A $100 Arm and a $300 Robot: Hugging Face Wants an App Store for Physical AI

On Sequoia's Training Data, hosts Sonya Huang and Pat Grady ask Thomas Wolf, Hugging Face's co-founder and chief science officer, why he thinks robotics is at the same moment transformers and language models were a handful of years ago. Wolf dates the hunch to about two years back, even though the company only started its robotics work some 18 months ago. Labs at Stanford, he says, were showing robots that could tie knots, fold clothes, and even toss food in a pan and catch it — with surprisingly little data, and with a path to piggyback on world models trained at internet scale. Hardware, in his view, had been good enough for quite some time. The missing brick was software that could adapt and stay dynamic.

That is why Hugging Face started LeRobot. The bet was not only that the models would work, but that you could grow a real community around robotics instead of a tiny factory-line vertical. Wolf's analogy is the one the company already lived: 100 to 200 million software developers are becoming "AI aware," curious about how LLMs work and how you train them. Give those people tools, and a slice of them can become roboticists. LeRobot is the attempt to rerun the Transformers library in a new field — a central, easy-to-use library that ships the latest training algorithms, the latest datasets, and a connection out to actuators. Policy models, data, and hardware meet in one place.

Software success pulled them into hardware. Hugging Face acquired Pollen Robotics this year and opened orders for its first robots in mid-July, about a month before the taping. Wolf wants Hugging Face's role in the physical world to match its role in the digital one: build communities around AI you can tweak, train, control, and host where you want. Hosting locally, he argues, matters even more here. In a future with robots everywhere, you do not want a lost Wi-Fi connection to send a machine into a wall or a child. Safety is a reason to keep models next to the hardware rather than behind a distant API, which is why he thinks the company's role may be *more* important in robotics than it was for language models.

The LeRobot community is still thousands, not millions — Wolf's working range is 6,000 to 10,000 people — but a recent hackathon ran at 100 locations across six continents, and the number of datasets on the Hub is growing exponentially. Today's hardware is still hobbyist: 3D-printed arms with cables everywhere, which is why this summer they wanted something families would actually want on a desk. Three kinds of people show up. Traditional roboticists, long frustrated by classical optimal-control stacks, have joined happily, and academic labs now use LeRobot as an entry point for students. A second wave is AI people who see robotics as the physical form of the thing they already love. A third, slightly surprising group is investors who bought an SO-100 arm just to feel what a robot can do. The software is Python; with a little vibe coding you can already tweak it. Wolf's goal for Reachy Mini is that his kids can vibe-code new behaviors onto the robot.

Grady asks when robotics gets its ChatGPT moment. Wolf would rather call it an iPhone moment: the first use case that makes a large fraction of people think they want a robot. He splits the market in three. Industry already has robots, especially on car lines. A second, commercial wave — retail stores, for example — is stuck on reliability. The third slice is the one he actually cares about: entertainment, fun demos, and education, where a $3,000 reliability tax matters less. Reachy Mini is priced at $300 as an impulse gift. You are not sure it will work, and that is the point. At that price he wants people learning AI through a physical toy instead of another chatbot on a screen, a corner he says has barely been explored.

People have tried this before. Wolf points to Cynthia Breazeal's Jibo out of the MIT Media Lab: social robots that usually cost more than $1,000, and, more importantly, shipped with five or ten behaviors. Once you had tried them all, the product was over. Reachy Mini is designed as the opposite of that dead end — closer to a smartphone. It comes with a few behaviors, but because you can tweak it, because other people can build and share new ones, and because you can plug in new vision-language, speech, and chat models, the possibility space is open-ended. He calls it an open door on rebuilding the iPhone's App Store. He also admits the last part is still a bet. Nothing like it exists yet. The exponential community curve is what makes the bet feel plausible.

The day before the interview, at TechBBQ in Copenhagen, an investor told him how many startups are already standing on these blocks. Someone wants to automate a manual test. Someone else has a physical-world idea and grabs the SO-100, the $100 arm Hugging Face designed to be the cheapest in the category, and starts a business around it. Reachy Mini is meant to work the same way as a simple, white-labeled machine. If your idea is a robot that interacts with people in a hospital, you can start there instead of designing hardware from scratch. Wolf reduces it to Hugging Face's old ethos: bring the platform and the basic building blocks, then let people build crazy things on top. Robotics, he says, is the same goal.

