Humanoid

Choosing the platform is the short part. This is the shape of the work that follows, in the order it usually happens, because the second question after "which one" is "and then what".

The safety concept comes first, not last.

A machine of this size sharing a floor with people needs a risk assessment for your installation, a defined working envelope, a rule about who may enter it while the machine is active, and an agreed way to stop it. That assessment is part of the project and it needs a name against it on your side, usually whoever is already responsible for machine safety in that area. Starting it late is the single most common reason a deployment that works technically cannot be switched on.

Then the task is taught rather than described.

The application layer is built on ROS 2, and for most industrial work the reliable path is a taught sequence with vision used only where something genuinely moves. Where the task is too varied to teach directly, a person drives the machine through it in a headset while the system records, and those recordings become the training data. That is not a fallback, it is how this generation of machines is programmed.

Integration is where the schedule actually goes.

A humanoid that works in isolation and cannot be told what to do next by your existing systems is a very expensive island. Expect work on the interface to whatever holds the orders, whether that is a PLC, a warehouse system or a spreadsheet somebody maintains. This is ordinary integration engineering and it is estimable, but it is rarely in the number a buyer has in their head.

Layout follows runtime.

Published runtimes across these platforms range from around an hour and a half of walking to ten hours with hot-swap packs, so a machine that cannot cover your shift needs either a swap routine with somewhere to keep charged packs or a charging window designed into the process. Decide that before the cell is laid out rather than after.

Finally, the people.

Two or three operators who know how to start it, stop it, clear a fault and recognise when to call. A documented handover rather than a demonstration. Where nobody on site owns the machine, it quietly stops being used, and that has ended more deployments than any technical limit on this page.