Daybreak - A new class of gig-workers in India are teaching robots to do the dishes

Episode Date: June 25, 2026

Meet Ranjan. He works at Deloitte by day and spends his evenings strapping a camera to his forehead, recording himself doing household chores by evening. He's a physical AI trainer, a part of a growing gig economy built around creating training data to teach humanoid robots human behaviour.Reporter Sakshi Sadashiv joins host Rachel Varghese to break down how this supply chain works: from gig workers in Delhi, to the firms like Tesla and Nvidia that eventually buy their footage from the companies that vet, annotate and package it for sale. And why this feels like a pattern India has been in before.*Read Sakshi's original story here.*Take the Daybreak listener survey here.

So before we start today, we have been thinking a lot lately about this show, about what it is, what it could be, and about you, the people who show up for it every single morning. And we realized, we don't actually know that much about you. So we have made this survey. It takes about three minutes and we are genuinely asking what's working, what isn't and what you want more. off. The link is in the show notes and we promise to read every single response. Okay, let's get into it. Ranjan is a 20-something employee at Deloitte. But unlike most of his co-workers, his work date doesn't end at 5pm. Instead, when he gets home at 6.30, he showers, drinks coffee and then gets down to business. He straps the camera to his forehead and starts recording as he goes about his daily chores,

washing and folding laundry, cooking, doing the dishes. Once he begins though, he moves slower than usual. His movements are deliberate. His hands never leave the frame. He can't record faces, idle for longer than three seconds, or shoot in slow motion. These are the conditions of Ranjan's new gig. He is a physical AI trainer and for every task he does, he makes about 250 to 300 rupees. On good evenings, he even makes 1000. He is part of a new class of gig workers who freelance for data collection companies like Micro One, Ego Data, Human Labs and more to help train robots in human behavior. You may remember Pronto, the Bangalore-based home services startup that made headlines in May for allegedly sending workers equipped with cameras into customers' homes

so that they could record similar footage for physical AI training. Of course, Pronto's competitors like Urban Company and Snabit rushed to reassure their own customers that they would not be doing the same. But backlash aside, it's important to note that Pronto did not create this market. It was already there. In fact, companies like Tesla, figure AI and Nvidia, all building humanoid robots, by the way, had been purchasing footage made by workers like Ranjan long before the Pronto controversy ever made the news.

The thing is, this kind of labour supply pattern is not new to India. We saw it first with IT services, then with the BPO or business process outsourcing model. Each wave has brought the country employment and the opportunity to integrate into the global technology stack. Each wave has left India supplying the labour while others captured the intellectual property. To look at this wave more closely, I have with me in the studio my colleague Saakshi Sadashiv. She has spent days reporting on the story, speaking to recruiters and workers across the country, studying listings sprinkled across sites like Indeed and Nockery.com. What she found suggests that this market is only picking up steam.

Which means this is the question. Is physical AI training just the latest version of that same old Faustian bargain? Hi, Saksh. Thank you so much for joining us in the studio today. I know this is your first daybreak episode with us. So before we get into the questions, would you give us a brief introduction about you? Of course. Thank you for having me.

My name is Sakshi. I am a reporter with the Ken. I have been here for close to three months and I work from the first. Great. Thank you so much for that. So I wanted to start the story with where you start the story actually, where you talk about Ranjan, who is this physical AI trainer and for whom this is a side gig. Like he has a day job at Deloitte. So could you describe what his day looks like when he's doing the training? Of course. So Ranjan works full time at one of the big fours and does this after work. He straps an iPhone or. any sort of body camera basically to his chest and opens a task list. This task list has a set of instructions which means that your phone has to be angled at a certain, which I'll get into a bit later. The phone has to be angled in a certain way. Your hands, which are the most important thing in this process, have to be visible at all times. You cannot slow yourself, you cannot

not like fasten your speed at any point. And he sort of records himself. The speed has to be like consistent throughout. Like you can't slow down or speed up. Yeah, yeah. So you have to have like a consistent amount of speed. Oh. And they will sometimes have like certain tasks that they would want you to do a little bit more fast or a little bit more slow.

Right. Which I'll get into Y variation is important. And he sort of just records himself doing chores. So that is your folding clothes. It could be making tea. It could be loading a dishwasher. Making tea. Yeah.

