Patent Application
20220134543
The invention generally relates to programmable motion processing systems, and relates in particular to programmable motion (e.g., robotic) systems intended for use in environments requiring, for example, that a variety of objects (e.g., articles, parcels or packages) be processed (e.g., sorted and/or otherwise distributed) to several output destinations.
Many object distribution systems receive objects in an organized or disorganized stream that may be provided as individual objects or objects aggregated in groups such as in bags, arriving on any of several different conveyances, commonly a conveyor, a truck, a pallet, a Gaylord, or a bin. Each object must then be distributed to the correct destination location, as determined by identification information associated with the object, which may be determined by a label printed on the object. The destination location may take many forms, such as a bag or a bin or a tote.
Automated systems for order fulfillment have also been proposed. A challenge however, is that any automated system must be capable in many applications of handling a wide variety of objects of varying size, weight, volume, center of mass and shape, as well as objects and/or packaging having low pose authority (such as clothing packaged in plastic bags), or low position authority (such as round or cylindrical objects that may roll or otherwise move after being placed).
Further, many object distribution systems receive objects (e.g., SKU items, parcels, packages, etc.) from one or more shipping entities, for distribution to a wide variety of destinations. Such object distribution systems must therefore sometimes accommodate induction of objects for which little or no information is available. Such unknown objects may be processed by hand, but in an automated object processing systems, such processing by human personnel may be disruptive or otherwise impractical.
There remains a need for a more efficient and more cost effective object induction system for an object distribution system that distributes objects of a variety of sizes and weights into appropriate collection bins or trays of fixed sizes, yet is efficient in handling objects of such varying sizes, weights, compositions, and identities.
In accordance with an aspect, the invention provides an object induction system for assigning handling parameters to an object. The object induction system includes an analysis system, an association system, and an assignment system. The analysis system includes at least one characteristic perception system for providing perception data regarding an object to be processed. The characteristic perception data includes data relating to any of a weight, height width, length, weight, center of mass, object description, object category and at least one image. The association system includes an object information database with characteristic recorded data including data relating to any of a weight, height width, length, weight, center of mass, object description, object category and at least one image for a plurality of objects. The association system is for assigning association data to the object responsive to commonality with of any of the characteristic perception data with any of the characteristic recorded data. The assignment system is for assigning programmable motion device handling parameters to the indicia perception data based on the association data. The assignment system includes a workflow management system as well as a separate operational controller. The programmable motion device handling parameters include any of a vacuum pressure at an end effector, a size of a vacuum cup at the end effector, a maximum speed of movement of the end effector, a maximum angular acceleration of the end effector, a maximum linear acceleration of the end effector, a location at which to engage the object with the end effector, and a pose at which to hold the object by the end effector when grasped.
In accordance with another aspect, the invention provides an object induction system for an object processing system including at least one programmable motion device. The object induction system includes an analysis system, a handling parameter system, and an assignment system. The analysis system includes at least one characteristic perception system for providing perception data regarding an object to be processed. The characteristic perception data includes data relating to any of a weight, height width, length, weight, center of mass, object description, object category and at least one image. The handling parameter input system is for obtaining handling parameters input data regarding the object. The handling parameter data includes data relating to any of whether the object is able to roll following placement, whether the object is fragile, whether the object is stackable, whether the object is crushable, whether the object is deformable, whether the object is too thin to be processed, whether the object includes glass and whether the object is non-rigid. The assignment system is for assigning programmable motion device handling parameters to the indicia perception data. The assignment system includes a workflow management system as well as a separate operational controller. The programmable motion device handling parameters include any of a vacuum pressure at an end effector, a size of a vacuum cup at the end effector, a maximum speed of movement of the end effector, a maximum angular acceleration of the end effector, a maximum linear acceleration of the end effector, a location at which to engage the object with the end effector, and a pose at which to hold the object by the end effector when grasped.
