Humanoid robots vs AMRs in UK warehouses: which does which job?

No, and the question is the wrong shape. An autonomous mobile robot (AMR) is a better machine for moving goods along repeatable routes between fixed points, and it is cheaper, faster and more proven at that job than any humanoid on the market. A humanoid is a better machine for handling goods in spaces and at fixtures that were designed for a person: picking a tote from a shelf at chest height, loading a conveyor, working a station that has no automation interface. The realistic UK warehouse in 2026 uses AMRs for movement and is trialling humanoids for the handling that AMRs leave to people.

That split matters because the two are usually bought by different processes. AMRs are a mature purchase with reference sites, fleet software and a known return. Humanoids are a pilot decision, made to learn something specific about a task before any wider commitment. Treating a humanoid as an AMR alternative produces a slow, expensive AMR; treating it as a handling trial produces a decision.

An automated guided vehicle (AGV) follows a fixed guide, historically a wire, magnetic tape or reflectors, and stops when its path is blocked. An AMR carries its own map and sensors and plans a route around obstacles, so it needs no guide infrastructure and copes with a changing floor. A humanoid robot has legs rather than wheels and arms rather than a load deck: it is not primarily a vehicle at all, and its value is in reach and manipulation rather than in transport. Fixed automation, conveyors, sorters and automated storage and retrieval systems, moves the most goods per hour of anything on this page, and moves them only where it was built.

The UK regulatory framing follows the same lines. Driverless industrial trucks, which is where AGVs and most load-carrying AMRs sit, are covered by BS EN ISO 3691-4. Robots with arms, including humanoids, are covered by the industrial robot standards ISO 10218 and, for collaborative operation, ISO/TS 15066. Both sit under the Provision and Use of Work Equipment Regulations 1998 and the Health and Safety at Work etc. Act 1974. A humanoid pilot and an AMR purchase therefore need different risk assessments, not one shared document.

The table is a decision aid, not a scorecard. Read each row for the task you have, and let the machine follow from it.

At the ends of the AMR journey. A goods-to-person AMR fleet delivers a tote to a station where a person picks from it, and collects it when they are done. The person is there because the pick, the induction into a conveyor, the decant from a supplier case into a tote, or the machine tending at a station has no automation interface. That is exactly the handling a humanoid is designed for. Agility Robotics lists tote handling, palletising and depalletising, machine tending, and kitting and sequencing as Digit’s programmed workflows, and names GXO, Amazon, Toyota Motor Manufacturing Canada, Mercado Libre and Schaeffler among its deployments and partners. Those are the human-shaped gaps between fixed and mobile automation, not replacements for either.

The economic test is simple. If the station can be re-engineered so that an AMR or a conveyor serves it directly, that is normally the cheaper answer. If it cannot, because the fixture is shared with people, the building is leased, or the process changes every season, the humanoid is being asked to do the job a person does now, at that fixture, without changing it. Measure it against that person’s rate and error rate, not against a conveyor.

Whenever the task is movement. If the work is “take this from A to B and come back”, a humanoid is a slow, expensive way to do what an AMR does with a fleet manager, a charging schedule and a reference customer list. It is also the wrong answer where the operation can change to suit the machine: a warehouse being fitted out from scratch should design for conveyors and AMRs first and leave humanoids for the handling that remains. And it is the wrong answer for any task that needs raw rate above everything, because no humanoid maker publishes a sustained handling rate, and the honest expectation in a first pilot is a robot slower than the person it works beside.

The mirror case is also true. An AMR cannot pick from a shelf, cannot work a fixture built for hands, and cannot climb a mezzanine stair. Where a warehouse has tried to solve those with more AMRs and more conveyors and found the capital case does not close, the humanoid pilot is the next honest experiment.

Pick one task, write down how a person does it today with a rate and an error rate, and then test the candidate machine against that, not against a brochure. For an AMR that is a fleet quotation against a mapped route with your tote volumes. For a humanoid it is a supervised pilot at one station over two to six weeks, measured on handling rate, accuracy, uptime including charging, interventions per shift and the hours it actually takes off people. Agree the measures and an exit condition before anything arrives. The pilot that concludes “not yet” has done its job. That is how we structure a pilot programme, and the wider picture of what humanoids can do in a UK warehouse is in our main guide.