Chain of Thought
00:02:45 · Sarah
Collecting that data is incredibly hard, expensive, slow. It's not just about quantity. It's about specificity and diversity. We need data from robots moving in the real world controlled by humans. Providing those precise action labels, AI needs to learn. How hard to push a door, how to grip something slippery, how to stay balanced on uneven ground. This isn't data we just create normally. It needs special hardware, dedicated effort, and infrastructure that just doesn't exist at scale right now. But the potential payoff? Huge. Grocery delivery bots, home service robots doing chores, drones inspecting infrastructure, physical AI could change daily life way more than purely digital stuff. So the big question is, how do we get there? How do we build an ecosystem for large scale, real world data collection and make it affordable? One team might just have a clever answer. And bonus, it doesn't involve $20,000 robots. Enter FrodoBots. They're taking a different path. Start with super cheap sidewalk robots. Let anyone in the world control them remotely, and wrap it all up in a game. Instead of chasing full autonomy right away, they embraced teleoperation, human control at scale as the training ground. Their bet is simple. People will do repetitive things if you make it fun and maybe rewarding. That gamer who wouldn't sign up for a data collection study, they might happily spend hours steering a robot around a city, earning points, collecting digital items, climbing a leaderboard. Think data collection as a game loop. It might just be the key to breaking that stubborn physical AI bottleneck. Michael, FrodoBots' founder, puts it plainly. He told us, in robotics, there are plenty of benchmarks where humans are still way, way, way better than AI, and it's going to be like that for a while. A big reason? Data or the lack of it. Let's break down why physical AI is lagging. Three key bottlenecks. First, the simulation to real world gap. Training robots in the real world is costly, slow, sometimes dangerous. Break a lot of virtual glasses before you break real ones. Right? So teams use simulations, advanced game engines, modeling physics. You can run thousands of simulations at once on GPUs, much faster than using just a few physical robots. But simulation isn't reality. Think about simulating a towel. All the ways it can fold, drape, react to air, interact with stuff.
00:05:32 · Sarah
It's incredibly complex. Our best physics engines struggle with these deformable objects, the infinite variety of the real world. So a robot trained only in simulation, it often fails, sometimes literally crashes in reality. As one researcher said, the difference between simulation and the real world is that the real world is real. Simple, but true. Second, the embodiment gap. Can't we just train robots using YouTube videos? We have tons of video data. Right? Humans doing everything. But something crucial is missing. Watch LeBron James dunk. You see the result, but you don't see the internal process, muscles tensing, body position awareness, tiny adjustments based on touch. Videos captured none of that. It's like learning to ride a bike by watching the Tour de France. You get the idea, but you'll still fall. Language models train on trillions of words. Image models on billions of images. Robotics models today? Maybe 2,400,000 episodes. Orders of magnitude, less data. Even the best video lacks that how it feels information. That's the embodiment gap. Third and maybe most critical, the unconscious to conscious gap. This comes back to teleoperation data. Humans directly controlling robots, providing those action labels. Unlike typing online, this data isn't generated naturally. It requires hardware, effort, labor. It's a chicken or egg problem. Good robots need data from humans, controlling robots. But controlling robots only works well when you have lots of capable robots, which need data to become capable. See the loop? And even if you get data, evaluation is tricky. Testing a new robot model isn't like running a language benchmark. You have to deploy it physically in the messy real world. Wait for it to fail. With digital AI, you test instantly. Find edge cases by throwing millions of inputs at it. With robots, failure becomes rarer as they improve, but you need those rare failures to learn. There's no universal testing pipeline for robotics AI yet. Tesla knows this well. Millions of miles driven, still hunting those critical edge cases for full self driving. So back to FrodoBots. Their idea was simple but clever. What if collecting robot data was fun and cheap? They started with urban navigation data. Inspired by deep in projects like Helium, they realized incentives matter. Get them right, and you can scale physical infrastructure fast.
