Autonomous humanoid robots: who is really driving?. Offbrand Human

As of mid-2026, no humanoid robot you can buy does open-ended household tasks by itself. That is the whole answer, and it fits in one sentence.

The longer answer is that the good demos are real, but the phrase "autonomous humanoid robot" hides three different mechanisms, and one of them is a person wearing a VR headset in an office you will never see.

The three words sold as one

When a video shows a humanoid folding laundry, exactly one of these is happening, and the difference is the entire story:

Teleoperation. A human is driving. Either "remote driving" (a person controls every motion) or "remote assistance" (a person gives high-level commands). The robot is a puppet with good posture. Reference
Learned / semi-autonomous. The robot does a narrow, trained task unattended, having learned it from human demonstrations. Works until the situation drifts off what it was shown.
Scripted. Fixed, pre-programmed routines. Reliable, and about as flexible as a music box.

Watch 1X's own home-demo film with the three buckets in mind and try to name which one each scene belongs to; the film never says.

Marketing collapses all three into "autonomous." Your job as a buyer is to pull them back apart. The tell is simple: ask whether a person can drive it, and what that person can see and hear inside your home during a remote session.

Four-rung autonomy ladder sorting the three words apart: rung 1 scripted, fixed pre-programmed routines where most dance and demo videos live, illustrated with a photo of a Unitree H1 mid-stride on a concrete plaza, Unitree bipeds like it ship with zero home autonomy out of the box; rung 2 teleoperated, a human drives it in a VR headset, illustrated with a photo of the 1X NEO head, $20,000 or $499 a month, hard tasks handed to a 1X employee; rung 3 learned and narrow, illustrated with a still from Figure's staged Helix bedroom demo showing a Figure 03 leaning over a bed it is making, Helix learned from about 500 hours of human demos and is not for sale; rung 4 open-ended, new household tasks unsupervised, zero products here and so nothing to photograph.

The ledger: every major maker, dated

Robot	Autonomy today	What that means	As of
1X NEO	Teleop-assisted ("Scheduled Expert Mode")	Basic pre-trained behaviors; for anything hard, a remote 1X employee drives it via VR	2026-08
Unitree G1 / R1	Programmable platform	No autonomy out of the box; it does whatever you or a lab code onto it	2026-08
Figure 03	Learned (Helix VLA), industrial pilot	Most-autonomous-looking; deployed at a car factory, no consumer autonomy to buy	2026-06
Tesla Optimus	None you can buy; "not doing useful work"	Musk's own words, January 2026; no orderable product	2026-01
Boston Dynamics Atlas	Research autonomy demos	Impressive dynamic autonomy, but not a product for sale	2024 to 2026

Row by row, here is what sits behind those one-liners. (Atlas gets no section below: it is a research program, not a product, and Boston Dynamics does not pretend otherwise.)

1X NEO: a person in a VR headset, disclosed in the fine print

NEO is the flagship example of teleop sold as autonomy, and to 1X's credit the fine print says so. Its product page promises "basic autonomy for early owners" and then, in the same breath: "For complex tasks NEO doesn't know, an Expert from 1X can remotely supervise its actions at scheduled times." Fortune reported in February 2026 that NEO could not do household tasks autonomously and that 1X employees teleoperate it through VR headsets.

1X's NEO Beta standing waist-up in a wood-panelled room, hooded gray and black knit suit, blank black visor, both palms turned up so the bare mechanical fingers show, a maker hero still that says nothing about who would be moving those hands. — A 1X press still of NEO Beta, posed in a living room with its hands open and empty. Nothing in the frame distinguishes a trained behavior from an Expert in a headset. Photo: 1X

So the $20,000 (or $499/month) robot does simple trained things on its own and hands the hard stuff to a human in an office. That is the current top of the consumer autonomy ladder. It comes with a stranger who can see through your robot's cameras. Full status on the Neo tracker.

Unitree G1 and R1: autonomy is a kit you assemble

The Unitree bipeds ship with zero home autonomy. That is the product as sold. The G1 is a programmable platform. It walks and balances out of the box; anything useful is code you or a research lab write. Unitree explicitly tells individual buyers to understand the limitations first. Buying a G1 expecting an autonomous helper is like buying a Raspberry Pi expecting a finished app.

