Robots Are Getting Beaten Up, And It’s Kind Of Funny

In partnership with

In between the protest videos, the crackdowns on journalists, and the running coverage of Jantar Mantar, my feed handed me a strange little comic break these last few weeks: robots, having a terrible time. Food delivery bots in American cities are being kicked over and spray-painted and left face-down on the pavement.

Earlier, at a sports meet in Beijing, robots caught fire, charged at bystanders and were carried off on stretchers — the robots, not the humans. At a tech show in Taipei, one collapsed mid-demo and had to be draped in a black tarp and hauled away. Honda's Asimo, the little robot that once jogged onstage for world leaders, was retired after twenty years of doing nothing useful.

I laughed at the machines that were meant to replace humans and at the breezy confidence of an industry that assumed we'd simply make room on the pavement for them. I noticed I wasn't the only one in the comments section of these videos. Nobody mourns a kicked delivery bot. We find the falling, the faceplanting, the "safe collapse sequence" genuinely funny, and I think the laughter is telling us something. We are not afraid of these robots. We are not even that impressed by them. We are watching something expensive and over-promised meet the ordinary friction of the world, and lose.

Then I got curious about the robots themselves. The ones rolling around Philadelphia, getting kicked outside cheesesteak shops, are made by a company called Avride — a spinout of the Russian tech giant Yandex's self-driving division, now headquartered in Austin under a Dutch parent after Yandex sold off its Russian businesses in 2024. 

Each little bot is worth around $10,000, rolls at about five miles an hour, and when it gets into trouble, say, stuck at a kerb, flipped over, surrounded by humans or walls, it calls a remote human supervisor at Avride who takes over and steers it to safety. The autonomous delivery robot has a person watching from somewhere far away, ready to grab the controls.

It is a small thing, but it points to a question we should probably be thinking about: are human-shaped robots going to make the world a better, more peaceful place? Do we even need them? And if we do, at what cost, trained on whose labour, built for whom?

To get at that, let me take a detour through television.

There is a scene in Pluribus, the Vince Gilligan show on Apple TV, where human-looking workers quietly restock a supermarket. No one is arguing, no one cutting in line, no one yelling at the self-checkout. They place items one by one, without any fuss and zero complaints. In the show, an alien virus has turned nearly all of humanity into a shared consciousness, a hive mind, where billions of people think and act as one. These people are still human, technically. They just don’t have selves anymore. 

It is efficient and unsettling in equal measure. 

Gilligan, who has been openly critical of AI, built the show around the question: what happens when we hand over our work to something that promises to do it better, and quieter, than we can?

That is also, more or less, the promise the humanoid robot industry has been selling this year.

Over 40,000 humanoid robots were shipped globally in the first half of 2026, 97 percent of them from China. Venture capital poured $2.37 billion into the sector in a single quarter, a 288 percent jump over the previous year. Tesla says its Optimus is already working inside its factories. A Norwegian startup is selling a $20,000 home robot that will apparently make your tea. In Shanghai, a factory is producing one humanoid every ten minutes. In Beijing, they sort parcels overnight for China Post and pour coffee at kiosks.

So, how does a machine learn to make tea in the first place?

The answer is that it watches a human do it. The industry calls the footage "egocentric video" — first-person recordings of hands making tea, mopping floors, folding laundry, every mundane thing you might one day want a robot to do. Feed a machine enough of this, from enough different kitchens, and it slowly learns how objects behave in the physical world. One system, GEN-1, reportedly needed half a million hours of real-world footage before its success rate climbed from 64 to 99 percent.

Half a million hours is a staggering amount of footage. And a lot of it is coming from India.

As Decode reported in September, an estimated 20,000 to 30,000 workers across the country train these robots. Many of them are women working from home, strapping iPhones to their foreheads and filming themselves doing household chores. The clips move through local vendors and data startups before landing in robotics labs in California and Norway. For every usable hour, the workers earn between Rs 50 and Rs 250. The robots they are training sell for $20,000 to $100,000

The asymmetry here is important to note. India is where a great deal of this training labour happens, but it is not where the robots are headed. 

The domestic market for humanoid robots is small — around $212 million and growing slowly — and the handful of Indian players are at the very beginning: a Kerala startup, iHUB, recently became the first Indian company picked for NVIDIA's humanoid programme; a Mumbai startup announced a single factory deployment this month. The machines being trained on footage from Rajasthan kitchens are built for warehouses in Texas and homes in California. 

The labour is cheap here; the robots are not for here. If they ever do turn up in Indian homes, they will arrive as expensive imports, trained, in part, on footage filmed by people who could never afford one.

A woman who records herself doing chores told my colleague Hera Rizwan that she was thinking of buying gardening tools with her own money so she could film herself gardening. A task that doesn't exist in her home, because the robots want footage of lawns and dishwashers in an American suburban life. She would be spending her own wages so a machine on another continent could learn to move through it and she could make another extra hundred rupees every week. 

So: are these robots as efficient as the eerily competent humans shelving groceries in Pluribus? The money says they ought to be. The robots themselves say otherwise.

Those falling, burning, face-planting machines aren't outtakes. They are roughly the state of the art on a bad day, and the bad days are frequent.

The people building these machines are oddly candid about the gap. The founder of Unitree, one of China's biggest humanoid makers, told the World Robot Conference that current robots run at 30 to 50 percent of human efficiency.

The reason they keep failing, as the roboticists call it, is the "sim-to-real gap." A robot trained inside a simulation: flat floors, predictable objects, nothing wet or wobbling falls apart when it meets the actual world. A Stanford study found robots managing 90 percent success in simulation dropped to 12 percent on real tasks. One Forbes estimate put the real-world failure rate at 88 percent. 

