Picture the same building on the same shift. On one aisle, a mature AMR fleet moves pallets and totes on known paths with WMS integration you already trust. On another, a humanoid pilot reaches into a cart built for people, sequences parts, or tends a machine that was never redesigned for automation. The interesting question is not which robot looks cooler in a keynote. It is which constraint dominates: path geometry you can pave for wheels, or brownfield fixtures shaped like humans.That framing matters because CapEx committees often compare a humanoid pilot’s sticker shock to an AMR’s unit cost without aligning on the job. Moving a tote fifty meters is not the same job as inserting a sheet-metal part into a fixture or building a sequenced kit from unsorted bins. If you conflate them, you will either overbuy legs or underbuy dexterity.AMRs still own flat logisticsAutonomous mobile robots—and their older AGV cousins—won for good reasons. Navigation stacks, docking, battery ops, and warehouse-management integrations are mature. Throughput on known paths is predictable. Total cost of ownership models exist with years of fleet data across many vendors and industries. Safety playbooks and standards familiarity are older and clearer for many sites than the still-evolving application guidance around bipedal co-bots working near people.AMRs are weak where the building fights them: stairs, mezzanines, uneven thresholds, narrow human-only corridors, and fixtures that only a human-shaped arm-and-torso stack can reach without ripping out conveyors. If your bottleneck is moving mass along a flat path, wheels remain the default. If your bottleneck is finishing work designed for people, wheels alone often stop one step short—and that last step is exactly where humanoid pilots show up in automotive and logistics case studies.None of this means AMRs are “old.” It means they are specialized. Specialization is usually how automation wins.Where the humanoid form factor is the pointHumanoids earn CapEx when the environment is the constraint:• Reaching into racks and carts built for people without redesigning every shelf height and tote interface.• Machine tending and sequencing where adding a dedicated robot cell would mean months of fixture work and line downtime.• Multi-step manipulation—grasp, scan, place, push a cart—in one mobile body instead of a brittle chain of single-purpose machines.• Bridging islands of automation that AMRs can feed but cannot finish: the last meter of dexterity, not the long haul.• Brownfield speed — sometimes the humanoid’s job is to postpone a multi-million-dollar redesign while you learn which tasks are stable enough to harden later.This is why logistics and auto plants appear so often in humanoid pilots. The building already assumes human reach envelopes. A biped (or wheeled humanoid torso) is a bet that adaptability beats ripping out infrastructure—for selected tasks, not for every meter of floor. Treat selected tasks as the product requirement, not as a footnote.The hybrid reality of 2026Most serious sites will not choose “only AMRs” or “only humanoids.” They will run both. Digit-style bipeds for tote handoffs sit beside AMR undercarriage fleets that own long travel. Wheeled humanoids and wheeled-legged designs appear when stairs are rare but human-height manipulation is not. Semi-humanoid industrial platforms marketed for human-height work (for example wheeled bases with humanoid upper bodies such as those discussed alongside AEON- and Galbot-class systems) sit in the middle: they keep some reach advantages while dropping some of the bipedal energy tax covered in our battery article.CapEx modeling should compare RaaS humanoid hours against owned AMR fleets using labor hours displaced, quality escapes avoided, and fixture redesign deferred—not robot romance. Include energy stations, spare packs, remote oversight, and integration engineering. A humanoid that looks cheap per hour but needs two teleoperators and a battery tech is not cheap.Decision frameworkUse a short checklist before you fund a pilot:Factor — Favors AMRs / wheels — Favors humanoids / legsFloor flatness & elevation — Flat, mapped, few elevation changes — Stairs, mezzanines, human-only pathsFixture redesign cost — Cheap to redesign for automation — Prohibitively expensive brownfieldManipulation dexterity — Simple transport / docking — Multi-step grasp–place–scanUptime & battery ops maturity — Established docks and fleets — Need explicit swap/charge planSafety & standards scope — Existing AMR/AGV playbooks — New application assessment for bipedsTask variability — Stable, high-volume paths — High mix, human-shaped work cellsIf three or more rows lean wheels, expand AMR first. If three or more lean humanoids, scope a narrow pilot with clear intervention metrics, SKU lists, and exit criteria—not a building-wide rip-and-replace. Hybrid answers (“AMRs for travel, humanoid for the cell”) should be first-class options in the RFP, not afterthoughts.How to run the pilot so finance learns somethingA useful pilot publishes three numbers monthly: autonomous cycle percentage, interventions per hundred cycles, and labor hours displaced (or quality defects avoided) on the targeted task. Pair those with infrastructure costs actually incurred—docks, fencing changes, IT integration, training. If the pilot cannot produce those numbers, it is a science project, not a CapEx input. Expand only the task classes that clear your threshold; keep AMRs on everything else.TakeawayLegs win when the environment is the constraint; wheels win when the path is the product. For the next decade, most plants will run both. Browse HumanoidRobotList’s commercial category—and entries such as Digit, Atlas, AEON, and Galbot—to compare form factors, then model labor and fixture cost honestly before you pick a side. The winning architecture is usually a fleet mix, not a mascot.