is the covariance kernel. The expected value of B(B) is B(B), and the covariance kernel is the expected product of the deviations from the mean:

[0104] B(B,B′)=B[B(B)] and B(B,B′)=B[(B(B)-B(B))(B(B′)-B(B′))].

[0105] For any finite subset of real numbers B1, ..., B B , vector [B(B1), B(B2), ..., B(B B )] is a multivariate Gaussian distribution.

[0106] In practice, the main difference between a multivariate Gaussian and a GP is the starting or prior covariance. In contrast, in a Kalman filter (with a multivariate Gaussian), the prior covariance is usually a diagonal matrix: At the beginning, B B With B B On the other hand, in GP, ​​the core is the use of non-zero off-diagonal covariance terms due to the prior belief in the continuity of B. If B is continuous, then as B tends to zero, the covariance of B(B) and B(B+B) should converge to the variance of B(B): BBB B→0 B(B, B+B)=B(B, B), and the correlation is 1 in a small neighborhood of B.

[0107] This makes inference practical when only a portion of the function is observable, since observations of B are typically only at one sample point B at a time. When making inferences, the prior belief in continuity and therefore the covariance between B(B) and B(B+B) means that B(B)-B(B+B) should be small, and therefore the post-measurement difference in means B(B)-B(B+B) should also be small. Alternatively, if the off-diagonal of the covariance is zero, then observations of B(B) will not give any information about B(B+B), and learning B will take too long.

[0108] Fundamentally, the covariance prior is used to enforce continuity in the sampled function, which reduces the number of observations required to get a good estimate of the true underlying function.

[0109] Those skilled in the art will appreciate that many modifications and variations can be made to the above disclosed embodiments without departing from the spirit and scope of the invention.

Claims

1. A control system for an article handling system having a programmable motion device, the control system comprising: a parameter estimator for estimating article handling parameters of articles for which at least some parameters are known; a parameter explorer for identifying article handling parameters for articles for which no parameters are known; and A parameter adjuster is used to determine whether to use the parameter estimator or the parameter explorer to adjust the parameters of the item processing system.

2. The control system of claim 1, wherein the adjustable parameters include any one of robot speed, vacuum cup selection, vacuum pressure, and vacuum flow.

3. A control system as claimed in any one of claims 1 to 2, wherein the adjustable parameters include a perception process for generating grasp candidate positions.

4. A control system as claimed in any one of claims 1 to 3, wherein the adjustable parameters include thresholds that affect the behaviour of an item during transport.

5. The control system according to any one of claims 1 to 4, wherein the adjustable parameters include the moving speed of the shuttle.

6. A control system as claimed in any one of claims 1 to 5, wherein the adjustable parameters include margins used during packaging.

7. A control system as claimed in any one of claims 1 to 6, wherein the adjustable parameter includes attitude orientation.

8. A control system as claimed in any one of claims 1 to 7, wherein the adjustable parameters include tank dump parameters.

9. A control system as claimed in any one of claims 1 to 8, wherein the adjustable parameters are selected by the control system using optimization-based modeling involving sequential hypothesis testing.

10. The control system of claim 9, wherein the adjustable parameters are selected by Bayesian optimization that recursively estimates the function being optimized.

11. An article processing system, comprising: Programmable motion device; as well as A control system comprising a parameter adjuster that adjusts a parameter of the article processing system using any of: a parameter estimator for estimating article handling parameters of articles for which at least some parameters are known; as well as A parameter explorer is configured to identify, by exploration, an item handling parameter for adjustment and to adjust at least one item handling parameter provided by the parameter estimator.

12. The article handling system of claim 11, wherein the article handling system includes a plurality of sensing systems for monitoring for faults during article handling.

13. The article handling system of claim 12, wherein the failure comprises any one of a picking task failure, a transport failure, a packaging failure, an identification failure, and a case dumping failure.

14. An article handling system as claimed in any one of claims 11 to 13, wherein the adjustable parameters include any of robot speed, vacuum cup selection, vacuum pressure and vacuum flow.

15. An article handling system as claimed in any one of claims 11 to 14, wherein the adjustable parameters include a perception process for generating grasping candidate locations.

16. An article handling system as claimed in any one of claims 11 to 15, wherein the adjustable parameters include thresholds that affect the behavior of articles during transportation.

17. An article handling system as claimed in any one of claims 11 to 16, wherein the adjustable parameters include the moving speed of the shuttle.

18. An article handling system as claimed in any one of claims 11 to 17, wherein the adjustable parameters include margins used when packaging.

19. An article handling system as claimed in any one of claims 11 to 18, wherein the adjustable parameters include posture orientation.

20. An article handling system as claimed in any one of claims 11 to 19, wherein the adjustable parameters include bin dumping parameters.

21. An article handling system as claimed in any one of claims 11 to 20, wherein the adjustable parameters are selected by the control system using optimization-based modeling involving sequential hypothesis testing.

22. The article processing system of claim 21, wherein the adjustable parameters are selected by Bayesian optimization that recursively estimates the function being optimized.

23. A method for processing an article using a programmable motion device, the method comprising: providing a parameter estimator for estimating article handling parameters of articles for which at least some parameters are known; providing a parameter explorer for identifying item handling parameters for items for which no item handling parameters are known or for identifying item handling parameters for adjustment through exploration, and for adjusting at least one item handling parameter provided by the parameter estimator; as well as A parameter adjuster is employed to determine whether to use the parameter estimator or the parameter explorer to adjust parameters of the item handling system.

24. The method of claim 23, wherein the method further comprises: Failures during item handling are monitored, including any of picking task failures, transport failures, packaging failures, labeling failures, and case dumping failures.

25. The method of any one of claims 23 to 24, wherein the article handling parameters for adjustment are selected using optimization-based modeling involving sequential hypothesis testing.

26. The method of claim 25, wherein the item handling parameters for adjustment are selected by Bayesian optimization involving recursively estimating a function being optimized.

Images