Training data is where the two halves of this story meet. Research on robot learning suggests robots only learn to handle new situations after seeing enough varied ones — which is exactly why companies are so hungry for footage from as many different homes as possible. Whether a kitchen in Bharatpur actually helps a robot destined for a kitchen in Fremont is an open question, and not one the industry seems in a hurry to answer. But it is a reasonable thing to wonder about a supply chain that films Indian homes to teach machines about American ones.

One pattern does hold across every robot that actually works today: it works only where nothing surprising can happen. Add a wet floor, a toddler, or a doorway the robot has never seen, and the failure rate takes over.

Which brings us to the question the industry would rather we not dwell on: do we actually need the human-shaped robot at all?

The case for it isn't silly. The world is built for humans — stairs, door handles, shelves at chest height — so a machine in our shape can slip into our buildings without anyone redesigning them. The labour shortage is also real: Bain estimates a global shortfall of nearly 8 million manufacturing workers by 2030, sharpened by aging populations in the very countries placing the orders. For dangerous, dull, hard-to-fill work, the argument holds together.

But some of the field's most serious people think the human shape is mostly theatre. Rodney Brooks, who co-founded iRobot and has built robots for forty years, argues that today's humanoids don't even have the sensors to match human hands — a single palm carries around 17,000 touch receptors no machine has come close to reproducing. When the robotics firm KUKA was asked why they don't build humanoids, they said their job is to solve real industrial problems, and there's "no need to reinvent the wheel." The successful robots are the boring, wheeled, distinctly non-human machines that have worked for years.

That Avride bot in Philadelphia calls a remote human when it gets stuck. That is the whole industry in miniature. The robot that looks autonomous has a person behind it, someone you are not meant to see, usually somewhere cheaper. In Philadelphia it is a supervisor in an Austin control room. In the longer story, it is a woman in Bharatpur with a phone strapped to her forehead, filming herself make tea for a machine she will never meet, in a home that looks nothing like the one the machine is built for.

So let me come back to where I started. Are human-shaped robots going to make the world better, calmer, kinder?

On the current evidence, the honest answer is that we don't yet know what they are even for. The ones that work do the narrowest slice of work in rooms arranged for them. The ones meant to walk among us fall over, and get kicked when they don't. And the entire edifice rests on an older arrangement we already know well: the value collects in California and Shenzhen, the labour is done, cheaply and invisibly, somewhere else.

In Pluribus, the hive keeps insisting it only wants to help. It restocks the shelves, it fixes the houses, it asks for nothing but your surrender. Carol's refusal is never about whether it works. It is about who decided, and on whose behalf, that this was the better world. The humanoid robot industry is making the same offer, with the same confidence, and nobody has quite asked us whether we wanted it, least of all the people whose hands are teaching it to move.

Maybe we will need these machines one day. Maybe they will even be good. But before the money decides it for us, the questions worth asking are: who is being asked to build the future with their hands, who gets to own it when it arrives, and what, exactly, are we being promised relief from — and at whose expense?

Like Carol, I'd rather keep asking than agree to be helped.

On My Bookmarks

A WhatsApp group for women's safety, recast as a terror plot. A Delhi PhD student made a group to help female students travel safely to the October 10 Jantar Mantar protest. By nightfall, screenshots were on prime-time TV, her number and dozens of others were public, the rape threats had started, and the police were at her door by 7am. Decode's Hera Rizwan traces how an ordinary act of solidarity became, overnight, the thing used to identify, expose and detain the women in it. Read it on Decode

A women's safety project, cleared for riot control. The Internet Freedom Foundation got hold of the Delhi Police's never-published drone SOP through an RTI. The 75 drones were built by students of a women's university and launched for "women's safety" — but the document that governs them lists "crowd monitoring and riot control" among their uses, hands the only big drone to the district that contains Jantar Mantar, and sets no rule on what gets recorded or how long it is kept. Read it on Internet Freedom Foundation

France leans on the platforms. As cost-of-living protests spread across French cities, the government has pressed Snapchat, TikTok and others to pull "sensitive" content, demanded the identities of users said to be organising unrest, and floated restricting internet access in some areas at night. Reminder that the pressure to police protest through private platforms isn't only an Indian story, it's becoming the default democratic reflex. Read it on WIRED

How Jennifer Aniston’s LolaVie brand grew sales 40% with CTV ads

For its first CTV campaign, Jennifer Aniston’s DTC haircare brand LolaVie had a few non-negotiables. The campaign had to be simple. It had to demonstrate measurable impact. And it had to be full-funnel.

LolaVie used Roku Ads Manager to test and optimize creatives — reaching millions of potential customers at all stages of their purchase journeys. Roku Ads Manager helped the brand convey LolaVie’s playful voice while helping drive omnichannel sales across both ecommerce and retail touchpoints.

The campaign included an Action Ad overlay that let viewers shop directly from their TVs by clicking OK on their Roku remote. This guided them to the website to buy LolaVie products.

Discover how Roku Ads Manager helped LolaVie drive big sales and customer growth with self-serve TV ads.

The DTC beauty category is crowded. To break through, Jennifer Aniston’s brand LolaVie, worked with Roku Ads Manager to easily set up, test, and optimize CTV ad creatives. The campaign helped drive a big lift in sales and customer growth, helping LolaVie break through in the crowded beauty category.

Learn more.

Got a story to share or something interesting from your social media feed? Drop me a line, and I might highlight it in my next newsletter.

Was this email forwarded to you? Subscribe