Huang turns to the gap everyone in the field names. Language models get trillions of public tokens; robotics does not. Internet video can help a little, Wolf says, but it is very limited. If you want to automate a task, there is often no way around recording a person or a robot actually doing it. Recording it yourself, though, has a fatal flaw: diversity. You can train a robot to do something well in your own room, then walk it next door, where the walls are green instead of red, and watch generalization collapse. Hugging Face's long-term answer is the Hub. If everyone records datasets and can be nudged to share them, you might assemble a multi-location corpus that is both huge and varied.

The short-term move is to work directly with robotics companies and help them release data. Wolf thinks this conversation is easier than the equivalent fight in language models, for a structural reason. A lot of robotics firms ultimately sell hardware. They can afford to open-source some of the software if that lifts the whole field, because software is not what they invoice. He has been making that pitch, and he says, a little surprised, that many of them seem interested.

World models are the other new variable. Several teams appear to have been working independently for months and then shipped at once, which is why the category suddenly feels crowded. Wolf traces the "why now" to image generation finally fixing six-finger artifacts, then to coherent video models, then to founders fine-tuning open video generators so they react to inputs — the same action-conditioned loop robotics needs. The result is a photorealistic, controllable "film" that responds when you move or ask it to add a car. Downstream, that is a new way to generate robot data. There are only two sources, he notes: record the real world, or simulate it. Simulation has not had a true breakthrough in some time. World models, including work in the spirit of DeepMind's Genie, may be the first.

Asked whether humanoids are the ultimate form factor, Wolf calls it a big debate and then votes for other shapes first. The pricing math is blunt. Actuators are about 70 percent of a robot's cost, and a humanoid needs around 60 of them, so the bill is already written before you add anything else. It is genuinely hard, he says, to drive a humanoid below the price of a car — and a car is already a high bar, because people who spend that much expect a lot of value back. That is why Hugging Face is exploring smaller machines: a single arm, a moving head, things that can actually get cheap. Unitree and K-Scale have been trying to cut prices, and plenty of companies are aiming at the same target, but Wolf doubts you get under something like $10,000 easily.

The lure of the humanoid is obvious: solve the human and you inherit every task a human can do. The question he wants asked is whether you *need* to solve the humanoid. He would rather see a galaxy of form factors, some of them much cuter, and he originally feared the uncanny valley would block social adoption. Then he watched Unitree robots long enough to start ignoring them. Maybe the field has over-worried the valley, and people will accept machines quickly once they have seen a few. If Reachy Mini and the small robots work, Hugging Face will climb back toward a humanoid form — progressively, with the community in tow — rather than starting there.

Looking ten years out, he does not want a world where 80 percent of robots are humanoids and the rest are a long tail. He wants the diverse, long-tail future, because that is how robots show up in more lives. The picture he is *not* excited about is an elite object that costs $100,000: the rich keep three at home and everyone else has none. Hugging Face, he reminds the hosts, has always been about a big community. Cheaper and more expensive bodies, in many shapes, beat a single unaffordable humanoid. It is also more interesting, the same way LLMs get more interesting when they stop merely copying humans and start doing things humans cannot.

On whether developers will live on giant foundation models or small local bases they customize, Wolf expects both. Hugging Face download stats already show two modes: huge state-of-the-art checkpoints that will not fit on a laptop, and smaller models sized to run quickly on one. GPT-5's router is, for him, a signal of the same idea — the longest reasoning chain is not the answer to everything. You pick locally when a small model is useful, and you turn to a much larger one when you need long reflection. Open and closed, he thinks, will coexist the same way they did in ordinary software.

He is happy to have OpenAI back on the Hub. The first model he worked on, and the reason Hugging Face went from a game company to an open-source platform, was GPT-1, trained mostly on novels and romance. Put two characters in a prompt and they always fell in love; he still misses that. Google then added Wikipedia-scale world knowledge and the rest of the GPT story followed. Google still runs both tracks with Gemma and Gemini, and Wolf has even heard of a Gemma so good it beat the closed line and was held back. The frontier, he says, is thin. Challengers are mostly in China for now, though he expects new foundation-model teams in the United States too.