It could be opening cupboards. It could be, you know, picking up something and putting it down and things like that. The most important part of this is his fingers and how they move. The emphasis is on first person perspective. And the reason that is is because internet, which is optimized for engagement, doesn't necessarily have a lot of first person POV videos. You will have a lot of face, which is what internet rewards. But first person POV is not something that is very readily available.

So just for context, if a chat GPT had like lakhs of years of reading to scrape from, robotic companies are working with not even half or even. or even less than half of that data. They're working with almost like one-fourth of that data. And out of that one-fourth, you don't have even half of that being first-person POV. So they're really, really scrambling to have more first-person POV data. And how Ranjan helps in this entire supply chain

is that after recording, he'll upload that footage, which then goes through like quality checks and annotations. and it's sort of is a very surprisingly methodical for something that looks so ordinary and annotators then sort of figure out how to make sure that like you are able to

tag spatial awareness you are able to trace heat maps on the fingers and then that footage just sort of gets sold to bigger companies which is your Tesla your figure AI and others So, yeah. Okay.

So I wanted to pick up on what you said about how robotics doesn't have the same amount of data that, you know, like a generative AI LLM does. And you also refer in your article to the fact that robotics was an unglamorous cousin to agentic AI. Yeah. So how exactly did the gen AI boom change that and, you know, give birth to this new trainee job? yeah okay uh so robotics was always looked at with this lens of being an unglamorous cousin of agentic AI is primarily also a lot to do with how investors looked at it it is much more difficult to understand it's extremely hardware oriented right um which is what robotics signifies everything about right like hardware in general is something that is much more

difficult to scale. You have a lot of CAPEX that goes into creating something like that to it. And that has always been one thing. The other thing has been that, which is

a similar case in deep tech as well, that investors are now sort of picking up on understanding deep tech. It has sort of gone beyond that lens of being a purely optic-driven industry to more like understanding that deep tech is something that treats patient

capital. And that has sort of showed that like once LLMs showed that like foundational models could sort of generalize across many tasks across the world, companies sort of started to explore the idea. Right. If that could apply to the physical world as well. Like, I'm sure you and I both know a lot of people who use LLMs for very basic tasks as well, right? I could ask it to give me a shopping list.

Think of it as like a similar concept that people could replicate with the physical world, at least the at least the curiosity for it. That if we can replicate it into the physical world as well. And I think that is exactly where physical AI comes in. I think it has now moved from like not just being hardware, but it is now sort of moving into like, how smart can it be? Right.

How smart can a robot be? Can a robot understand that this is a hurdle and move past it without you being there? Yeah, not like a Roomba that keeps. Yeah. Like keeps like, you know, clanking into things. And one of the things that somebody had told me when I was reporting on the story, is one of the most basic things you have to solve as a robotic company is that robots are stupid and they keep clanking into each other.

So they don't have spatial awareness. They don't understand grip. That was very fascinating to me to understand that they're obviously going to be stronger than humans. So grip strength is something that is so innate to us. You understand. like I could look at something and I could understand that like

oh this is going to be heavy for me right exactly you know or if I like squeeze it it's going to break yeah right that is not something that robots can understand so they're not that smart yet which is where

a person like Ranjan helps you to understand how do you navigate spaces and Indian homes are tiny So nothing like spatial awareness than an Indian home because you are moving through so many things like I remember in the story when I was reporting on it one of the most important things is when he was folding clothes

he was sort of smootening out creases and that creases comes through this way of like you have to lift certain fingers for that crease to be let go And that is not something robots would understand, but they can only see and understand and sort of replicate how humans navigate the world. That's super interesting. So are there any examples of companies that have succeeded in training robots? Like, are we seeing success with this kind of a training model? Yes, it's still very early, I will say so.

But you have a lot of players in the ecosystem. But the amount of manufacturing intellect that you need to be able to scale a robotics company is something that China has been able to replicate, has had, not replicate, but has seen very good success in. There was this video that was trending about like the robots doing martial art shows. Yeah, like martial art, you will see like a lot of Chinese robots do backflips and martial arts and. So many other things that you would think that, like, I don't think I can do a backflip even. So, yeah.

