Can any humanoid I buy in 2026 do a full job with nobody supervising it?

Not reliably, and not generally. Constrained, single-task setups can run semi-autonomously under supervision (level 3), but continuous unattended general-purpose work (level 4) is still research, not a shippable product. Budget for a human in the loop.

How do I tell if a demo video is teleoperated?

Look for fluid recovery from surprises, single-take framing, jump cuts, and speed-ups, and ask directly: "Was this teleoperated, and what's the unattended success rate over 10 tries in a changed layout?" A serious supplier answers plainly.

If it's mostly teleop-assisted, why buy one now?

Because teleop is how the training data gets built. Buying now lets you collect task-specific demonstrations, develop the ROI case, and be positioned as autonomy improves — rather than starting from zero later. For research buyers, teleop plus an open SDK is the product.

What's the cheapest way to start learning this without over-committing?

A research-class platform with an open SDK. Unitree's R1 ($5,900 US list†) or G1 (from $16,000 US list†) let you teleop, log data, and train policies at MOQ 1 unit. Confirm current single-unit pricing and SDK tier via RFQ.

Does more DOF or more tactile sensing mean more autonomy?

No. High DOF (e.g. Fourier GR-2 at 53†) and dense tactile skin improve the quality of data you can capture and the dexterity a human operator can express — they don't make the robot autonomous by themselves. Autonomy comes from the policy trained on that data.

We only need it for a trade-show booth — are we overthinking this?

Probably yes, in a good way. For a demo unit, level-1 teleop or a scripted routine is the right, cheaper answer. Prioritise reliability and stage presence over autonomy, and don't pay for research-grade capability you'll never switch on.

Teleoperation vs Autonomy: What Humanoid Robots Can Actually Do in 2026

The single biggest expectation gap

The gap between what a humanoid demo shows and what a humanoid delivers on arrival is the most common reason buyers feel misled. It usually isn't the supplier lying. It's that "the robot folded the shirt" and "the robot autonomously folds arbitrary shirts on demand, unsupervised, all shift" are separated by three levels of capability that a 30-second clip cannot distinguish.

Get the vocabulary right and most of the disappointment disappears. There are four levels.

The autonomy spectrum

Level	What runs the robot	What it can handle	Where it is in 2026
1. Pure teleoperation	A human operator, live, via VR rig / motion capture / gloves	Anything the operator can do through the robot's hands — dexterous, adaptive, but 1:1 human effort	Shipping and reliable. The workhorse for data collection and remote tasks
2. Learned skills (imitation learning)	A policy trained on recorded teleop demonstrations	The specific skills it was taught, in conditions close to training (same objects, similar layout, similar lighting)	Shipping in narrow form. Brittle to changes it wasn't trained on
3. Task-specific autonomy	A tuned policy + perception for one job in a fixed cell	One repeatable task, self-run, with engineered fixtures and a safety envelope	Early pilots. Works when the environment is constrained hard
4. General autonomy	On-board reasoning that adapts to novel tasks/objects	"Just tell it what to do" across open-ended work	Not here. Research demos only; do not budget for it

The honest 2026 summary: most humanoids you can buy operate at level 1 or level 2, and the best factory pilots are pushing into level 3 for a single task. Anyone selling you level 4 is selling a roadmap, not a product.

What impressive demo videos actually show

A polished humanoid clip is a real robot doing a real thing. The question is how, and the video rarely tells you. Read every demo against this checklist:

Is there a human in a VR headset off-camera? If the motion is fluid, reactive, and recovers gracefully from a slip, it is very likely teleoperated (level 1). Fluid recovery from surprise is the current tell of a human in the loop.
How many takes? A single clean run says nothing about success rate. A level-2 policy might succeed 6 times in 10; the video shows take 7.
Did the environment change between the demo and the "same" task? Learned skills degrade fast when the object, position, or lighting drifts from training.
Is it sped up, and is the cut continuous? Jump cuts often hide re-grasps, resets, or an operator stepping in.
What's the claimed vs. shown metric? "3.3+ m/s demonstrated"† for a Unitree H1-class machine is a peak in a controlled run, not a sustained working speed under payload.