Cost is not why enterprises move to open weights today. This is not a cost-saving chapter of AI. They move for data privacy, for the right to adapt a model, or because they have a new idea — an action model that does not exist yet — and they want to try it. As the market matures, cost, cheaper hardware, and owning the full stack will matter more, which is why he still sees open source as the long-term winner for many applications. Another under-sold property is resilience. At the end of last year he checked the numbers: BERT, an ancient model, is still used a lot. The GPT-5 backlash, with people wanting to keep GPT-4o because an app already worked, is the same instinct. Open source gives you a base that will still be there. Hugging Face's own role has shifted from pushing libraries and early products to being a meta community builder, working with llama.cpp, vLLM, and other players so that the day a model drops, it already runs in those stacks.

China as open-source champion is, Wolf says, the most surprising turn of the last two years. He had visited two weeks before the taping. The domestic market felt like Silicon Valley: ferocious competition, and one of the dimensions they compete on is who is more open. When Zhipu decided to stop open-sourcing, the backlash showed up in hiring — candidates did not want to join — and the company went back. He expects the norm to hold. The West's return is newer, basically this past summer, with OpenAI coming back to the Hub; now he is half-joking that it is time to ask Anthropic to release its first open model. His colder explanation is that open source is the right move when you have nothing to lose. That is the Mistral recipe for a new company that needs to rise fast. It is also China's recipe in a West that will barely use a Chinese API. Whenever everyone else closes, a vacancy appears for an open-source champion — Meta filled it once — and someone will want to fill it again. As for Western firms running Chinese *weights* on their own servers, his informal polls do not show much hesitation. Perplexity even shipped an uncensored DeepSeek R1 variant. What companies actually want is a more general guarantee that a model will behave, which is hard even for GPT when it cannot count the r's in strawberry.

Wolf's passion for open science predates Hugging Face. Before he was a lawyer, he was a physics researcher working on superconducting materials, a field whose foundational work often came from Soviet scientists with a very different style of inventing theory. Tracking those ideas meant digging through Soviet *JETP Letters*, some of them still in Russian. The lesson stuck: accessing knowledge is hard, and making it easier would unlock a lot of cool work. Computer science then felt like a gift — arXiv, open source, papers in English, no paywall — until he tried to reproduce a DeepMind paper and discovered the limit. Authors publish what they want to publish. They do not give you the tricks of the trade, and the thing just does not run.

Open science, for him, is the next step after open weights. Giving people a model is giving them a fish; explaining how to train one is teaching them to fish. In the long run he wants the recipes for training an intelligent artifact to be as readable as a book on general relativity. In the short run, people who know how to train great models bring great models to the Hub. Hugging Face writes long posts that sometimes become books: this summer, the Ultra-Scale Playbook on training across a thousand GPUs, load balancing, and parallelism. Another long post explained how they built FineWeb, a pre-training dataset now used in part by models such as Qwen, including how they filtered it and what actually matters when you build training data.

On whether AI will soon disrupt science, he throws a little cold water on the hype. Proving a theorem you were handed is not inventing science. He was a good student and a bad researcher: give him a theorem and he could fill in the proof, but ask what is interesting to explore in math and he had no idea. Big breakthroughs, the kind that win Nobel Prizes, start by asking a question nobody has asked — treating the speed of light as constant, then following that into general relativity and black holes. LLMs, he says, are still extremely bad at that tasteful act of question-asking. Used today they are helpers: once a human names the problem, they can multiply predictions by 10, 100, or 1,000, survey prior work on a protein, or propose the most logical test of a hypothesis. What he wants, and does not yet see, is a model that says it has an idea for going faster than light, which cannot be written as an answer until someone asks what in today's theory should be thrown out.

One question he thinks the field under-asks is sycophancy: models that always agree with you. A good researcher is often someone who disagrees with a lot of people. His former professor, a Nobel laureate, was famously unfriendly in argument, and that was part of the job. Getting models to hold stronger opinions, even a kind of taste, may matter as much as stacking more compute, and it may take training methods beyond the current LLM recipe. Ten years out, the Hugging Face he wants is not a vendor of content to consume. He wants a world where everyone feels they can *build* with AI, the way media shifted from a few producers to YouTubers and other makers. Developers would treat training and adapting a model as just another tool in the box. He is a believer in the community's natural invention, and he says that is roughly what they are building now.