But even in the U.S., you have a lot of Silicon Valley giants also sort of focusing on it. Elon Musk had said that, like, he would want, he thinks that robotics is going to be a much more significant part of revenue moving forward, which is where Tesla is also looking at. So you have companies like Tesla, you have agility robotics, which is backed by Amazon. You have Figure AI, another startup who has done very well. And they're sort of all collecting and using like real world, like behavioral data. These are also subsequently the biggest buyers of the data that India produces. I wanted to go back to something you mentioned that we would get to later, which is basically about how it's important to have variation in the way that you do tasks.

And part of it is what we already talked about, being the grip strength, the way that your fingers move, and just how specific some tasks can get in, because obviously the real world is not, you know, small sailing all the time. So, yeah, could you explain, like, what does that look like in practice? Yeah. So variation is the most important thing. for a data collection company. Right. And that has been a data collection trend

for as long as data collection existed. And you see this like keep coming up as like an LLM conversation of like how smart are they really. A similar function. One is the dearth of data. They genuinely don't have enough data. Right.

And two is that within that data, the quality of the data is supremely important. Across all the data collection workers that I spoke to, when they first started out, their videos would constantly get rejected because they were not doing it properly. Even when I tried my hands on doing it,

I didn't do it properly. Is there like a constant feedback loop also over here? Like you send it to your recruiter and then they give you... Yes, so some companies have daily videos that you have to send them. Right. Some companies will ask you to send them weekly. But coming back to variation, the most important way to look at variation has been that

one is the speed of it. One is, of course, that would constitute as one part of the variation where you are sort of doing it more slowly, more deliberately and a little bit more hurriedly. But of course, they prefer the slower the better because a robot can't exactly make out everything from a fast video. Right. Right. Like as slow as you can.

And then just sort of like speed it up to a certain point. The other variation. Is this so that robots can also understand speed? Like ideally you would not want a robot to be doing their chores slowly. Yeah, yeah, yeah. But you would want them to have like an understanding of like what exactly. what exactly like say for example i'm washing dishes right uh i will go into the crevices in a certain way i will fold my hand over the over the vessel in a certain way and then when i'm washing it i will make

sure i don't spill water on me you don't want a robot to do that either yeah right uh so you will sort of like that that sort of movement is very important to understand um that That is one type of variation. The other type of variation is actually very interesting is that one of the person when I spoke for this story has a washing machine. So he will record himself putting the clothes into the washing machine. And as everybody does, they will just do like the pause and start button. And that's about it. Like that's all the data collection that he has to do for the washing machine bit.

But one of the other person that I spoke to in the story is. Priti, she washes her clothes by hand. And that is gold for data collection company. Because it is such a laborous activity that you will not only be training, pressing buttons is one of the easiest things robots can learn. Right. You know. But the laborous task of washing something by hand, like clothes, especially because you

will tilt them, you will beat them up, you will put soap on it, you will put it under the water, you will like wrangle it. Wrangling is a very important task because it's not something that you can like trace it completely. Right, right, right. But it is so much force. Right. Right.

Like you only stop wrangling either because you don't have enough strength to wrangle it more or because you will break it. Right. So it's such an important task. And wrangling is one of those tasks that you don't have enough of. Pressing buttons is very easy. Picking up a glass and putting it down, it's very easy task. Like I remember she had told me that like she had earned a lot of money out of that one specific task because it's such a laborious task.

And a lot of money, I mean you are like 900 rupees. I mean, that's a lot for it. But for a task, it was like a lot of money. That's sort of a variation of hand tasks that, like, you don't get on the regular. So that is why it becomes so important for them to be able to do that and get that data collection. Oh, you also mentioned, like, that one of the problems is that they have, like, the homes, these trainers live in are small and, you know, they have limited space. And I'm quoting you here. they have limited shows and finite imagination.

So I'm curious as to why you mention imagination here. So I remember one of the person I was speaking to for the story had said that we, like he has a really tiny home. And he was basically saying that like these people keep asking for variation. And what am I supposed to do in my one room kitchen? You know? So variation, which is where the finite imagination comes from. Can I do this task differently?

Right. Can I fill my water bottle in a different way? That's also a task that you get paid to do. Yes. So, like, that's what I'm saying, right? How exactly you have limited space already and these companies don't want you to do the same task again and again. Like, you can do folding clothes. but you have to do it differently this time.