In accordance with a further aspect, the invention provides an object induction system for an object processing system including at least one programmable motion device. The object induction system includes an indicia perception system, an analysis system, a handling parameter input system, a non-transitory storage medium, and an assignment system. The indicia perception system is for providing indicia perception data regarding identifying indicia of an object to be processed. The analysis system includes at least one characteristic perception system for providing perception data regarding the object to be processed. The characteristic perception data includes data relating to any of a weight, height width, length, weight, center of mass, object description, object category and at least one image. The handling parameter input system is for obtaining handling parameters input data regarding the object. The handling parameter data including data relates to any of whether the object is able to roll following placement, whether the object is fragile, whether the object is stackable, whether the object is crushable, whether the object is deformable, whether the object is too thin to be processed, whether the object includes glass and whether the object is non-rigid. The non-transitory medium is for storing the characteristic perception data and the handling parameter input data as associated with the indicia perception data. The assignment system is for assigning programmable motion device handling parameters to the indicia perception data. The programmable motion device handling parameters include any of a vacuum pressure at an end effector, a size of a vacuum cup at the end effector, a maximum speed of movement of the end effector, a maximum angular acceleration of the end effector, a maximum linear acceleration of the end effector, a location at which to engage the object with the end effector, and a pose at which to hold the object by the end effector when grasped.
The following description may be further understood with reference to the accompanying drawings in which:
The drawings are shown for illustrative purposes only.
In accordance with certain aspects, the invention provides a system and process for inducting objects into an automated object processing system. For example, personnel at warehouse receives a SKU that it had not previously processed. The human personnel scans the barcode to identify the object. Then the person scans the object with a machine that measures the height, width, length and weight of the object (such as with a Cubiscan scanner sold by Quantronix, Inc. of Farmington, Utah). The personnel can then add additional information about the SKU (such as whether it's a fragile item) that might aid in either the classification of the SKU, or the determination of handling parameters. That additional information may be entered in a GUI or by scanning a sheet showing pictures and barcodes (as discussed below). Then the personnel decants the case of that new SKU into a tote, and because the whole process was associated with the SKU barcode and the tote barcode, the database now has information about the tote contents.
In the background, the weight and dimension data, as well as the additional information provided by the personnel, is used to compute at a separate server whether the SKU is eligible for automation, and what the handling parameters are (e.g., optionally involving machine learning comes in to extrapolate on a new SKU). These parameters are then used to route the SKU to manual picking cells, or specific automated picking cells.
Current object processing warehouses have existing equipment and processes entirely built around the manual processes that have been in use for decades. While some automation systems force the warehouse to adapt around the new automated equipment (e.g., add lines on the floor for robots to use to drive around, use X scanners, put items at Y height . . . etc.), systems of various aspects of the present invention provide equipment that adapt to existing warehouses.
When adding robotic picking systems into existing warehouses there need to be systems in place to prepare items for robotic picking. Some warehouses receive truckloads of sealed cardboard boxes and process whatever is inside entirely manually (e.g., a person will move the box, cut it open, and place individual items into new boxes). In these cases, the robot is not a direct replacement for a person. Other process optimizations need to be made to guarantee efficiency and functionality of the robots. Tasks that generally need to be done include the following. First, objects eligible for robotic picking need to be chosen. Some items will be too heavy, too large, too small, too light, etc., while others will be packaged improperly. Eligibility can sometimes be known by using item data, but other times, the object must be visually inspected as received from vendors. Second, the objects must then be placed in optimal containers to be conveyed and presented to the robot for robotic picking. Robots cannot pick objects out of just any container. Objects tightly packed into a cardboard box, for example, may be difficult to pick using a robot.
Systems disclosed herein enumerate ways to deal with these two tasks (and other tasks that could be considered to be value add for the warehouse and/or for robotic picking). Many of the disclosed ideas combine automation and manual processes, using a balance to capture the strengths of each. The outcome is a net reduction in labor with an increased system efficiency. A goal is to increase throughput rate.
