Tesla's Cybercab Hit the Road — Then Ran Its Wipers on a Sunny Day

The short version: Tesla's Cybercab — a two-seater built with no steering wheel and no pedals — started taking paying passengers in Austin, Texas in September 2026. Within a week, riders had filmed it running its wipers on a cloudless 95°F day, dropping people off in the middle of traffic, and hiding a virtual joystick in the touchscreen. Federal regulators opened an investigation. But the Cybercab wasn't the most important story that week. XPeng switched on a humanoid production line, Agility Robotics published an actual payback period, and Unitree's stock fell by half. Robotics is shifting from who films the best demo to who can show the math.

TL;DR
① Cybercab is Tesla's purpose-built autonomous two-seater — no steering wheel, no pedals. Paid rides started in parts of Austin, with 45 registered in Texas.
② The Verge's reporter waited about 45 minutes for a $15 ride under two miles. The wipers came on during a sunny, 95°F day.
③ Riders filmed Cybercabs stopping in busy roads without pulling to the curb. Others found a hidden virtual joystick on the center screen.
④ NHTSA is auditing Tesla's self-certification — in the U.S., automakers declare their own compliance and the government checks afterward.
⑤ Same week: XPeng opened a dedicated IRON humanoid line in Guangzhou, Agility said its next Digit could pay for itself in roughly 1.1 years, and Unitree shares fell ~53% from their first-day high. Those three together are the real story.

The source material here is The Rundown AI's Robotics newsletter from September 10, 2026. After reading the whole issue, I don't think the headline story is the most interesting one — so I'll finish it first, then explain why. Every number below links to its original source so you can check it yourself.

What actually happened with the Cybercab

First, what the Cybercab is. It's a car Tesla designed specifically to drive itself: two seats, no steering wheel, no pedals. Not a retrofit of an existing model — built from the start on the assumption that no human will ever touch the controls. Tesla wants it to become the workhorse of its robotaxi fleet.

In September 2026 it started carrying paying passengers in limited parts of Austin, with 45 registered in Texas. Less than a week in, the oddities started showing up:

• A 45-minute wait for a $15 ride under two miles. That's from The Verge reporter Mack DeGeurin's hands-on test.
• Wipers on a cloudless day. Same ride, 95°F and sunny, and the wipers switched on mid-trip.
• Drop-offs in live traffic. Several riders posted videos of Cybercabs stopping in busy roads to let passengers out, without pulling to the curb.
• A joystick hidden in the screen. Despite the car having no controls by design, riders found a virtual joystick on the center display with forward, reverse, stop, horn, and door buttons. To be precise: the reporting shows the interface exists but does not establish whether riders can actually drive the car with it, so this shouldn't be read as "passengers can take over."

NHTSA isn't investigating the wipers — it's investigating self-certification

Here's the part most coverage skips. U.S. vehicle safety law runs on self-certification. An automaker declares that its vehicle meets Federal Motor Vehicle Safety Standards and puts it on the road; the government audits after the fact. That's very different from the EU or Japan, where a vehicle type has to be approved before it can be sold.

The Cybercab reached public streets through that route. The complication is that a number of those federal standards assume the vehicle has things like a steering wheel, mirrors, and pedals — so a car deliberately built without them sits in a grey area.

NHTSA has now opened a case to examine the technical data and the process behind Tesla's self-certification. Ars Technica put it plainly: the car has been deployed and is already under investigation. Worth separating, though: this is a regulator reviewing data, not a finding of non-compliance.

That hidden joystick is sensitive for exactly this reason. If some command interface does exist in the car, the argument that "this vehicle needs no human control, so those standards don't apply" is harder to make cleanly. But that's an open question right now, not a conclusion — the investigation will decide whether it matters at all. Meanwhile Tesla is still expanding to more vehicles and more locations.

The other three stories from the same week

1. Boston Dynamics veterans are building robots you'd want to hug

Boston Dynamics is the company behind the backflipping humanoid and the Spot robot dog. Two of its veterans left to start Dynamic Creatures: former strategy chief Marc Theermann and former Atlas researcher Farbod Farshidian.

Their direction is close to the opposite of their old employer's. Boston Dynamics chases athletic capability. Dynamic Creatures is building mobile robot characters that roam theme parks, hotels, and entertainment venues and interact with the people they meet. In-house prototypes so far include Danielle, a poodle, and Hugo, a Tyrolean mountain troll.