Right. But how exactly do you fold quotes differently? Right. Right. Like how exactly, like you don't really have, this is not a creative work. But as all employers are, they're asking for creativity here. And how exactly would you do creativity in a situation like that, right? And that is where he, so he, like, when he had started out,

he would have like a lot of a lot of tasks to do but now that he has reached a point where he doesn't like six months into what he has been doing he sort of spends like these 30 40 minutes just figuring out that I'm a lot of what I do you know okay so one thing that I wanted to go back to that you mentioned in your article is about how a lot of companies that also rely on gig work like the whole thing that happened with Pronto and also Doodash in the US and a couple of other companies in India. How exactly are they jumping in on this trend and what are they trying to, you know, gain from this? All of these companies are tech companies. So to want to collect data and utilize that data is in their DNA.

Right. Right. And one of the things is also that. that I know that Pronto and all came about to be, like, DoDash I know has been that. But for example, Uber India does this very interesting thing where they, if you are a driver with Uber and if you are idle, that means you haven't accepted any ride. Oh, right. Or if you are just sort of like, like just idle, but you're active on the app itself. you can annotate data for them.

A way to do that is that they ask drivers to annotate the roads. As in, is there a pothole here? Is there streetlights here? Is there like a right turn here or a left turn here? Of course, this is a way of crowdsourcing, which is how also Google Maps work. That Google Map also, like, you must have seen that, like, Yeah, yeah.

Is this a traffic area or whatever? These are very common mapping technologies. So it's a way to monetize that time. And I know Uber USA, like the global firm, sells this annotation data to marketing agencies. Right, okay. I'm not sure if Uber India does. Okay. But I know that Uber globally does like incentivize their drivers.

And then that like that data ends up getting so. to marketing agencies because, say, for example, if you have a really pretty billboard somewhere and the driver is like, oh, it's right next to this luxury apartments, the marketing agency can put up, I don't know, a spa advertisement there. And I also think it's important for them to do it. Yeah, right? Like, I would play the devil's advocate here and say that, like, you, if a competitor does it before, you market share is gone. So this brings me to my final question because we're also running out of time. Pretty early on in the story, Sarkshi, you describe this kind of a new gig work that's coming up.

And then you also compare it to the IT and BPO wave that we saw in India, where India makes considerable wealth through this kind of a labor. But the higher intelligence of it, so to say, is built elsewhere. Could you explain what you mean by that? Yeah, so the reason I would make that comparison, and I would argue that it's a comparison to be made, is that India is sort of once again supplying a very critical input into like a technology boom. And that critical input for any global technology boom that we see, right, we always sort of end up producing the labour and producing. And I would argue that that labor, without that labor, none of these companies can stand.

Exactly. If you do not have the labor that an Indian, like, produces with that sort of data, you wouldn't be able to train these robotics at all, right? Like, you won't make them intelligent. And that has something that has always been in India, right? in IT for that matter it was services even in DPO it was labor right and in generative AI it became annotation and again data collection and that now is like behavioral data for robotics and the difference is that while India generates the inputs the intellectual property which I would argue is the most important part of it still is

largely for the other guys to take. Yeah. Right? For the global guys to take. I would bring in a very important company, her. It's called, it's a company which is US-based called Skid AI. It's building robotic, robotics like foundational model. So it's interesting because it's not building the body.

They don't care about the body. They don't want to build a humanoid. They are building a foundational model. that robotic companies can use for themselves. And it raised like $1.4 billion in January this year. And within like seven months of coming alive, the company was valued over $14 billion. And the entire thing is it's a recurring revenue at the end of today, right?

Like since they are building a foundational model, a Tesla can buy it, a figure I can buy it. All they have to do is. make sure that their foundational model is compatible and can figure itself out in any humanoid body

and where do you think the software data is coming from? It's coming from here. Now with that very uncomfortable realization or I mean confirmation of a realization we've had many times actually I think we're at the end of the episode.

Thank you so much, actually, for joining us and sharing so much about your story. To our listeners, I'd like to say that there are a lot of interesting pictures about what exactly these AI data trainers are recording in the original story, which I'll be linking in the show notes. Thank you so much for listening, and I will see you on Monday. Daybreak is produced from the newsroom of the Ken India's first subscriber-focused business news platform. What you're listening to is just a small sample of our subscriber-only offerings.

A full subscription offers daily long-form feature stories, newsletters and a whole bunch of premium podcasts. To subscribe, head to the ken.com and click on the red subscribe button on the top of the Ken website. Today's episode was hosted and produced by my colleague Rachel Vargis and edited by Rajiv Sien.