The $18,000 Humanoid Has a Human Inside: Astribot’s U.S. Launch, Nucleus’s 40% Teleoperation Reveal and 10 Questions Before a Robot Joins Your Team (AI Trends, 3 October 2026)

What’s trending in AI on 3 October 2026: humanoid robots stopped being a keynote fantasy and started being a purchase order. At the IROS 2026 robotics conference in Pittsburgh, Chinese firm Astribot brought its T1 humanoid to North America at $18,000, available to order now. Two days earlier, European startup Nucleus posted nearly two hours of uncut factory footage and disclosed something most robot makers avoid saying out loud: in that demo its humanoids were 60% autonomous and 40% teleoperated by people. And Tesla said it halved the memory on its next AI chip because that was the only way to get enough supply to build Optimus robots in volume. Put together, these stories answer a question every business buyer should ask: when a humanoid robot shows up at your site, who is really driving it, and what does it send home? This guide covers what launched, what the numbers say, the privacy and security catch, and ten questions to ask any vendor before a robot joins your team.

Key takeaways

Humanoids now cost about as much as a used car. Astribot’s T1 starts at $18,000 in the U.S., and 1X’s NEO home robot is priced at $20,000 or $499 a month.
Many “autonomous” robots still have a human inside the loop. Nucleus says its factory demo was 40% teleoperated; Astribot’s T1 ships with VR teleoperation; 1X sends a remote operator when NEO can’t do a task.
That’s a feature, not a scandal. Human corrections are the training data that make robots more capable, the same playbook self-driving cars used. The problem is when buyers don’t know it’s happening.
A teleoperated robot is a remote-access device with arms. It carries cameras, depth sensors and microphones, connects to a vendor’s cloud and can be steered by someone you have never met.
China builds almost all of them. Chinese firms made about 97% of humanoid deliveries in the first half of 2026, which raises data-residency and supply-chain questions for Western buyers.
The real robot workforce is still industrial arms. About 500,000 industrial robots were installed in 2025 versus roughly 7,000 humanoids sold for professional use.
For your business: treat a humanoid like a privileged remote user plus a camera system. Ask who can operate it, what it records, where the data goes and how to cut it off.

Humanoid robots got cheaper and more honest this week. Many still rely on a person somewhere else to finish the job.

1. The week humanoids got a price tag

Astribot T1, $18,000. Shenzhen-based Astribot used IROS 2026, held in Pittsburgh from 27 September to 1 October, to bring its T1 to North American buyers. The robot stands about 1.55 meters tall, weighs around 66 kg, has 23 degrees of freedom (not counting its hands or grippers) and lifts up to 5 kg per arm. It uses a cable-driven design meant to make its movements more compliant around people. It runs Astribot’s Lumo-2 foundation model, which the company has shown packing assorted household items into a backpack on its own, and it ships with an SDK, APIs and a natural-language tool that turns a text description into a robot app. U.S. customers can order now. It is positioned for household tasks and as a developer platform, not as a finished factory worker.

Nucleus, paid by the hour. Nucleus, which came out of stealth in August after putting humanoids into a German factory within 90 days, does not sell robots at all. It charges customers by the hour for completed physical work, using modified Unitree G1 humanoids. On 1 October it posted almost two hours of uncut footage of its robots picking parts, loading shelves and moving carts, hesitations and human interventions included. CEO Melvin Schwarz joked that the team “worked very hard to make it this boring.”

The rest of the field. 1X’s NEO home humanoid is priced at $20,000 up front or $499 a month, with deliveries targeted for 2026 from a factory in Hayward, California that opened in May. The U.S. Army has its eye on Alex, an armored humanoid from the Florida Institute for Human and Machine Cognition that won the Army’s xTechHumanoid competition and uses a hybrid of AI and human control. And Elon Musk said Tesla cut the memory on its upcoming AI5 chip from 144 GB to 72 GB, and on AI6 from 216 GB to 144 GB, because it was the only way to secure enough chips for Optimus production. That last detail ties humanoids directly to the AI-driven memory shortage squeezing everything from phones to servers.

2. The 40% nobody used to mention

For years, humanoid demos have been short, edited clips. Viewers rarely knew whether a robot was acting on its own or being puppeted by a person in a motion-capture suit or VR headset nearby. Nucleus broke that pattern. By its own count, the robots in its two-hour video worked autonomously about 60% of the time, and a human teleoperator handled the other 40%. The company stresses that the figure applies to that demonstration only, and that its robots now routinely run four-to-six-hour shifts and finish jobs end to end.

Nucleus is not alone; it is just unusually candid. Astribot’s T1 includes VR-based teleoperation and feeds data from both autonomous work and teleoperation back into central model retraining. 1X describes an “expert mode” in which a 1X employee remotely guides NEO through tasks it hasn’t learned, with the owner’s approval. Reviewers have estimated NEO’s autonomy at launch in a similar range to Nucleus’s demo. Even the Army’s Alex is described as “hybrid autonomy.”