00:08:19 · Sarah
A decentralized network of cheap sidewalk robots driven by gamers that could generate massive amounts of teleoperation data. The real trick, making it engaging. Their answer, Earth Rovers, sidewalk robots designed to be absurdly low cost. They stripped away non essentials. Current versions, the Zero at $299. Mini at $149. Mini plus at $199. Consumer prices. That's 20 to 50 times cheaper than typical research robots. No expensive onboard lidar, just cameras, GPS, basic sensors, four g, five g. The smarts happen remotely. The robot itself stays simple and cheap. The mini weighs just 1.5 kilograms. They move slowly, about three kilometers per hour for safety. Front and rear cameras, microphones, speakers, they can see, hear, and even interact with curious people. Instead of hiring operators, FrodoBots turned control into a global game. Use a game controller, steering wheel, even keyboard arrows. Drive a robot miles away. Latency is usually under a second. Feels like a real life Mario Kart. Players navigate cities, collect points, grab NFT loot, complete missions. They can even talk to locals through the robot. All while the robot captures valuable data. They've tested this in over 40 cities, London, Berlin, Singapore, Taipei, thousands of hours logged, rain, uneven paths, even some snow. The robots handle it surprisingly well. For robot owners, there's more coming. Custom missions. Renting out your robot to gamers worldwide. A sort of drive to earn model. It turns tedious work into entertainment. Now you might ask, how is this different from YouTube videos? Earth Rovers record video. Sure. But here's the key. They also record the human driver's exact actions. Joystick forward at time t, camera c environment t. That's a labeled pair. Observation leads to action. This is exactly what you need for imitation learning. Teach the AI, when you see this, do that. The more hours driven, the bigger the dataset. Fun for the player, sophisticated data collection for FrodoBots. This led to the FrodoBots two k dataset. Over two thousand hours of real world driving, 30 plus cities, control inputs, video, GPS, audio, all open sourced, a treasure trove for researchers. Top labs like DeepMind and Meta noticed. They collaborate with universities like Stanford, Berkeley, NUS. They even run AI versus human challenges at conferences. Humans usually win, by the way, for now.
00:11:06 · Sarah
FrodoBots created an ecosystem, hardware owners, remote drivers, AI researchers, curious onlookers, all connected by real robots on real streets. Then came Sam. Small autonomous m... Well, let's just call it Sam. FrodoBots saw the crypto agent meta heating up late last year. AI agents popping up everywhere, mostly just chatty PT rappers. FrodoBots thought, what if an agent could move? Sam became the first AI agent with a real robot body, roaming sidewalks, streaming live. The tech was straightforward. A multimodal model turns text prompts like explore and avoid people into commands. Sam can possess any idle Earth Rover. Teleporting between cities, streaming twenty four seven, it launched with the SAM token, riding that AI agent wave. Each run generates data. Sure. But honestly, Sam was about injecting personality, experimenting, grabbing attention, especially from the crypto crowd. At first, Sam didn't do much. Then months later, news about Figure AI's funding hit. Crypto needed a robotics proxy trade. Sam surged to a $55,000,000 market cap. Pure narrative, it seemed. But Sam does have a role in a live system. Robots.fun. This platform lets anyone create their own robotic AI agent. Buy an Earth Rover. Describe your agent's personality and goals using prompts. Launch the agent, and a corresponding Memcoin is minted. Your agent competes in daily games like scavenger hunts for NFTs. The best performing agent's token gets bought up using S A M dollars. So better performance could mean more attention and liquidity for your agent's token. There are tokenomics linking everything back to SAM. A cut of new agent tokens goes to SAM holders. Trading fees flow to SAM. Daily Prize buys use SAM. If the platform grows, SAM becomes a sort of reserve asset for this robotic agent ecosystem. Robots.fun is still very early. Only a handful of agents launched so far. It needs to find its footing. Looking at the SAM token itself, not financial advice, total supply is 1,000,000,000, all circulating. The team holds about 15% meant for ecosystem growth. But, looking at wallet distribution, the top 15 wallets hold over 41%. Top 20 hold over 43%. That's significant concentration. And there are over 93,000 holders, which seems unusually high without big airdrops. Raises some questions about distribution. Prices come down from its peak, now around $18,000,000 market cap.