A SKU is a stock keeping unit and represents a kind of item. Each kind and size of shampoo represents a different SKU. An object (e.g., unit or an each) is one of something. A case is usually a cardboard box containing multiple units of the same SKU. Quality assurance (QA) typically refers to a part of a retailer's organization tasked with ensuring that the right goods get to the right places without damage. The warehouse management system (WMS) is the computer system that communicates with the robot cells. The WMS tells the robot what orders to fulfill, e.g., “pick one item out of tote with barcode Z, and put it into an empty container, then indicate what its barcode is.” A robot control system (RCS) centralizes information about SKUs, and conveys to all robots operating at a site what to do when it receives a SKU—new or previously encountered. An automated storage and retrieval system (AS/RS) is an automated system that stores and retrieves totes; in the context of warehouses for store replenishment or e-commerce, these totes contain one SKU, or are subdivided into compartments, each of which contain one SKU.
There are two important parameters: automation eligibility and handling parameters. Automation eligibility involves using a flag that indicates whether to send a given SKU to a robot or not (or to a type of robot if there are multiple different types of robots). The handling parameters involve (where an object is automation eligible), what parameters such as robot speed, suction cup, or vacuum pressure and any other parameters to be employed while handling the SKU. These two parameters need to be set for all SKUs that the robot cells could encounter. Warehouses often have 100s of 1,000s of SKUs, so it is often intractable to set these parameters manually.
In accordance with various aspects, systems and processes are provided for determining the eligibility for automation of SKUs by performing the evaluation on-cell. In one method, the warehouse management system (WMS) assumes that all new SKUs are automation eligible and can be routed to a robot. The robot then attempts to pick all SKUs and detects from (a) drops; (b) mis-picks; (c) plural-picks; (d) inability to pick the tote to empty; or (e) any other measure that indicates performance was out of norm, that it did a bad job picking a particular SKU. In this instance the SKU or the tote containing the SKU could be flagged and sent automatically to QA (building QA, decant station, or other location) to make an automation eligibility decision manually.
There can be an optimization or scheduling of when to do these evaluation picks described above. There is a risk of wasting time doing the first evaluation picks, because the robot systems may mis-pick, drop, etc. These new SKUs could be evaluated at a time in the day where the systems are more idle in accordance with an aspect. These new SKUs could be evaluated while weighing the costs (potential for error and manual intervention at the cell) and benefits (expanding the number of SKUs that can be automatically picked) while hitting operational measures (meeting daily throughput targets), and dependent on measures like how often a given object is likely to be picked based on information about SKU velocity (i.e., how often ordered).
When in this mode of evaluating a new SKU, the system may also operate more conservatively than with previously encountered SKUs. The robot may move more slowly, or otherwise change parameters to, for example, take a little bit longer. The system may also perform exploratory motions, e.g., jiggling the item, without actually doing a full transfer motion, so as to further evaluate handle-ability of the SKU. During these evaluation picks, the cell might notify human personnel or operators that it is in an evaluation mode, so that they may more quickly come to the rescue if needed.
In accordance with another aspect, system and methods involve using customer return data or other customer feedback (good or bad). Such systems and methods may use customer return information to inform handling parameters or automation eligibility. In this case, the WMS sends categorical or qualitative information from the receiving customer or retail store to the RCS such as: happy with item as received, item was damaged, or received wrong number of items. This information could be associated with a specific SKU, or with a specific customer order in the case of e-commerce, or with a case/box of items sent to a store. When the exact unit can be determined, then all the sensor data and handling parameters for that specific unit can be associated with other objects, and used as inputs to machine learning algorithms that better choose handling parameters to avoid damage, or generally improve customer satisfaction.
In accordance with a further aspect, systems and methods are provided that use a pick and decant test station for training of human personnel and robots. In this instance, instead of assuming all new SKUs are automation eligible, new SKUs are examined by a separate automated system. This system may, for example, automatically pick the items repeatedly to test their automation eligibility. This would be a test cell that is not in the usual flow of goods through the warehouse. The operation may be bin-to-bin, which is to fill a left-hand tote with the SKU, and transfer all the units to a right-hand tote. Once the left-hand tote is empty, transfer those items back to the right-hand tote; repeat many times to determine compatibility and other pick parameters on new SKUs. Errors at this cell are much less expensive, and don't impact operations. Alternatively, this may be a semi-automated or fully manual process. A manual gripper test-bench may be employed to try picks (equivalent of hose w/suction cup on it), which would evaluate whether the system could grip the item. At this system, a human personnel could also train the robot how to grasp items. The test bench can be used for human personnel to determine decant decisions (best face up), including for example, other directions for decant (max cases per tote, etc.), and whether the objects could be processed without the test bench.