In Fortune's interview, Theermann said something I keep coming back to: the company will likely never build a photorealistic robot. The reasoning isn't hard to follow — the closer you get to a real human, the deeper you fall into the uncanny valley.

Commercially, they're launching as Boston Dynamics' entertainment and hospitality partner, and say they're working with an unnamed large theme park and a retailer. Technically, they're building on existing platforms like Spot for now, which means all of the effort goes into natural movement and guest interaction rather than new hardware. Disney's free-roaming Star Wars BDX droids already proved the concept works; Dynamic Creatures wants to sell the same thing to other IP holders and venues.

2. XPeng opened a dedicated humanoid production line

XPeng is a Chinese EV maker. It just switched on a production line in Guangzhou dedicated to its IRON humanoid. The first unit assembled there walked off the line under its own power, and XPeng says over 80% of the line's processes are automated.

The timeline: mass production targeted for the end of 2026, with initial deployments in XPeng's own stores and campuses before commercial deliveries start in 2027. On specs, IRON has 76 degrees of freedom (21 in each hand alone) and runs three of XPeng's in-house Turing AI chips, delivering up to 2,250 TOPS.

Here's XPeng's own video on why IRON moves the way it does:

Video: XPENG official YouTube channel.

I don't think the specs are the point. The factory is the point. Humanoid startups tend to get stuck in the same place: building one or two units for a demo video is impressive, but delivering thousands is a different business entirely. XPeng already mass-produces cars, so the supply chain, in-house chips, automotive-grade quality control, and manufacturing know-how all carry over.

Funding isn't the constraint either. XPeng's robotics arm raised more than $900M at a valuation above $6.3B just weeks earlier, to take on Tesla, Unitree, and UBTech.

3. Amazon's Zoox is fighting Waymo with marketing

Amazon's Zoox also runs robotaxis, and like the Cybercab its vehicle was designed from scratch without a steering wheel or pedals — bidirectional, so it doesn't need to turn around.

In San Francisco, Waymo is beating it badly. The numbers are blunt:

• Zoox: roughly 100 robotaxis, about 10,000 U.S. rides a week.
• Waymo: roughly 4,000 vehicles, more than 500,000 fully autonomous trips a week.

That's about a 50x gap. So per The New York Times, Zoox changed tactics: stop competing on technology, start selling the experience. Influencer deals, ad campaigns, sponsored festivals, an SFMOMA partnership, and events at its Union Square Rider Lounge. Free rides are being used as marketing too, shuttling people from the lounge to eight restaurants and tourist stops like the Painted Ladies.

The goal is clear enough: make riding a Zoox something you try when you visit San Francisco. Waymo has already shown there's real demand for driverless rides, but the economics of robotaxis remain largely unproven. Zoox rides in San Francisco are still free, so whether attention converts into high utilization and paying customers is an open question.

Everything else in that issue

• Samsung is reportedly racing to unveil its first in-house general-purpose humanoid prototype at CES 2027, though it hasn't confirmed this.
• Lyft is matching riders with fully autonomous Waymo vehicles through its app in Nashville — the first city where the same Waymo fleet can be booked from two platforms.
• Unitree shares fell ~53% from their first-day high, wiping $35B from the peak valuation, though the stock is still more than 3x its IPO price.
• Tesla may soon offer Robotaxi rides to customers whose cars are being serviced, instead of loaners or Uber credits.
• Agility Robotics said its next-gen Digit humanoid could cost factories about $400K over five years while paying for itself in roughly 1.1 years.
• Luum, founded by an exoskeleton robotics pioneer, is rolling out robots that do eyelash extensions, already operating in some Ulta and Nordstrom locations.
• Trump's new 100% tariff on some imported drones is reportedly rattling police and firefighters who rely on cheaper Chinese models.
• UForce, a London drone startup built from Ukrainian battlefield manufacturers, is reportedly in talks to raise about $500M at a nearly $5B valuation.
• China is accelerating efforts to move humanoids from labs into military training, exploring reconnaissance, logistics, patrols, and combat roles.

What I actually took from this issue

Honestly, the Cybercab stumbling doesn't surprise me. Early autonomy always looks like this — Waymo got filmed frozen at intersections too. That kind of thing gets fixed.

What made me stop was three items landing in the same week: XPeng building a production line, Agility publishing a 1.1-year payback figure, and Unitree's stock falling by half.

