Despite a challenging year for global logistics in 2020, investments in warehouse automation technologies witnessed a whopping 57% increase, hitting $381 million in the first quarter. Warehousing automation solutions drove deal activity in logistics as investors made huge bets on emerging robotics technologies for warehousing and supply chain.
Investors’ growing interest in startups dealing with warehouse automation coincides with the surge in ecommerce orders and changing consumer buying behavior. Smaller batch (single item) orders have become more popular post-pandemic and consumers now expect better visibility into the order fulfillment process.
It is no surprise then that warehouse automation is fast becoming a competitive advantage not only for ecommerce companies but all businesses reliant on efficient logistics.
This article will explore what warehouse automation is, what it entails, and what machine learning based software tools can push you forward in the world of warehouse automation.
Warehouse automation is a collection of technologies that enable organization logistics to orchestrate the movement of goods into, within, and out of warehouses to customers with minimal human intervention. Warehouses are ideal candidates for automation because they have multiple repeatable, process-oriented, and human error-prone operations. Think pallet stacking, order picking, receiving, doing inventory, etc. Automation helps warehouse processes become more consistent, accurate, and overall a profitable business unit.
The COVID-19 pandemic and its resulting impact on supply chain and logistics have served as a catalyst to warehouse transformation. With companies having to operate with decreased labor force, automation served a critical purpose. Conveyor systems and other robotic process automation technologies kept warehouses running even without significant human involvement.
Larger macroeconomic factors like rising domestic and international wages led companies to realize the ROI on one-time automation costs. Recent research suggests that industrial robot prices might drop by as much as 60% by 2025, making automation even more cost-effective.
The commoditization of enabling technologies like machine learning is also driving increased adoption of warehouse automation across key markets. LogisticsIQ suggests that the global warehouse automation market could double in size by 2025. Advanced robotic solutions could account for up to 41% of the $27 billion in annual spend. Warehouse automation can broadly be categorized into two components:
Process automation focuses on taking manual processes and varying workflows and adding automation components that reduce workload and manual hours.
In addition to process and physical automation, warehouse automation systems can also be categorized as basic augmentation (low level of automation like conveyor belts), software-based automation (process automation), RPA, and advanced automation (where a significant part of day-to-day operations have been automated).
You don’t need huge Amazon level fleets of autonomous mobile robots or massive automated conveyor belts to see the benefits of automation through machine learning based software. These tools can provide a ton of value your warehouse management system and pipelines you’re already using, or can become an entirely new piece of the backend puzzle.
Processing incoming invoices, vendor documents, and product labels requires manual effort and task understanding to take the information from these documents and input them correctly into backend systems. Most warehouses process a ton of these documents per day which requires workers to take time away from more important and time restrained tasks.
Automated document processing software leverages a number of machine learning and deep learning techniques to extract the valuable information from these documents and automatically update backend systems and databases. These models can be trained to handle a huge variety of document formats, sizes, and qualities of picture which allows for a higher accuracy in the real world. The pipeline setup can be as quick as snapping a picture of the document or label on a phone, and the software extracts the information and updates the backend systems in less than 3 seconds.
Here’s a look at the different technologies that go into a standard warehouse document automation system.
Deep learning based OCR algorithms allow us to extract the text from documents in a warehouse environment. These algorithms can read difficult text off box labels, handwritten characters, text from fuzzy or poorly taken images, and product images.
We leverage powerful pretrained algorithms like TrOCR as a backbone for our architecture then use in-domain fine-tuning to boost the accuracy of the product for your specific warehouse use case.
We use a deep learning based information extraction framework that allows us to extract text positioning, text size, key phrases, classification, and other parameters to understand the relationship between text fields in the document (‘QTY’, ‘Size’, ‘Item’ above) and the specific text that correlates.
We’re able to create reasoning through the relationship learned and support a number of downstream processes that require multiple variables or higher order variables.
Once the data has been extracted from these documents and mapped to any respective fields in the document, or some internal rules, we need to map to actual fields in our warehouse backend or database. Oftentimes the fields extracted for a document are either:
- Not exactly what we call them in our internal database for some business use reason. This happens when our internal system calls fields something different then what the document calls the same information
- The format that these fields are extracted in from the document does not fit our internal structure. This commonly happens when either the document or internal database combines or splits data into different fields.
We use a number of different NLP algorithms to postprocess this data into use case specific output.
Automated invoice matching allows you to automatically match text information for standardization across documents such as purchase orders, receipts, invoices, and the inspection report. This surprisingly manual task can be automated out using the same pipeline as above to extract key data like line items, product titles, quantity, price, and contract terms and validate it across all the required documents.
Product recognition software uses computer vision techniques like OCR and object detection to recognize areas where a product is located in an image or video stream, then uses embedded vector similarity to match the detected product area to a specific SKU or unique identifier in a database. This allows for the automation of a number of warehouse and logistics operations.