The weight sensing conveyor system includes a conveyor section 22 that is mounted on rollers 24, 26, each of which is mounted at both ends on a pair of load cells 28, 30 (only one of which is shown at one end of each roller 24, 26) as shown in
With reference to
Such systems therefore, that provide weight sensing in the presentation conveyor may include one or more load cells or weight sensitive mechanisms embedded into the surface on which objects are presented to a programmable motion device such as an articulated arm. Each object's weight and/or observed density (weight/volume) as may be estimated using the programmable motion system's cameras or range sensors that can perceive volume.
The response evaluation section 16 of
Further, if the roller(s) 52 are rotating at a faster rotational velocity, the system may determine an inertial value for the object 58 as the roller(s) engage and discharge the object from the roller(s). A wide variety of further characteristics may also be determined or estimated, such as for example, center of mass (COM) using the roller(s) in combination with the perception unit(s) as discussed herein and further below. The system may further use the perception units and roller(s) 52 (together with computer processing system 100) to determine whether the object is a collapsible bag, and/or whether the presumed object 58 is actually a plural-pick (includes more than one object), again, using the perception unit(s) in combination with the roller(s) by observing whether the objects move apart and/or whether the shape of the object changes as it rides over the roller(s) 52. In accordance with further aspects of the invention, the transport rollers 50 may be replaced by conveyor sections that stand below the height of the perturbation rollers 52.
In accordance with a further aspect, systems and methods are provided that generate informative handling parameters in semi-automated fashion. In this instance, a human personnel inspects the SKU, takes measurements such as recovering weight and product dimensions, then also provides categorical information about the SKU, and then a software system receives all this information, and then determines automation eligibility and handling parameters (using regression, machine learning, or other algorithms). The human personnel scans the item barcode so as to get the SKU, and then puts it in a weighing and dimensioning system (e.g., again a Cubiscan scanner as noted above). Then the human personnel tags all relevant categorical information about the SKU, characteristics such as: fragile, brittle or easily damaged (which might imply about how aggressive to move or push down on item), in a plastic bag (which might imply how quickly to move), rolls (which might roll when placed), not stackable (which might mean don't put this on the bottom of a pile), is glass (so don't handle by robot, or handle very gently), is very thin (so either don't handle if not compatible, or use special type of gripper), and opens when picked up (like a book or jeans, and so shouldn't be handled by a robot).
The separate operational controller 70 may then send routing instructions 76 as well as handling parameter information for the object to any of a variety of programmable motion devices 90, 92, 94 (as well as data indicating that the object is ineligible for automated handling 96). The system may also periodically flag for quality control previously enrolled SKUs to be re-weighed and analyzed for their dimensions as shown at 98.
The end-effector may be used with a programmable motion device in an object processing system.
The object processing station 122 includes an infeed conveyor section 124′ that circulates selected supply bins 126′ from and back to the infeed conveyor 124 using the diverter bi-directional conveyors 136. The end-effector of the programmable motion device 132 is programmed to grasp an object from a supply bin 126′, and move the object to deliver it to a desired destination bin 130′ on the destination conveyor load area 128′ by placing or dropping the object into a destination container 130′ on the destination conveyor 128′ at the destination conveyor load area. The supply bin 126′ may then be returned to the input conveyor 124 and, optionally, brought to a further processing station. At the processing station 122 therefore, one or more vendor supply bins 126′ are routed to an input area, and the programmable motion device 132 is actuated to grasp an object from a bin 126′, and to place the object into a selected destination container 130′. The processed vendor bins 126′ are then returned to the common input stream on the conveyor 124, and the destination container 126′ is moved further along the destination conveyor 124.