Unitree's studio render of the G1 holding a deep crouch on a plain lilac-gray sweep, white panels, black head with a lit blue visor arc, which is the sort of balance and joint range you actually get out of the box before any code is written. — A maker render of the G1 crouching. Poses like this are the shipped feature set: balance and motion, with autonomy left as an exercise for the buyer. Render: Unitree

Figure 03: the most autonomous, and you cannot have it

Figure is chasing real autonomy hardest, and its architecture is worth understanding because it is the template everyone copies. Figure's Helix is a Vision-Language-Action model with a two-speed brain: a 7-billion-parameter vision-language "System 2" running at 7 to 9 Hz for understanding and planning, feeding an 80-million-parameter "System 1" transformer that fires motor commands at 200 Hz for reflexes. Figure says it learned broad tasks from about 500 hours of demonstration data. It looks like the future. It is also deployed to BMW, not homes, and there is no consumer price. Most autonomous, least buyable.

Figure's own footage is the best available calibration for what the learned rung of the ladder looks like: fluent on a staged task it was trained for, and silent on everything it was not.

The bedroom in that film is a demo. The job Figure 03 actually reported for is a "logistics workflow at BMW Group Plant Spartanburg", which Figure posted itself in June 2026.

Figure says Figure 03 has started performing a logistics workflow at BMW Group Plant Spartanburg (@Figure_robot, 2026-06-30)

The dated maker statement behind the Figure 03 row in the ledger, and the work it names is a factory logistics workflow rather than anything domestic.

Tesla Optimus: the autonomy that is not doing anything

Optimus generates the most "autonomous robot" search traffic and the least autonomy you can verify. In January 2026 Musk admitted "no Optimus robots are doing useful work" at Tesla. There is no orderable product and no price. When you see an Optimus clip, assume nothing about autonomy until Tesla says which of the three words applies. Timeline and slippage.

Tesla's We, Robot promotional image of an Optimus handing a drink to two women on a yellow patio sofa at an evening backyard party, string lights and other guests behind it, a chore scene staged with no word on what is driving the robot. — Optimus serving drinks at a staged evening party in Tesla's own We, Robot imagery. The picture shows the chore and, like the clips, never names which of the three words applies. Image: Tesla

Why teleoperation is the business model

The remote human is doing two jobs: finishing the chore, and generating the data meant to eventually make the remote human unnecessary.

Modern humanoids learn by imitation learning, copying human demonstrations recorded as "state-action pairs" (what the robot saw, what the human did). The simplest form, behavioral cloning, has a famous flaw called distribution shift: the moment the robot drifts off the path it was shown, it faces situations no human ever demonstrated and has no idea how to recover. Fixes like DAgger work by having a human keep correcting the robot in exactly those new situations. Reference

Read those two facts together and the strategy is obvious. Every teleoperated session is labeled training data. Putting a teleop-dependent robot in your kitchen in 2026 is how a company harvests the demonstrations it needs to eventually make the robot autonomous. Early buyers are, functionally, paying to train the model that replaces the operator. That is a defensible way to build a robot company. It should also be printed on the invoice.

None of this is a reading between the lines. Pushed on teleoperation by a reply on X, 1X answered that it is "simply a tool to collect high quality robot data to make great AI".

1X says teleoperation is simply a tool to collect high quality robot data to make great AI, and that fully autonomous humanoids has always been the path (@1x_tech, 2025-06-10)

The reply in full. Alongside the data-collection admission, 1X claims full autonomy "has always been the path", with no date attached.

The ceiling nobody is admitting fast enough

Even if the data pipeline works, there is a wall. Melonee Wise, former chief product officer at Agility Robotics, put it bluntly: "the reality of the situation is that currently AI is not robust enough to meet the requirements of the market" (IEEE Spectrum). Walking is a solved-ish demo. Reliable, general manipulation, the grabbing and handling that a household actually needs, is not, and no amount of smarter chatbot brain fixes a gripper that fumbles. (Why that is so hard: Why hands are the hardest part of a humanoid.)

What to actually ask before you buy

Skip "is it autonomous." Ask these:

Can a remote human drive it, and when? (If yes, it is teleop-assisted, plan for the privacy surface.)
What tasks does it do unattended, specifically, and how narrow are they?
What happens when the task drifts off-script, does it stop, or does it need a person?

Answer those and you know exactly which of the three words you are buying. For which humanoids are even orderable, see Humanoid robots you can actually buy in 2026 and the live tracker.