This architecture takes traditional object recognition a step further and provides a bunch of huge benefits for a warehouse environment:
1. The dynamic nature of comparing product area to a database allows you to evaluate products at a deeper level than you can with traditional object recognition. This lets you identify exact SKUs or sizes for the same overarching product.
2. The dynamic natural also allows you to support new products or box designs without needing to retrain the models at all. In traditional object recognition you would need to retrain the pipeline each time you want to add a new product or box design to the possible matches. With product recognition you just need to update the database images to add entirely new products or boxes.
3. You can get started with much less data and can scale to many more unique SKUs and IDs.
We deploy this as a cloud based rest API that allows you to use the pipeline on any website, phone app, or backend system with a camera. The pipeline returns product recognition results quickly and can provide alerts, confidence scores, and other management information.
Here’s an example from a real use case of how the benefits of this architecture can be used. Let’s say we’re a distributor of Coca Cola products that supports all Coca Cola products all year around. Both of the below boxes have different SKUs and must be recognized as different products.
To tell these two products apart with object detection we would need 100s of image examples with varying noise levels and multiple rounds of training. Any time the box changed to a seasonal design or throwback design we would have to train again.
To support this new box design for the 12 can pack we would have to collect images and retrain again. With the product recognition architecture we just have to add this image to the database and associate it with the unique ID of the original 12 can pack. The constant need for updating to new box designs, adding new SKUs, or other large changes to the product we are trying to recognize grows as we try to scale the softwares use.
Warehouses and Distribution Centers are always looking to better utilize their space based on current and future inventory. These businesses become overwhelmed due to seasonal changes, unforeseen late or early inventory, and poor space management. This results in poor efficiency in terms of stock going in and out, higher costs, and more manual labor required per inventory action. Planning and optimizing the required space is a difficult task to solve manually as the number of variables such as total warehouse space, forecasted inventory, demand optimization, and others make it difficult to model.
Warehouse space and slotting optimization algorithms work to setup the problem as a multi-variable optimization problem while using human input to determine the objectives and constraints of the slotting optimization. These algorithms have been shown to have huge ROI in warehouse operations such as 120-150% increase in current storage space, 37% productivity rate increase, and a huge reduction in operational costs (study).
The key gains come from a learned understanding of these concepts:
1. How much space a specific item(s) requires relative to how often it is accessed or moved.
2. The velocity of input to and output from the warehouse for a specific item relative to the above points based on historical and forecasted data. This data is often used in a simulation type environment to optimize available space for new products coming in and the shifts in volume in different seasons. One of the cool benefits is that your sales data acts as the only data needed for this step, which makes it easy to get off the ground.
3. What are the key objectives and constraints to this specific use case. This can be anything such as certain products having to be grouped together, number of laborers required to move a product, color of the product etc. We’re huge fans of focusing heavy on this step as with most ML/Ai problems a clear definition of the objectives makes it easier to reach a successful end.
Automated sortation systems have been around for years in manufacturing. These systems use RFID scanners and sensors to sort and divert products on a conveyor belt. Automated sortation systems are crucial to the entire fulfillment process, from picking to receiving and even shipping. These systems can integrate nicely with the product recognition architecture from above to sort products without RFID. Let’s look at the different types of sortation systems and applications:
data matching software uses natural language processing techniques to compare various product data fields to each other for similarity. These algorithms predict the similarity based on a learned relationship between these data points with a level of variance and noise. The cool part of how modern versions of these algorithms are designed allows for you to add or subtract fields on the fly and still reach a high level of accuracy. An example of this can be seen in our customer record matching system that allows fields like email, phone number, or username to be blank and still reach a production level system.
This software integrates perfectly with the document processing and product recognition use cases from above to further enhance the automation possibilities. We’ve fine-tuned the models to compare incoming product data from external sources such as vendors or sellers to an internal database of SKUs or unique IDs. This allows you to manage the incoming products and invoices from multiple sources to a single identifier while these external sources have different formats or field requirements. Here’s a UPC code matcher that can match extracted codes from documents or product labels with a huge noise variance level. The model even generates a reason for why these two UPC codes are match!
The first thing people usually associate with warehouse automation is improved ROI. However, in a modern warehouse, automation does much more than just help contain costs. The Amazon-ification of supply chain and logistics means that customers now have high expectations from brands when it comes to order fulfillment. Same-day deliveries, comprehensive tracking capabilities, and easy returns have become the price of entry for companies in the retail logistics business. The trend is gradually starting to take shape in B2B environments too.
Here’s why you should consider automation your warehouse operations:
Warehouses and distribution centers have an unique opportunity to develop their competitive advantage by leveraging machine learning and IoT for automation. However, identifying what sort of automation is needed and integrating it into your WMS is key. You need a solution that ensures that every step of your warehouse operation is optimized for automation.
Want to learn more about implementing a successful warehouse automation project? Book a free consultation with us to understand how our warehouse automation experience can fuel your growth.