The system 120 may also include one or more perception units 138 located on or near the infeed conveyor for identifying indicia on an exterior of each of the bins, providing perception data from which the contents of the bin may be identified, and then knowing its relative position on the conveyor 124, track its location. It is assumed, in accordance with an aspect, that the bins of objects are marked in one or more places on their exterior with a visually distinctive mark such as a barcode (e.g., providing a UPC code), QR code, or radio-frequency identification (RFID) tag or mailing label so that they may be sufficiently identified with a scanner for processing. The type of marking depends on the type of scanning system used, but may include 1D or 2D code symbologies. Multiple symbologies or labeling approaches may be employed. The types of scanners employed are assumed to be compatible with the marking approach. The marking, e.g. by barcode, RFID tag, mailing label or other means, encodes a identifying indicia (e.g., a symbol string), which is typically a string of letters and/or numbers. The symbol string uniquely associates the vendor bin with a specific set of homogenous objects. Based on the identified code on an infeed bin 124, the system may either permit a bin 124 to continue along the infeed conveyor 124, or may direct the selected bin 126′ onto the selected infeed conveyor 124′.
On the selected infeed conveyor 124′ at the object processing station 122, the perception system 134 assists (using the central control system 200—e.g., one or more computer processing systems) the programmable motion device 132 including the end-effector in locating and grasping an object in the infeed bin 126′. In accordance with further aspects, each object may also be marked with a visually distinctive mark, again such as a barcode (e.g., providing a UPC code), QR code, or radio-frequency identification (RFID) tag or mailing label so that they may be sufficiently identified with a scanner for processing. The type of marking depends on the type of scanning system used, but may include 1D or 2D code symbologies. Again, multiple symbologies or labeling approaches may be employed on each object.
The system 120 further includes a site intake perception system 140 that includes a plurality of perception units 142, 144, 146148 that are directed downward onto the one or more objects in each infeed bin 126 on the infeed conveyor 124, as well as a weight sensing section 139 of the conveyor 124 under the perception system. Further, the weight sensing section 139 may further include a vibratory device 137 for shaking the bin in order to cause objects within the bin to spread apart from one another within the bin as discussed in more detail below. The perception system is mounted above the conveyor into each bin of objects to be processed next looking down into each bin 126. The perception units, for example, may include, a camera, a depth sensor and lights. A combination of 2D and 3D (depth) data is acquired. The depth sensor may provide depth information that may be used together with the camera image data to determine depth information regarding the various objects in view. The lights may be used to remove shadows and to facilitate the identification of edges of objects, and may be all on during use, or may be illuminated in accordance with a desired sequence to assist in object identification. The system uses this imagery and a variety of algorithms to generate a set of candidate grasp locations for the objects in the bin as discussed in more detail below.
The perception system 140 includes among the perception units 142-148 scanning and receiving units as well as edge detection units for capturing a variety of characteristics of a selected object of the whole bin. Again,
In accordance with further aspects, the scanning and receiving units may also be employed to determine a density of the collection of objects in the bin, which is compared with a known density of the identified SKU multiplied by the known number of objects in the bin from knowing the object's mass and volume. The volumetric data may be obtained for example, using any of light detection and ranging (LIDAR) scanners, pulsed time of flight cameras, continuous wave time of flight cameras, structured light cameras, or passive stereo cameras.
In accordance with further aspects, the system may additionally employ edge detection sensors that are employed (again together with the processing system 200), to detect edges of any objects in a bin, for example using data regarding any of intensity, shadow detection, or echo detection etc., and may be employed for example, to determine any of size, shape and/or contours as shown in
Again, the operations of the system described above are coordinated with a central control system 200 that again communicates (e.g., wirelessly) with the articulated arm 132, the perception systems 134, 138, 140, as well as in-feed conveyors 124, 124′, bi-directional conveyors 136, destination conveyors 128, 128′ and any diverters. This system determines from symbol strings the UPC associated with a vendor bin, as well as the outbound destination for each object. The central control system 200 is comprised of one or more workstations or central processing units (CPUs). For example, the correspondence between UPCs or mailing labels, and outbound destinations is maintained by a central control system in a database called a manifest. The central control system maintains the manifest by communicating with a warehouse management system (WMS). The manifest provides the outbound destination for each in-bound object.