None of this makes the demo fake. It makes the demo a demo. Treat it as evidence the hardware can physically perform the motion — not as evidence the robot will do the job unattended on your floor.

Teleop-assisted is a feature, not a bug

Here is the part most buyers miss: teleoperation is not a fallback for robots that "can't do it yet." It is the mechanism by which they learn to.

Modern humanoid skills come from imitation learning, and imitation learning needs demonstrations — thousands of them, captured by humans driving the robot through the exact task, in the exact workspace, with the exact objects. Every teleop session is a data-collection session. The robot that a human pilots today is generating the training set for the semi-autonomous policy that runs tomorrow.

This is why dexterous hands and tactile sensing matter so much to the data flywheel, and why the spec sheets are worth reading closely. A hand like the Unitree Dex5-1 (20 DOF, 94 touch points†) or the PaXini DexH13 GEN2 (≈978 ITPU taxels at 0.01 N†) exists partly to capture rich contact data during teleop that a simpler gripper cannot. High-density skin such as Daimon Robotics' DM-Tac W (≈40,000 units/cm²†, POA) feeds the same pipeline. See /components/dexterous-hands/ and /components/tactile-force-sensors/ for the current catalog.

So when a vendor says "it's teleop-assisted," the correct read is not "it doesn't work." It's "it works with a human in the loop, and that loop is also the training loop." That is exactly the right architecture for a 2026 pilot whose real goal is to build a dataset and a business case, not to eliminate an operator on day one.

Matching the level to your use case

The autonomy level you need is a function of why you're buying, not which robot has the flashiest video. Three buyer profiles, three very different right answers.

Your situation	Autonomy level you actually need	Sensible hardware class (2026)	What "success" looks like
Research platform	Level 1–2, plus open SDK and joint-level access	Unitree G1 (23 DOF, from $16,000 US list^†) or R1 ($5,900 US list^†); Fourier GR-2 (53 DOF^†, POA) for whole-body work	You can teleop it, log data, train and deploy your own policies. Autonomy is your research output
Factory / logistics pilot	Level 3 on one narrow task, human-supervised	EngineAI PM01 ($12,000^†); AgiBot Yuanzheng A2 (project pricing, ≈$100–190k reported by variant^†) with a platform integrator	One task runs semi-autonomously in a fixed cell; a human handles exceptions; you're building the dataset and the ROI model
Demo / exhibition unit	Level 1 is fine — reliability and presence beat autonomy	Whatever looks and moves well on cue; scripted or teleop routines	It performs its routine every time on the show floor. Nobody cares that a human is driving

The mistake is buying a research platform expecting a deployment appliance, or budgeting a pilot as if level-4 autonomy shipped. A G1 is an outstanding platform because it's open and you can teleop and train on it — not because it will autonomously run your line out of the box. Conversely, a demo buyer overpaying for a 40-DOF research humanoid to wave at a trade-show crowd has bought capability they'll never expose.

For a fuller read on what the Chinese humanoid field can and can't do this year, see /guides/state-of-chinese-humanoid-robotics-2026/; for the buying mechanics, /guides/how-to-buy-robots-from-china-2026/.

Why your RFQ `use_case` decides everything

This is why our RFQ form asks whether you're buying for research, pilot, deployment, or demo before it asks anything else. It's not a marketing funnel — it changes the honest answer to every downstream question:

Research → we steer you to open-SDK platforms with joint-level control and teleop rigs, and we don't oversell autonomy you'd rather build yourself.
Pilot → we scope one narrow task, a fixed cell, and a data-collection plan, and we set the expectation that a human supervises. We'll tell you if your task is too open-ended for 2026 level-3 autonomy.
Deployment → we get specific and cautious. Fully unattended, general-purpose humanoid work is not a 2026 product; we'll say so rather than quote you a fantasy, and often propose a supervised pilot as the honest first step.
Demo → we optimise for reliability, appearance, and a scripted or teleop routine, and save you the cost of research-grade autonomy you won't use.

A quote written against the wrong use_case is how expectation gaps get baked into a purchase order. Tell us the real reason you're buying, and the spec, the price band, and the honesty of the autonomy claim all fall into place.