What a humanoid robot can actually do today

"What can a humanoid robot do" is usually answered by a demo reel, and the more useful answer is the published payload, joint and sensor figures for the actual platforms on offer.

Payload per arm is the number that decides whether it can do the job, not degrees of freedom.

The A2 Ultra carries about 2 kg per arm across 40 active degrees of freedom; the A3 carries 5 kg per arm with 320 Nm of peak joint torque; the G2 carries 5 kg per arm across 26 force-controlled joints. A platform with more joints is not automatically the one that can lift more, and the payload figure, not the joint count, is what actually rules a task in or out.

Runtime decides whether it covers a shift or needs a routine built around it.

The A3 runs up to 10 hours on hot-swap packs; the A2 Ultra runs roughly 3 hours standing or 1.5 hours walking. A platform on the shorter figure is not disqualified for longer work, it needs a swap or charging window designed into the layout from the start rather than discovered on day one.

Sensors decide whether it can find something, not just move precisely.

A platform without LiDAR or depth sensing executes a taught sequence at a fixed position reliably; it does not locate a part whose position varies or a route nobody mapped. The A2 Ultra and G2 both ship with a full sensor suite as standard; other platforms in this catalogue vary by trim, and checking which sensors are actually on the specific SKU ordered matters more than the word "autonomous" on its own.

What no spec sheet answers is how the platform handles this specific task, which is what a two-week proof of concept is actually for.

A short trial answers whether a task is technically possible on the real parts in the real space, reliably enough to call it a pattern rather than luck. It does not answer questions that only show up over months, like long-run reliability, which is a different measurement on a different timescale entirely.

A humanoid robot today does what its payload, runtime and sensor figures say it does, on the specific SKU ordered, and a two-week trial is the honest way to find where a task sits against those numbers rather than assuming from a demo.