With reference to
With further reference to
The scanning system 162 is positioned above a detector 200, and an object 202 (such as a portion, e.g., edge or corner, of a shipping bag) may be positioned over the detector 200. When the scanning system 162 is positioned adjacent the object 202, the x-ray scanner is engaged, and heater power is supplied to the cathode electron emitter 204. A high voltage (e.g., 30-50 kV) is thereby applied between the cathode end 188 and the anode end 186. The electric field produced by the applied high voltage accelerates electrons from the electron emitter through the vacuum to the x-ray producing target 194. The intensity of the x-rays produced at the target increases with increasing high voltage, electron beam current, and atomic weight of the target material. A portion of the x-rays produced in the target exit the tube via the x-ray transmission window 196, and exit the housing 182 via the x-ray output region 184 of the housing 182. The high voltage at the cathode end is typically provided as a negative high voltage (e.g., −50 kV) and the voltage potential at the anode end is typically provided at a reference ground potential of the system. This permits the anode end 186 of the tube 180 to be coupled directly to the housing 182. The x-ray tube 180 may be packaged in a reciprocating device that includes a high voltage power supply and a power source to drive the electron emitter.
Depending on the power level adjustments, the scanning system 162 may be used to determine any of materials and density of both the shipping bag and/or any contents. For example, the system 162 may be used in certain embodiments to distinguish between low density polyethylene (0.917-0930 g/cm3) and high density polyethylene (0.944-0.065 g/cm3). Such a system may also be used to determine whether a density of an object is too great for an articulated arm to be able to safely lift or move the object, or in other aspects to determine comparative responses to a variety of known materials for machine learning purposes.
The scanning system 164 is shown in
With reference to
The system may then return to a text/image/bag analysis routine (step 1004) in
Further, and with reference to
Additionally, and with reference to
The system may then (and with reference to
The system may then employ x-ray analysis of edges (step 1086), x-ray scattering (step 1088), and apply x-ray detection analysis (step 1090) to characterize any scattering or transmission detection to further characterize the object. The system may also (and with reference to
In accordance with further aspects, the invention provides strategies for decanting by robot, automated system or human personnel, in order to improve picking performance. Decanting is the operation of transferring units that arrive in a case, to a tote; the tote then gets forwarded to one or more robot cells directly, or is the storage unit for an automated storage and retrieval system (AS/RS). The performance of the robot cell can depend on the organization of objects inside the tote. If the objects are neatly tiled and adjacent to one another, then it can be difficult for grasp detection algorithms to generate a pick that picks only one item. For example, if the system captures a picture of tote from above, then chooses where to draw circle (for where the suction cup goes), then the chances that the circle contains two objects may be high (and thereby result in a plural-pick). In accordance with various aspects therefore, the robotic or automated decanting station, or station for improving pickability might, for example, shake the tote to reduce the effects of tiling, and to for example, spread out the objects in the tote to help the robot make better grasp decisions. In accordance with various further aspects, human personnel may select some grasp points on an image to help train the robot as disclosed, for example, in U.S. Pat. No. 10,625,432, the disclosure of which is hereby incorporated by reference in its entirety.
In accordance with further aspects, multiple special induction systems for SKU tuning may be provided similar to the SKU induction system 10 of
In accordance with further aspects, the system may provide sub-optimal picking performance or characteristics about the SKU that don't match information in the database, and automatically tags the SKU or tote to be routed to QA stations. This may be done where the operational controller 70 detects suspicious data and asks human personnel to weigh or measure an object (SKU) either using the system disclosed above or a Cubiscan scanner noted above that is in communication with the station for use. In accordance with further aspects, human personnel may categorize a menu of packaging types, or other items used in compatibility such as employing a crush factor, which is a measure of susceptibility of an object being crushed.
