Factories have been automated since the 1970s, so AI is landing only where classical automation always broke: unfixtured parts, unseen defects, and machines with no failure history. A grounded look at what is in production versus what is still a single-task pilot.
In BMW's Spartanburg plant in South Carolina, a Figure 02 humanoid spent part of 2024 doing one job: picking sheet metal parts and placing them into a fixture in the body shop. Not welding, not painting, not roaming the line. One insertion task, in one cell, alongside conventional automation that has been doing higher-throughput work there for decades.
That deployment got a lot of coverage, and the coverage mostly missed the point. The interesting thing isn't that a humanoid stood on a factory floor — it's why the task was scoped that tightly. Factories are the most automated environments on earth and have been since the 1970s. What AI is changing in manufacturing is not "robots arrive." Robots arrived fifty years ago. What's changing is the narrow band of tasks where the part isn't in a known pose, the defect isn't in a known catalog, and the sequence isn't in a known program — the exact places where classical industrial automation has always fallen over.
Here's the honest state of play across the four areas where AI is actually deployed in manufacturing and logistics today, what the technology under each one really is, and where each one currently stops.
| Application area | What the AI actually does | Maturity | The real blocker |
|---|---|---|---|
| Warehouse & fulfillment robotics | Grasp planning, item ID, fleet routing | Production at scale — hundreds of thousands of units | Edge cases in the long tail of SKUs |
| Visual defect inspection | Supervised/anomaly-detection vision on line cameras | Production, widely | Few defect examples; drift when the line changes |
| Predictive maintenance | Anomaly detection on vibration/thermal/current signals | Production, mixed ROI | False-positive fatigue; labels are rare by definition |
| General-purpose humanoids / VLA policies | End-to-end vision-language-action control | Pilot — single tasks, single cells | Reliability per attempt, cycle time, and safety certification |
The gradient here is not about how impressive the AI is. It's about how much structure the environment provides. Fulfillment centers were redesigned around robots. Factory floors were not redesigned around humanoids, and that's most of the difficulty.
Amazon is the clearest case, because it's the largest and the most documented. Its fulfillment network runs a fleet of mobile and manipulation robots numbering in the high hundreds of thousands — Amazon has publicly cited more than 750,000 and has more recently described the fleet as approaching a million units. The named systems are worth distinguishing, because they're doing very different kinds of work:
The AI content varies enormously across that list. Proteus is mostly a navigation and safety problem, well-served by classical planning plus learned perception. Sparrow is the one that genuinely needed modern machine learning: picking an arbitrary item out of a jumbled tote requires estimating a grasp on an object whose pose, deformability, and packaging you did not specify in advance. That's a learned-perception problem, and it's the reason Amazon licensed technology from Covariant and brought its founders in-house in 2024 — Covariant's pitch was precisely a general-purpose picking model trained across many warehouses rather than programmed per-site.
The newer layer is fleet-level. In 2025 Amazon described DeepFleet, a foundation model for coordinating robot traffic across a facility — treating routing as a learned sequence-prediction problem rather than a hand-tuned scheduler. Amazon's own figure was around a 10% improvement in fleet travel time. Take single-vendor numbers with the usual caution, but the architectural direction is real and it's the same direction as everywhere else in AI: replace a stack of hand-written heuristics with one model trained on the operational log.
Where it stops: the long tail. A picking model that succeeds 99% of the time on a catalog of hundreds of millions of items still fails tens of thousands of times a day, and every failure has a physical consequence — a dropped item, a jammed cell, a human intervention. The economics of warehouse robotics are dominated not by average success rate but by mean time between human interventions.
Automated optical inspection is the least glamorous and most successful application of AI in manufacturing. A camera watches a line; a model flags parts that look wrong. It works because the setup is nearly ideal for supervised vision: fixed camera, fixed lighting, fixed part geometry, high volume.
The wrinkle that makes it an AI problem rather than a machine-vision problem is data imbalance. On a well-run line, defect rates are measured in parts per million. You cannot train a conventional supervised classifier on a class you have forty examples of. So production systems lean on two workarounds:
Where it stops: drift. The model is trained on a line configuration, and lines change — a new supplier's raw material, a different lighting ballast, a seasonal humidity shift in the plant. Inspection models degrade silently, and the failure mode is the dangerous direction: false negatives look like a clean line. Every serious deployment I'd trust has a sampling-based human audit running underneath it precisely to catch this.
