Image credit: Image via source article. Used under fair use for news commentary. · source
Picture a warehouse floor with 200 autonomous mobile robots zipping between shelves. Impressive, right? Now picture the same floor six months later: half those robots are idle during demand lulls, the software can't coordinate with the conveyor system, and the operations team is manually overriding the WMS twelve times a shift.
I've seen this exact scenario play out at three different facilities in the past year. The throughput numbers looked great on paper. The actual operational reality was a mess.
This disconnect between headline metrics and real-world performance is finally getting attention from the industry's biggest players. Daifuku, the Japanese material handling giant with roughly $5 billion in annual revenue, is now publicly arguing that the entire way we measure warehouse automation is broken. And a recent Supply Chain Dive analysis suggests the fix isn't buying more hardware. It's rethinking the software layer that ties everything together.
Throughput, measured in units picked per hour or orders fulfilled per day, has been the gold standard for warehouse performance since the first conveyor belt was installed. It's simple. It's measurable. And it's increasingly misleading.
The problem is that throughput measures peak capacity under ideal conditions. It doesn't capture:
Related coverage
More in Industrial
Consumer tech sales are booming while industrial computing quietly falls behind, and I'm not sure anyone's paying attention.
Robert "Bob" Macintosh · 1 hour ago · 3 min
After decades of chasing throughput numbers, the industry is finally figuring out what I've been saying since the 90s: the software matters more than the hardware.
Mark Kowalski · 3 hours ago · 6 min
Home Depot's pushing DeWalt deals while the factories making those tools are running skeleton crews of robots. Something doesn't add up.
Robert "Bob" Macintosh · 3 hours ago · 4 min
Amazon and Best Buy are slashing prices on storage, but if you're building out a robotics workstation or edge computing setup, the deals aren't quite what they seem.
How well the system adapts when demand spikes unexpectedly
Whether different automation subsystems actually communicate with each other
The labor hours spent on manual interventions and workarounds
Energy consumption during low-demand periods
Recovery time after equipment failures
From my time building hardware at Fanuc, I learned that the spec sheet number and the real-world number are often different things entirely. A robot arm might be rated for 120 picks per minute in a controlled test environment. Put it in a warehouse with variable lighting, inconsistent SKU sizes, and a WMS that sends conflicting commands, and you're lucky to hit 80.
Daifuku's argument, which I find compelling, is that warehouses need to shift from measuring throughput to measuring what they call "system intelligence." That's a somewhat vague term, but the underlying idea is concrete: how well does the entire operation respond to changing conditions?
Look, I'm generally skeptical when automation vendors introduce new buzzwords. But the concept here maps to real operational challenges I've observed.
System intelligence, as Daifuku frames it, involves three capabilities:
Real-time adaptability. Can the warehouse management system dynamically reallocate resources when conditions change? If a picking zone gets backed up, does the system automatically route orders elsewhere, or does a human have to intervene?
Cross-system coordination. Most warehouses run multiple automation technologies: AMRs from one vendor, conveyor systems from another, picking stations from a third. These systems often speak different languages. True system intelligence means they coordinate without manual translation layers.
Predictive operations. Rather than reacting to problems after they occur, intelligent systems should anticipate bottlenecks, equipment failures, and demand shifts before they impact throughput.
The challenge is that none of these capabilities are easy to measure. You can't put "adaptability" on a spec sheet the way you can put "500 units per hour." This is where the industry's metrics problem becomes self-reinforcing: we measure what's easy to measure, which incentivizes vendors to optimize for those metrics, which further entrenches the measurement approach.
The Supply Chain Dive analysis makes a point that, honestly, I think the industry has been slow to internalize: the value of warehouse robotics increasingly lives in the software layer, not the hardware.
This isn't to say hardware doesn't matter. It obviously does. But the marginal gains from faster motors or more precise sensors are diminishing. The marginal gains from smarter orchestration software are, in a way, just getting started.
Consider a typical AMR deployment. The robots themselves are largely commodity hardware at this point. The differentiation comes from:
Path planning algorithms that minimize congestion
Task allocation systems that balance workload across the fleet
Integration middleware that connects to existing WMS and ERP systems
Simulation tools that let operators test changes before deploying them
AI, specifically machine learning models trained on operational data, can improve all of these. A well-trained model can predict which zones will see demand spikes based on order patterns. It can identify when a robot is likely to fail based on subtle performance degradation. It can optimize pick paths in ways that would be impossible to program manually.
The catch, and there's always a catch, is that these AI systems require massive amounts of high-quality operational data. Most warehouses don't have the data infrastructure to support this. They're running on legacy WMS platforms that were designed for batch processing, not real-time machine learning.
This is where I remain somewhat skeptical. The rhetoric is shifting, clearly. Major players like Daifuku are talking about system intelligence. Warehouse operators are, according to multiple industry surveys, increasingly prioritizing software capabilities over hardware specs when evaluating automation investments.
But the procurement process hasn't caught up. RFPs still ask for throughput guarantees. ROI models still focus on labor cost reduction. The metrics that matter for getting a deal signed are not the same metrics that matter for long-term operational success.
I've talked to warehouse managers who understand this disconnect perfectly well. They know that a system with lower peak throughput but better adaptability might be the smarter investment. But they can't justify it to their CFO using traditional financial models.
The companies that figure out how to bridge this gap, how to quantify system intelligence in terms that finance teams understand, will have a significant competitive advantage. It's too early to say who that will be.
If you're evaluating automation investments right now, here's what I'd suggest based on, well, everything I've seen go wrong over the past few years.
Demand interoperability upfront. Don't accept proprietary systems that lock you into a single vendor's ecosystem. Ask specific questions about API availability, data export formats, and integration with third-party WMS platforms. If the vendor gets cagey, that's a red flag.
Test under realistic conditions. Pilot programs should include demand variability, equipment failures, and integration with your existing systems. A pilot that only demonstrates peak performance under ideal conditions tells you almost nothing useful.
Invest in data infrastructure. Before you buy robots, make sure you have the systems in place to collect, store, and analyze operational data. This might mean upgrading your WMS, implementing IoT sensors, or hiring data engineering talent. It's not glamorous, but it's foundational.
Rethink your metrics. Start tracking things like recovery time from disruptions, manual intervention frequency, and cross-system coordination latency. These won't replace throughput, but they'll give you a more complete picture.
Budget for software. The industry rule of thumb used to be that software was maybe 20% of a robotics deployment budget. That number is probably too low now. Plan for ongoing software licensing, integration work, and AI/ML development.
None of this is revolutionary advice, basically. It's common sense that the industry has been slow to adopt because the old metrics were easier and the old approach worked well enough during the e-commerce boom years.
But "well enough" isn't going to cut it in a more uncertain demand environment. The warehouses that thrive will be the ones that measure what actually matters, even when it's harder to measure.