In accordance with further aspects, the system may evaluate tote features that would improve picking performance. For example, the system may note tote color or texture, e.g. so that the bottom of the tote has a known texture or background. The color (e.g., of a background or inner surface) of a tote may be used for identification purposes such as classifying a tote, or may provide a known background against which objects may be more readily identified. Further, a textured inner surface (e.g., wavy) may provide a non-flat bottom to aid in picking, and/or may include a shape (e.g., inserts) that keep items away from the edges, providing a more bowl shape again to aid in picking. In accordance with further aspects, the system may provide totes with holes and push blades that can push objects within the tote to aid in picking.
In accordance with further aspects, the system may provide feedback for informing a supply chain of better packaging options. Some of this information could be used to provide an automated what-if recommendation for vendor feedback. Human personnel would take a photograph of an example of bad packaging to generate report to be sent back to vendors. Also, the system may generate grasps from the photo to let the associate see what grasps would look like.
Parameter estimation may also be conducted throughout the induction processing, for example, during training and scoring modes. The scoring mode uses the current handling parameter estimation models to generate planning parameters for a single SKU or a set of SKUs. An estimator application will be called by product manager upon receiving new or updated SKU data.
The training mode will have the following functionality: training of parameter estimation models, offline evaluation of newly training model (see below), comparison of offline performance of the new model with the one of production model and decision if the new model is promoted to production, and generation of planning parameters for all SKUs in Product collection in data storage.
The data storage will store a separate test set of experiments with the known numbers of drops, plural-picks and damages. In this set each test SKU will have experiments with multiple cup sizes and multiple scale durations. This test set will be a subset of damage and 24/7 experiments. A model (production one or newly trained one) will choose for each test SKU the best parameters among those that have experiments in the test set. Then a training application will count the number of bad events generated by the chosen parameters.
For each SKU in the Product collection, the data storage will store two sets of planning parameters, one for induction and single-pick cells, and the other one for plural-pick cells. A large database will also store two serialized models, one for induction and single-pick cells, and the other one for plural-pick cells. The training will run both on-demand and periodically, for example after uploading a new training data to data storage, human personnel will be able to launch a training process (both data upload and launch of training will be done through training application console). Additionally, the training may be registered as a scheduled job and will run for example, once a week.
The datastore 278 also communicates with a parameter estimator application 284, and in particular provides, for example, pick results and results of damage experiments to a training module 290, which provides parameter estimation models and results of offline evaluation back to the datastore 278. The datastore 278 also provides parameter estimation models to a scoring module 292 of the parameter estimator application 278, and the scoring module 292 is in communication with the product manager 286, providing planning parameters for each type of cell, and receiving from the product manager 286 product information. The product manager 286 also receives notifications for generating parameters for existing products, as well as new product information and update information for existing products from a workflow management system 288.
Closed-loop training is provided in accordance with a further aspect employing a final production model that is used for scoring, and does not have access to drop and plural-pick data and further predicts bad event score from SKU features only. This plural-pick and drop data is referred to herein as privileged features. This data can be used during training but is not available during scoring. The following table summarizes four types of data that will be used in training and scoring modes.
Dataset A is obtained from dataset C by removing privileged features. Parameter estimation application will use self-training approach for training production model of
The analysis module 370 provides data 376 that removes the privileged features to provide dataset A (labeled training data) 378, which is in communication with a further training module 380. The output of the training module 380 is provided to a candidate prediction model 382, and the candidate prediction model 382 is in communication with a further scoring module 384, which receives dataset B (unlabeled scoring data) 386, and provides the SKU parameters 388.
Those skilled in the art will appreciate that numerous modifications and variations may be made to the above disclosed embodiments without departing from the spirit and scope of the present invention.
The present application claims priority to U.S. Provisional Patent Application No. 63/107,680 filed Oct. 30, 2020, the disclosure of which is hereby incorporated by reference in its entirety.
| Number | Date | Country | |
|---|---|---|---|
| 63107680 | Oct 2020 | US |
