How Jason Ma did it: He's Building $600M Robots That Folds your Laundry
Imagine you walk into a laundromat and a robot has been folding laundry non-stop for 24 hours. That's what Jason Ma and his team are working towards: fully autonomous, single-task, commercial-grade robots. They just raised $120 million to get there.
Jason is the lead author of multiple award-winning papers and was recognized internationally for his research. He had offers to return to Google DeepMind, Nvidia and Meta, but turned them all down to start Dyna Robotics. Instead of chasing the sci-fi humanoid — robots that look and move like humans — he's building robots that actually work: folding towels, stacking, packing in the real world.
At Dyna they're building general purpose robots that power the future of the physical economy: AI-powered robots that can do any task in any business or home scenario. To start out they've deployed robots in restaurants, gyms and fitness centers. Jason's view is that the bottleneck for useful robots isn't the body — it's AI and software. Humanoids right now aren't actually very useful, and the cost and hardware readiness is a big factor, so the company first focused on off-the-shelf hardware you can buy for a couple thousand dollars and developed the AI on top of it.
The breakthrough was a 24-hour napkin test: nearly 800 napkins folded with a 99% success rate. Most robot demos are brittle — it takes many shots to get one video that works, and prior works often hit only 70% or 80% success. As Jason puts it, if you try to fold 10 t-shirts and only succeed eight times, that's good enough for a demo but not for real-world deployment. Getting a robot robust enough to run for 24 hours straight is a technical barrier that hadn't really been solved before their work.
Dyna's model is a robotics service business — they don't sell the hardware, they rent the robots out at several grand a month, on par with or cheaper than typical labor cost in the United States. Jason's bigger lesson, moving from research to founder: building a startup is actually quite hard, and the best way to succeed is to build research and product at the same time.
What you'll hear
- Why not humanoids — robots as hardware aren't currently mature and are way too expensive; the real bottleneck for useful robots is AI and software
- Off-the-shelf hardware, custom AI — buying robot arms for a couple thousand dollars and developing the AI on top so they can fold napkins and do packaging at very high success rate
- The 24-hour napkin test — nearly 800 folded with 99% success rate, versus prior works at 70% or 80%, and why robustness over a long duration is the real technical barrier
- A funny failure — the robot pulling many napkins out of the stack at once, and later pulling a napkin too fast so it slipped off the table
- Lab to laundromat — what changes when you leave an air-conditioned office: overheating, bad Wi-Fi, and the operation challenge of trusting a robot on a customer site
- The robotics service model — renting robots at several grand a month instead of selling hardware, on par with or cheaper than labor cost in the United States
- Why he left big tech for a startup — the best way to make an impact in robotics is at a startup, because robotics is a research problem to the big labs but not something they want to solve right away
Key claims from this episode
Chapters
Quotes from this episode
the bottleneck for useful robots is AI and software
— Jason Ma, on why he started with everyday tasks instead of humanoids (00:43) these humanoids right now they're not actually very useful
— Jason Ma, on the state of humanoid robots (00:47) getting these robots to be very robust and can sustain a long duration of like actually doing a task is a technical barrier that hasn't been really solved before our work
— Jason Ma, on the 24-hour napkin test breakthrough (00:24) our goal is to power the future of the physical economy
— Jason Ma, on what's next for Dyna (00:21) the best way to make an impact in robotics is at a startup
— Jason Ma, on why he took the founder leap (21:52) the best way to succeed is to build research and product at the same time
— Jason Ma, on why Dyna does both research and product (00:55)
Themes Jason returns to
- AI and software is the bottleneck — not the hardware; humanoids aren't mature, but off-the-shelf arms plus the right AI can already be very useful
- Robustness over demos — typical demos are brittle and take many shots; the real bar is a robot that works at high success rate for a long duration in the real world
- Build research and product together — the feedback loop from product to research and research to product is what makes AI products good and sticky
- Pick the right problem — having good taste for which problems your robots should solve, and not going so deep in one vertical that you can't move to another
- General over specialized — one model trained on combined datasets for many tasks, much like language models that can chat, write code and do many things