Only one vendor has published a hard number. For the rest, the share of human control is something you have to ask about.

3. Why the human is there, and why that’s not a scandal

It would be easy to read “40% teleoperated” as “40% fake.” That misses how robots learn. A humanoid working in a real warehouse or kitchen meets thousands of situations no simulator anticipated. When it gets stuck, a remote operator takes over, finishes the task and, in doing so, records exactly the demonstration the AI model needs. Nucleus says its system logs camera images, depth, robot motion, contact signals, human corrections and the final outcome together, so every intervention becomes training data. 1X and Nucleus both compare this to how self-driving companies used remote assistance and safety drivers to build up the data that eventually let cars drive themselves.

There is also a hard economic reason. A robotics executive quoted by Interesting Engineering described the “90-10 problem”: simulation can get a robot to roughly 90% performance, but closing the final 10% eats most of the budget. Deliberately making an expensive robot fall down stairs to collect data on rare accidents isn’t practical, so real-world human help fills the gap. The honest version of this business model is the one Nucleus chose: say how much help the robot needs, show the footage, and charge for finished work rather than promising autonomy you don’t have yet.

The risk is not teleoperation itself. It is undisclosed teleoperation. If you don’t know a person can see through your robot’s cameras and move its arms, you can’t decide who that person should be, what they are allowed to see, or what happens when their account is compromised.

4. The numbers: hype versus the robots already at work

Humanoids are growing fast from a tiny base. The International Federation of Robotics counted about 7,000 humanoids sold for industrial and professional service use in 2025, against roughly 500,000 conventional industrial robots installed that year and about 5 million robots now working in factories worldwide. Using a broader definition, a report cited by Euronews counted 19,100 humanoids delivered in the first half of 2026 alone, with about 60,000 expected for the full year and 500,000 a year by 2030. Chinese makers accounted for about 97% of first-half deliveries, led by Agibot with roughly 8,400 units.

Industrial arms still outnumber humanoids by about 70 to 1. The forecasts assume that gap closes quickly.

The hardware is the hard part. A humanoid needs more than 40 joints, each with its own motor, gearbox, bearing and sensor, compared with about seven for a typical industrial arm. German supplier Schaeffler estimates actuators make up about half of a humanoid’s cost, and European researchers told Euronews that some low-cost components wear out in a week or two of heavy use. That matters for buyers: an $18,000 sticker price says little about the cost of keeping a robot running for three years. Ask for maintenance terms, expected component life and spare-part lead times before you sign.

There is also a counterpoint worth taking seriously. Kumar Sokka, CEO of Acre Security, argues that the real machine workforce already exists in the form of industrial robots, and that its biggest risk is thousands of machines with standing access that nobody actively manages. University of Pennsylvania roboticist George Pappas offers the balance: optimize the industrial robots you have, but place strategic bets on robots built to work in spaces designed for people, such as homes, hospitals and stores.

5. The security and privacy catch: a remote-access device with arms

From a security point of view, a teleoperated humanoid combines three things businesses already struggle to govern: a camera and microphone system, a cloud-connected IoT device that receives over-the-air updates, and a remote-access session controlled by an outside party. Then it adds something new: physical force. A hijacked laptop can leak data. A hijacked robot can also open a door, move inventory or hurt someone.

The robot is the visible part. The data pipeline and the remote operator are where most of the risk sits.

Some vendors are building sensible controls. 1X says NEO’s teleoperators need the owner’s explicit approval, faces are blurred by default, owners can set no-go zones and streams are encrypted. Astribot says customers can choose whether to contribute data from their robot’s work to model training, and that the T1 does not learn on its own in the field; updates come from central retraining, delivered over the air or on site. Those are good starting points, but they are vendor promises, not audited standards. We saw this summer how quickly a “connected camera on a person” became a workplace policy problem with AI smart glasses; a camera on a walking robot raises the same questions with more reach.

Risk	How it shows up with humanoids	Control to ask for
Unknown remote viewers	Teleoperators see live video of your floor, staff, screens and paperwork	Named, vetted operators; per-session approval; session logs and recordings you can review
Training-data leakage	Demonstrations of your processes and layouts flow into the vendor’s shared models	Training opt-out by default; data residency and retention terms in the contract
Hijacked control channel	Stolen operator credentials or a vulnerable cloud let an outsider steer the robot	Phishing-resistant sign-in for operators; isolated network segment; local emergency stop that works offline
Malicious or faulty updates	An over-the-air update changes behavior across a whole fleet at once	Signed updates, staged rollout, the ability to defer or roll back
Physical safety	A robot near people can drop loads, collide or be driven unsafely	Defined work zones, speed and force limits, a safety assessment before go-live
Supply-chain and jurisdiction	Hardware, firmware and cloud may sit in a country whose laws allow government data access	Clarity on where data is processed; a review against your regulatory and customer obligations

Six risks of teleoperated humanoids and the controls to ask for. Sources: vendor disclosures from 1X, Astribot and Nucleus; Delana analysis.