The pitch is irresistible: sensors on a machine, a model that learns its normal vibration and thermal signature, an alert before the bearing fails instead of after. Siemens' Senseye product and a dozen competitors sell exactly this, and it does work — on high-value, well-instrumented, failure-expensive assets.
The reason it hasn't eaten the world is a labeling problem that's structural rather than fixable. Supervised failure prediction needs failures. A critical asset that fails twice in five years gives you two positive examples. So most real deployments are unsupervised anomaly detection wearing a predictive-maintenance label — they tell you "this machine is behaving unusually," not "this bearing will fail in eleven days."
That distinction determines the ROI. An unusualness alert has to be triaged by a maintenance engineer. If the false-positive rate is high enough that engineers start ignoring alerts, you've built a very expensive way to generate noise. The deployments that succeed are the ones that constrain scope hard — a specific failure mode on a specific asset class, with a physical model of that failure informing the features, not a generic "put a model on every motor" rollout.
This is where the genuinely new AI is, and where the deployment reality is thinnest. The shift is architectural. A classical industrial robot cell looks like this: engineers define the part pose, program the trajectory, and any deviation is an exception that halts the line. A vision-language-action (VLA) model collapses that stack — camera frames and a task instruction go in, joint actions come out, with no explicit pose estimation or trajectory programming step.
The lineage is short and recent. Google DeepMind's RT-2 showed that a vision-language model fine-tuned on robot trajectories could generalize instructions it hadn't been trained on; Gemini Robotics and Gemini Robotics-ER extended that with explicit spatial reasoning and shipped through partners including Apptronik. Physical Intelligence released π0, a flow-matching VLA trained across a diverse fleet of robot embodiments, and open-sourced weights — which is why it's become the default research baseline. Figure moved off its OpenAI collaboration in early 2025 and shipped Helix, its own in-house VLA. NVIDIA's Isaac GR00T line packages a humanoid foundation model together with the simulation stack used to generate training data for it.
Two things about this cluster are worth separating from the marketing.
Language models had the internet. Robot policies have nothing comparable — every training trajectory requires a physical robot moving through physical time, and the data doesn't transfer cleanly between robot bodies. The field's three answers are teleoperation (expensive, slow, high-quality), simulation with domain randomization (cheap, fast, and burdened by the sim-to-real gap — contact dynamics and deformable materials remain the hardest things to simulate faithfully), and cross-embodiment training (pool data from many robot types and hope the shared structure outweighs the mismatch). Open datasets like Open X-Embodiment exist precisely because no single lab can brute-force this alone.
A policy that completes a task 95% of the time is a genuinely impressive research result and an unusable production system. At a 30-second cycle time, 95% means a failure roughly every ten minutes. Classical automation in the same cell runs at rates where failures are counted per shift, not per ten minutes. This is the actual reason the BMW deployment was one insertion task — not conservatism about humanoids, but the arithmetic of what per-attempt reliability does to a line that only makes money when it runs continuously.
Then there's certification. Industrial robot safety lives under ISO 10218 (parts 1 and 2, substantially revised in 2025) with ISO/TS 15066 covering collaborative operation — speed-and-separation monitoring and power-and-force limiting. Both frameworks assume you can characterize a robot's behavior in advance. A learned policy whose action distribution shifts with the visual scene is an awkward fit for a regime built around deterministic, analyzable motion. In practice this is handled by wrapping the learned policy in classically-verified safety envelopes — speed limits, force limits, geofenced workspace — which is sensible engineering and also a hard ceiling on what the policy is allowed to do.
The pattern across all four areas is the same, and it's the opposite of the usual narrative. AI isn't replacing industrial automation; it's being inserted at the specific joints where classical automation required certainty that the real world doesn't provide — the pose of an unfixtured part, the appearance of an unseen defect, the health of a machine with no failure history.
Three things follow if you're evaluating any of this: