Step inside one of today’s high-tech warehouses, and you might marvel at the high-speed conveyors, voice-operated picking headsets, or fleets of autonomous mobile robots (AMRs) bustling about. But you’d be hard-pressed to point out any concrete examples of one of the most advanced technologies in the facility: artificial intelligence (AI).
Although it’s fast becoming an industry buzzword, AI is little understood outside of engineering circles, and its impact on logistics operations is hard to trace. But the truth is, the technology is already widely used, powering everything from the conversational interface on the smartphone in your pocket to the warehouse management system (WMS) that controls the flow of goods through the DC.
So if you can’t see the AI in your warehouse, how can you get a handle on it? That is, how do you select a good system, judge its effectiveness, and measure its impact on your business over time? To get answers to these and other questions, we asked some experts to share their thoughts about AI and the warehouse.
To begin with, organizations that want to be successful at adopting AI have to change their basic approach to buying warehouse technology, says Peter Chen, co-founder and CEO of Covariant, which develops AI for commercial devices like robotic picking arms.
That’s because AI operates in a fundamentally different way from previous generations of logistics and material handling tools. Twenty years ago, logistics managers chose hardware—such as forklifts or conveyors—based on quantifiable attributes like speed, strength, and durability. As technology progressed and they began to select software—like a warehouse control system (WCS) or a WMS—they added criteria like cybersecurity, tech support, and ease of upgrades to the list. And now to buy AI systems, they need to adopt a new set of strategies, he says.
There are a couple of reasons for that. For one thing, AI differs from other technologies in that it becomes more, rather than less, effective over time—in direct contrast to, say, hardware that slowly breaks down with use or software that eventually becomes obsolete. What sets AI apart is that it doesn’t rely on “programmed intelligence,” Chen says. “With AI, you have intelligence that is not preprogrammed; instead, it learns from data and learns from experience. As opposed to static behavior, it learns from its own trial and error, and improves over time.”
In Covariant’s case, that learning curve enables machines like robotic arms to handle an ever-evolving and expanding range of items without requiring software upgrades or engineering studies, Chen says. Instead, the arm experiments with a wide array of stock-keeping units (SKUs) and slowly refines its ability to grasp items of various types, whether it’s apparel, grocery items, pharmaceuticals, or cosmetics.
Another factor that differentiates AI from other technologies is that companies get the best results when they start as soon as possible. Just as financial advisers tell clients to start investing early in life so their savings can grow through compound interest, AI works best when it has time to learn and develop. That contrasts with the typical hardware-buying strategy of waiting to refresh or replace equipment until the vendor rolls out the latest version. “The best way to buy AI is to get going as early as possible, because it can start learning ASAP,” Chen says. “Roll out your first site as quickly as possible so [the system] can collect data and start learning. The goal is to gather vast amounts of data, then develop analytics and actionable insights, so it compounds the results of AI adoption.”
Measuring the results is a critical step in justifying any warehouse purchase, but it comes with an added challenge for AI because artificial intelligence typically operates “behind the scenes,” says John Black, senior vice president for product engineering at Brain Corp. The San Diego-based firm develops AI software and analytics to run AMRs from third-party manufacturers, with a focus on the automated floor-cleaning robots found in factories, DCs, retail stores, and office buildings.
Just as most people don’t know what type of microchip is powering their personal computer, most users of AI-powered devices can’t pinpoint exactly which functions rely on artificial intelligence. That makes it tough to gauge how well the technology is working, particularly because AI is typically held to a pass/fail standard—if a machine’s logic makes a single mistake, the entire device is seen as defective. For example, as an AMR cruises through a DC, it executes dozens of AI-enabled steps along the way, from localization and navigation to data gathering and analytics. If it fails at any one of those steps, then the AMR is basically useless. “You have to get all the way there,” Black says. “You can get most of the way there, and that is interesting, but it’s not enough to get a [return on investment]” for the company that bought the AMR.
“[AI] has to be nearly perfect. The measure is, how much time can this robot go without an intervention? You can send an employee over to fix a problem on an AMR, but every touch [diminishes the system’s return]. The goal is no-touch autonomy,” he says. “What you’re paying for with automation is accuracy and repeatability. If you have to have a person babysitting it, essentially you’ve just changed their job to overseeing the task and haven’t truly repurposed that employee from a labor standpoint.”
By that measure, AI works best when people forget they’re even using it, agrees Mike Myers, director of solutions at Third Wave Automation. The company incorporates its AI into reach trucks built by partner companies, allowing those forklifts to become autonomous vehicles.
Myers points to AI that has run for years as a basic “rules engine” in the accounting software many people use to file their personal tax returns. More recently, some developers of tier-one warehouse management systems have applied AI to the complex puzzle of managing fulfillment operations in a busy e-commerce DC. “And in a WMS, the AI is invisible in how it works. That’s how you know things are effective—when people don’t have to go into the WMS; they can just go to the end points” and follow the software’s guidance, he says.
Striking a balance between automated decision making and human oversight is key to generating a solid ROI (return on investment) from an AI system, Myers says. But to measure how independently the AI in your warehouse is performing, you need to know exactly what it’s doing. And that can be a challenge.
A common misconception about AI is that it acts as “general intelligence,” functioning like a sentient robot in a Hollywood movie, Myers observes. But the truth is that most AI performs a series of small jobs, as opposed to pondering big questions like the meaning of life. “AI is in the vehicle navigation, the high-level route planning, and the sequencing of tasks in a facility, and it’s also in Siri on your iPhone,” Myers says. But as impressive as a tool like Siri is, it works through a series of machine learning and language processing steps, not through an umbrella of overall awareness, he explains. “So ‘general intelligence’ AI is not necessary for practical use cases; you can break up all those cases to achieve each step.”In the end, the best way to measure an AI system’s impact on your logistics operations is to go back to the classic supply chain yardstick—the key performance indicator (KPI). “KPIs don’t change, whether you’re looking at cost per unit, SLA [service level agreement] adherence, or whatever,” Myers says. “Consistency in meeting those numbers is a measure of effectiveness. The AI is just a component, one machine in the entire system. But because AI is self-improving, [the fact that you’re] making progress toward those KPIs is how you know it’s working.”