Two parts of this should feel familiar. The remote operator is a non-human-identity problem with a human attached, the same issue we flagged when AI agents started getting employee IDs: every account that can act inside your business needs an owner, least privilege and a kill switch. And the fleet-wide update channel is a software supply chain, the same weak point attackers abused in the incidents covered in this morning’s look at Microsoft’s 2026 Digital Defense Report.

6. Where humanoids make sense today, and where they don’t

Good fit now: repetitive pick, place and carry work in spaces built for people, where a fixed robot arm would need costly re-tooling; pilots paid per hour of completed work; research and developer teams exploring physical AI.
Wait and watch: customer-facing roles, homes with children or vulnerable residents, anywhere sensitive documents or screens are in view, and any task where a single mistake is dangerous or expensive.
Usually better solved another way: high-volume, fixed-layout tasks that conventional industrial robots, conveyors or cobots already do faster and cheaper.

If you are weighing a pilot, the pay-per-hour model has a big advantage: the vendor carries the cost of the robot’s mistakes and of the human operators behind it, and you pay only for work that gets done. It also makes the true level of autonomy visible in the invoice. For broader context on where AI agents already beat traditional automation, see our guide to agentic AI versus traditional automation.

10 questions to ask before a humanoid joins your team

Get the answers in the contract, not the sales demo.

What share of the work will a human do? Ask for the autonomy rate on tasks like yours, how it is measured and how it has changed over the last six months. “Fully autonomous” without a number is a red flag.
Who are the teleoperators? Employees or contractors, in which countries, with what background checks, training and supervision.
How is each remote session approved and logged? You want per-session approval or clear rules, a visible indicator on the robot when a person is in control, and logs or recordings you can audit.
What does the robot record, and what is kept? Video, depth, audio, maps of your facility. Ask about retention periods, deletion on request and what happens to the data if you cancel.
Does our data train shared models? Insist on an opt-out, ideally as the default, and on contract language that your process demonstrations won’t be used to train robots for competitors.
Where is data processed and stored? Get the countries in writing and check them against your privacy, customer and regulatory obligations.
How are operator and admin accounts protected? Phishing-resistant sign-in, least privilege and fast revocation. A stolen teleoperator login is a stolen pair of hands.
Can we isolate it and stop it locally? The robot should sit on its own network segment, and an emergency stop must work even if the internet or vendor cloud is down.
How are updates signed, tested and rolled out? Ask whether you can defer or roll back an over-the-air update, and how the vendor protects its build and update pipeline.
What does it cost to keep running? Maintenance contracts, expected component life, spare-part lead times and insurance. Compare buying against paying per hour of completed work.

If you already run cameras, IoT sensors or AI tools that see your workplace, fold humanoids into the same governance rather than inventing a new process. Our guides to multimodal AI for SMBs and AI agent security cover the consent, privacy and access rules that apply here too.

Frequently asked questions

How much does a humanoid robot cost in 2026?

Prices have fallen into car territory. Astribot’s T1 starts at $18,000 in the U.S. and is available to order now, and 1X’s NEO home robot is priced at $20,000 or $499 a month. Some industrial providers, such as Nucleus, don’t sell robots at all and instead charge by the hour for completed work. Ongoing maintenance, parts and insurance can add significantly to the sticker price.

Are humanoid robots really autonomous?

Often only partly. Nucleus said its humanoids were about 60% autonomous and 40% teleoperated in a two-hour uncut factory video published on 1 October 2026. Astribot’s T1 includes VR teleoperation, and 1X uses remote human operators for tasks its NEO robot hasn’t learned. Human help is used to finish tasks and to collect the training data that makes robots more capable over time.

What is teleoperation in robotics?

Teleoperation means a person controls a robot remotely, usually by seeing through its cameras and steering it with a VR headset, motion controllers or a computer interface. Robot companies use it to complete tasks the AI can’t yet handle and record those demonstrations as training data, similar to how self-driving car companies used remote assistance during early deployments.

Are humanoid robots a privacy or security risk?

They can be. A teleoperated humanoid has cameras, depth sensors and often microphones, connects to a vendor’s cloud, receives over-the-air updates and can be controlled by a remote person. Businesses should ask who the operators are, how sessions are approved and logged, what data is recorded and where it is stored, whether it trains shared models, and whether the robot can be isolated and stopped locally.

Who makes most humanoid robots?

Chinese companies. According to figures reported by Euronews, Chinese makers accounted for about 97% of the 19,100 humanoid robots delivered worldwide in the first half of 2026, with Agibot alone delivering about 8,400. European firms such as Schaeffler and Bosch supply many of the precision parts, such as actuators, which make up roughly half of a humanoid’s cost.

Why did Tesla cut memory on its AI5 chip?

Elon Musk said Tesla reduced AI5 memory from 144 GB to 72 GB, and AI6 from 216 GB to 144 GB, because it was the only way to get enough memory supply for Optimus humanoid production while cutting costs. He said memory bandwidth is unchanged, so performance should barely suffer. The move reflects the wider AI-driven memory shortage.