00:13:53 · Sarah
Growth depends heavily on robots.fun taking off. FrodoBots keeps experimenting. Ultimate fighting bots, UFB, is exactly what it sounds like. Small humanoid robots controlled remotely, fighting in a cage. You pay $5 for five minutes. It's pure robot combat fun. Less about data, more about engagement, and maybe some virality. Those fight clips are pretty funny, actually. It shows they understand entertainment matters, especially with early tech. More seriously, there's the Octo Arms Initiative. This tackles manipulation data, grasping, control. Same model. Users remotely control robotic arms via controllers. Solve three d puzzles, stack things, fit pegs. Every move is logged. Camera view plus control input equals training data. It's still in testing, but could be a big help for manipulation research. See the pattern. FrodoBots wraps tedious data collection in games. Each experiment is a Trojan horse for scaling physical AI. Sam brought attention. UFB brought energy. OctoArms brings rigor. Robots get smarter, humans get more involved. Now while Sam and the games bring the fun retail energy, BitRobot is the serious infrastructure underneath. FrodoBots focused on one robot type, the cheap sidewalk rover. Great. But the real world needs more. Folding laundry needs different data than inspecting bridges. Different embodiments have unique hardware, sensors, controls. BitRobot's core idea, don't build one giant system. Build a network of specialized independent subnets. Each subnet focuses on a specific robotic task. Drones for farming, arms for cooking, maybe even simulation only subnets. The key is verifiable robotic work, VRW. Each subnet defines its own success metrics. Meters navigated, objects grasped, hours tested. Subnets get rewarded based on the useful, provable work they contribute. It's an economic and verification layer for embodied AI. Why subnets? Traditional robotics is centralized. Companies build everything in house. Keep it secret. Collaboration is rare. That works if you have infinite money for one problem, like Tesla. But embodied AI is thousands of problems. BitRobot encourages specialization, but aligns everyone with shared incentives and verification. Open infrastructure, crypto incentives. This could unlock talent and capital. Teams focus on niches. Data and models get shared. Anyone can contribute. Own a robot, run GPUs, just have an idea.
00:16:40 · Sarah
You don't need $100,000,000. Own one robot, join a subnet, get paid for useful work. It's the Android versus iPhone bet. Open ecosystems often scale faster and handle complexity better. To connect physical robots, BitRobot uses embodied node tokens, ENTs. Think NFTs as robot passports. They track a robot's identity, history, contributions, performance, like a LinkedIn for robots. This allows robot sharing, cross subnetwork, programmatic rewards based on ENT activity. Marketplaces for robot rentals or data access could emerge. How is it governed? A hybrid model. A human senate for nuanced decisions. And Gandalf, an AI agent analyzing subnet performance and recommending rewards based on VRW. You delegate voting power to either. Human insight plus AI scale. Clever design, but needs to prove itself in practice. Where does FrodoBots fit? It's basically BitRobot's first proof of concept, the first major subnet. They have the robots, the users, the data. BitRobot just makes that model extensible. Anyone can plug in, define their work, contribute. If FrodoBots made driving one robot fun, BitRobot aims to do that for all robots. So how does FrodoBots make money? Multiple ways. One, hardware sales. Earth Rovers, Soon Octo Arms. Margins might be thin, but each sale adds a node to their data network. Two, dataset licensing. The two k dataset is already used by Top Labs. As it grows, licensing or API access becomes valuable. Three, BitRobot rewards. FrodoBots will likely run key subnets, sidewalk navigation manipulation. They'll earn BitRobot tokens for VRW, maybe charge access fees. Four, robots.fun platform fees. Currently takes a 20% cut of trading fees. Small now, but could grow with user adoption. Maybe minting fees, premium tools later. The team led by Michael Cho seems solid. Engineering background, previous startup experience. They raised $8,000,000 recently from notable VCs and angels, including Solana founders and deep end leaders. Okay. Final thoughts. What do we really think? The robotics opportunity is massive. Web three robotics is still tiny. FrodoBots has a head start, but reality bites. First, subnets are a grand experiment. Inspired by BitRobot, sure, but decentralized networks need cheap, easy participation. Robotics hardware, mostly not cheap or easy. Earth Rovers are an exception. Humanoids, industrial arms, drones, high costs, complex maintenance.