They all point the same direction. That Agility number is one of the very few public humanoid ROI figures anyone has been willing to commit to, and publishing it means being willing to have customers check the math against real output. Unitree's drop is the other side of the same shift: the market is repricing companies that film well but haven't shown they can earn.

This industry used to compete on demo videos. Now it's being asked for a balance sheet. I have no idea who can actually produce one, which is exactly why I think the next year is worth watching.

Why this matters if you don't work in robotics

The most useful thing in this issue, for me, wasn't about robots at all. It was the gap between "the demo looks great" and "it delivers reliably" — which is the same gap you hit with AI tools every single month.

A new tool ships with a beautifully edited launch video, and you walk away thinking it could replace half your workflow. Then you run it on your own client work and find the specific place it falls apart. The Cybercab runs its wipers on a sunny day; an AI tool hits the edge case it was never built for, on the project you actually have to deliver.

My own rule: after watching the demo, run the tool on one small job you genuinely have to ship before letting it into your workflow. If it survives one real delivery, it's worth learning properly. The tools I actually use and that hold up under deadline are collected in the Zeona toolbox.

There's a second angle here too. If you publish content, how stories like this get cited by AI search follows patterns you can actually work with — I broke that down in my complete guide to GEO (generative engine optimization), plus a case study with real results.

FAQ

Q1: What is the Cybercab, and how is it different from the Model Y robotaxis Tesla already runs?

The Cybercab is Tesla's purpose-built autonomous vehicle, with no steering wheel and no pedals. The robotaxis Tesla has been running in Austin are modified Model Ys that still have full human controls and safety personnel on board. The Cybercab is a two-seater designed on the assumption that no human will drive it, and Tesla intends it to be the workhorse of its robotaxi fleet.

Q2: Why are regulators investigating Tesla? It just turned on the wipers.

The investigation is about self-certification, not the wipers. U.S. law lets automakers declare their own compliance with federal safety standards and put vehicles on the road, with the government auditing afterward. NHTSA is examining whether the technical data and process behind Tesla's declaration hold up — particularly for a car deliberately built without the equipment several of those standards assume. The hidden virtual joystick riders found complicates it, because it suggests some command interface may exist in the vehicle — though the reporting doesn't establish whether riders can actually drive with it. And an investigation means a regulator is looking, not that Tesla has been found in violation.

Q3: How does the Cybercab differ from Waymo?

The main differences are sensor strategy and where the vehicle comes from. Waymo uses modified production cars equipped with LiDAR and other redundant sensors, running roughly 4,000 vehicles and more than 500,000 trips a week in 2026. Tesla runs a vision-only approach and builds a dedicated vehicle for the role. The scale gap is large right now: only 45 Cybercabs are registered in Texas.

Q4: Are humanoid robots really entering mass production in 2026?

A production line exists, but "can be built" and "can be sold" are still separate questions. XPeng's IRON has a dedicated line with mass production targeted for end of 2026 and commercial deliveries in 2027, and Agility Robotics has published cost and payback estimates (about $400K over five years, roughly 1.1 years to pay for itself). At the same time Unitree's stock fell ~53% from its high, which suggests the market is repricing the category. Actual delivery volumes and repeat orders are the things to watch.

Q5: Where can I read the original reporting?

The source is The Rundown AI's Robotics newsletter from September 10, 2026 — you can read the original issue or subscribe. Every item in this post links to the primary reporting it drew on (The Verge, Ars Technica, Fortune, The New York Times, Business Insider, and others), and those are worth reading first-hand.

Sources and image credits

This article is a summary and commentary based on The Rundown AI, "Tesla's Cybercab is already getting weird" (Robotics newsletter, September 10, 2026), written by Jennifer Mossalgue, Rowan Cheung, Shubham Sharma, and Zach Mink. In-article images are copyright Dynamic Creatures, XPeng, and Zoox respectively, as published in that issue; the cover image is copyright Tesla. The Cybercab photo is by Avda via Wikimedia Commons, licensed CC BY-SA 4.0. Video is copyright the XPENG official YouTube channel.

Primary reporting for each item:
· Cybercab hands-on: The Verge
· Hidden virtual joystick: @MatthewSkrzypc1 on X
· NHTSA investigation: NHTSA press release, Ars Technica
· Dynamic Creatures: company site, Fortune
· XPeng IRON line: XPeng newsroom, XPENG YouTube
· Zoox marketing push: The New York Times, Zoox Journal, Fortune

If you want to talk through how AI tools fit into your actual workflow, or need help making your content and site visible to AI search, get in touch.