Humanoid systems engineered for real work

Most companies arriving on this page do not need a humanoid, and the fastest way to be useful is to say which ones do.

If the task happens in one place, on parts that arrive in the same orientation every time, a fixed industrial arm will do it faster, more repeatably and for less money. That has been true for forty years and a general-purpose machine does not change it. If the task is moving boxes along a fixed route on a flat floor, a mobile robot or a conveyor is the answer. If the task is inspection along a route with stairs, gravel or grating, a quadruped is the answer, and it is a considerably smaller purchase.

What a humanoid is for is work in a space built for people, too varied to fixture and too low in volume to justify rebuilding the space. Machine tending across several machines at different heights. Picking mixed items from shelving sized for a human reach. A task that changes every few weeks, where reprogramming has to cost less than retooling. Somewhere a person does five different things in a shift and only two of them are worth automating.

The other legitimate reason is timing. Some companies buy now to have hands on the technology and people trained before it becomes ordinary. That is a defensible decision as long as it is the stated one. It goes wrong when it is presented internally as a productivity project and then measured as one.

If you recognise your process in the second paragraph rather than the third, say so on the contact form and we will tell you what we think you should be looking at instead. It costs us a conversation and saves you a year.

A category with this much video attached to it deserves a plain account of what is solid and what is not.

Solid.

Walking and driving over real floors: bipedal platforms handle thresholds, ramps, gravel and stairs, and wheeled ones carry more on a flat surface. Mapping and localisation in a space they have been shown. Speech interaction good enough for a reception desk or a guided tour. Manipulation of known objects from taught positions, which covers a surprising amount of real industrial work.

Improving, not guaranteed.

Grasping arbitrary objects in clutter, where the success rate depends heavily on the objects and where a demonstration on someone else's parts tells you very little about your bin. Long autonomous sequences with nobody watching. Anything that requires the machine to notice its plan has failed and choose a different one.

Still research.

A full shift of unsupervised general-purpose work. No supplier ships that today, whatever the footage suggests, and one who implies otherwise is a reason for caution rather than a reason to buy.

The practical bridge is teleoperation.

A person in a headset drives the machine through the task while the system records what they did, and those recordings are what an autonomous policy is later trained on. The VR teleoperation kit is a catalogue item rather than a research curiosity, and on a first deployment it is often how the work actually gets done while autonomy is built up behind it.

Runtime decides more than it should.

Published figures across the range run from roughly an hour and a half of walking to ten hours with hot-swap packs. A platform that cannot cover a shift needs either a swap routine or a charging window designed into the layout, and that is a floor-planning decision rather than a purchasing one. It is the question we ask earliest, because the answer changes the whole proposal.