Key takeaways
- Bitrobot uses crypto incentives to create a decentralized network of subnets for robotics AI research, similar to Bittensor's architecture.
- Robotics data collection differs from other AI fields because sensor data cannot be easily scraped from the web like text or images.
- Frobots operates RC robots in over 30 cities globally that gamers remotely control while collecting valuable action label data for AI training.
- DePIN networks enable distributed capital formation and coordination among hardware owners, researchers, and data collectors without requiring single corporate ownership.
- The gaming interface allows web2 gamers to contribute training data unknowingly while playing missions that resemble Mario Kart and Pokemon Go mechanics.
Why robotics needs crypto
Ashton Addison: What is Bitrobot?▶ 00:25
Jonathan Victor: The goal is very ambitious. We want to use crypto incentives to solve embodied AI, or basically robotics AI. We intend to do that by creating a network of subnets, similar to Bittensor. Each subnet within Bitrobot will work on specific robotic tasks that are meaningful and will move the needle for robotic research.
Ashton: Why does robotics suit a DePIN model, and what sets it apart from other parts of AI?▶ 01:42
Jonathan: In Web3 we're often hammers looking for nails. If you start from the problem and reason back to where blockchains might be useful, you find more interesting problems. Robotics is different from other parts of AI because there isn't a giant corpus of data you can pull off the web. There hasn't been 20 years of people spewing out sensor data. Companies like Tesla and Figure own the whole pipeline: they raise the capital, build the hardware, and collect their own training data in their own labs or warehouses. To create an open equivalent, you need to coordinate many distributed parties: capital owners who contribute robots to a shared fleet, researchers who don't have billions of dollars of hardware, and people who bounty out tasks or data sets. Crypto lets each participant be rewarded for their contribution without one company owning the entire pipeline. The subnet model also lets us focus people on specific tasks, like robotic arms folding laundry or urban navigation.
Evaluation might be an even bigger bottleneck than the lack of data.
From a robot game to subnets
Ashton: How do contributions work in practice?▶ 05:43
Jonathan: We started building sidewalk robots almost three years ago. This is our latest version. It's like a toy RC car with a 4G SIM card, so you can remote control it from anywhere. You could sit in New York and drive one around in Africa. We structured it as a game, a bit like Mario Kart or Pokémon Go. Gamers look through the robot's camera, and we give them missions, like picking up loot boxes and Solana NFTs. Most of the players are Web2 gamers who don't even know there's a crypto angle. As they play, we collect data: the video stream, but more importantly the action labels, what the human is doing. That data set is very rare, and it's very valuable for robotic research. We have robots in over 30 cities, and we've worked with research labs including Berkeley and Stanford. Last year we co-organized an AI versus human competition with researchers from DeepMind. You can think of the last two or three years as subnet one: the game, which incentivizes gamers and people who buy a robot that costs about $150. Subnet two could be a much bigger version of that competition, incentivizing researchers and compute providers. The output of subnet one is a data set, and the output of subnet two is an AI model. Crypto has a unique advantage in doing capital formation globally, very quickly.
Ashton: How do you turn that data into robots that people see changing their lives?▶ 13:08
Jonathan: Set expectations properly. These big innovations are like the integral of many S curves together. You shouldn't expect to go from A to B overnight. The thing to watch is progress on specific tasks. We spend a lot of time talking to researchers to find the most important near-term problems with no clear path to a solution, and then prioritize which tasks to collect data for. Urban navigation is one, and robotic arms and manipulation are another. There's also a paper showing that embodied AI models trained on data from many different sources perform better than models focused on a single task. That's part of why we're bullish on multiple subnets focused on different tasks.
A reality check on timelines
Ashton: AI is improving fast. How quickly will people see that in robotics?▶ 15:54
Jonathan: People hear "AI" in embodied AI and expect the same iteration speed as the latest LLM. In the digital world you need data, compute and talent. When a new model drops, within half a day everyone has a good idea of how good it is, so evaluation is never the bottleneck. Robotics is different. Say Tesla has driven a million miles with no crashes or disengagements. Have they reached level five, or are they still at level two? What if there's a crash next month? You can't tell how good your model is in simulation. You have to test it in the physical world. So evaluation might be an even bigger bottleneck than the lack of data, and a lot of people are going to be disappointed that it takes longer than expected. Robotics could be the final frontier on the path to AGI. It will draw more attention and funding, and it needs a lot of capital and a lot of physical robots.
Ashton: Why is gaming the right starting point?▶ 21:00
Jonathan: In my view, a robot taking over a chef's job at a restaurant is at least a decade away. For most use cases, 99% success isn't good enough. You probably need Six Sigma, 99.9999%. We've seen this in self-driving. I sat in a Google self-driving car in 2015, ten years ago, and thought it was coming. We're still talking about production. In the meantime, gaming has real entertainment value, and it has often been where new technology gets pushed forward. We have GPUs doing AI today because GPUs ran games for decades. Gaming could be the wedge that ties things over while we collect data in the real world.
Gaming could be the wedge that ties things over while we collect data in the real world.
Solana, partnerships and getting involved
Ashton: How can people contribute right now, and what comes next?▶ 23:38
Jonathan: We've just announced the concept of Bitrobot, and the whitepaper will be published in a couple of weeks. In the coming months we'll have a more concrete proposal and move into testnet. Bitrobot will be built on Solana. In the coming weeks we'll also announce partnerships with other DePIN projects. These are early showcases of what future subnets could be: projects with a lot of GPU compute, existing projects like ours with sidewalk robots, and projects with large gaming communities. Many of these partnerships also involve research labs, because without researchers, even if you provide compute, nobody runs the GPUs.
Ashton: Why Solana for a robotics DePIN network?▶ 26:22
Jonathan: Three years ago I was very anti-crypto. Then we discovered Helium two months into the project. They minted a token out of thin air, crowded in so much capital, and did it in a geographically distributed way. I was mind-blown and decided I needed to learn how to do this. A week later there was a Solana hackathon in Singapore. I didn't even know what a smart contract or an L1 was. We've been building on Solana ever since. For what we do, success mostly depends on our hardware and networking stack, but Solana has been a decent place to build. It has survived many setbacks, and there are now very big DePIN projects on it. We also have backing from Solana Ventures.
Ashton: What else can people get involved with in the near term?▶ 28:58
Jonathan: The NFTs are not a limited 10,000-piece collection. They're in-game assets used for utility. We also have Sam, an embodied AI agent with his own memecoin. Sam is built on Virtuals and live-streams 24/7 under the Twitter handle "Sam is moving". Anytime you look in, you see an AI driving one of our robots in a random city, and it's 100% autonomous. This is not financial advice. It's just a memecoin with no pretense of utility. Crypto has two extremes. One is heavy capital formation, which is what we're building with Jonathan and the Protocol Labs team. The other is memecoins, a fickle community but a great way to gain attention very quickly.
This interview has been edited for length and clarity. Recorded February 2025.

