I, Robot

Humans have had a close relationship with humanoids and automatons for a long time.

Humanoids were first described in the ancient myths of Homer. Hephaestus, the Greek god of fire and blacksmith, was said to build “handmaidens wrought of gold in the semblance of living maids”, who had “understanding in their hearts, and speech and strength”. In the Renaissance, Leonardo da Vinci sketched designs for a mechanical knight that could sit, stand, and move its arms. Today, we have Figure 01, which makes dishes and gives you an apple.

It’s one very long arc of history. There’s something peculiar about us trying to mold a human-like figure and put it at our service for thousands of years.

Automatons (mechanisms that automate a sequence of actions) first appeared way before the Greek myths – it was around 1500 BCE in Egypt, when clepsydra, the first water clock, was created. It was similar to a traditional water clock (which uses water flow to measure time) but the chime was activated by the figurine instead of a simpler mechanism.

Leading up to the 20th century, countless automatons and other proto-robotic machines performed practical or less practical tasks (like party tricks). Then, in the 1950s, George Devol created Unimation, the world's first robotics company, and sparked a wave of industrial robots.

Automated Horses or Cars?

When we set out to write this article, we asked the question, what is the area in crypto that has only one or two teams working on it? We didn’t come up with many ideas. But our investment in Frodobots suggested that we should discuss robotics, or RWAs, as we’ll henceforth refer to them (real-world agents).

It’s only a matter of time before crypto investors (AKA ponzi connoisseurs) realize how much money they can deploy into robots on blockchain. Just imagine the possibilities. Layer-1 for robots. Storage for robotic data. Marketplace for humans to sell their data to robots. Decentralized compute network to train the robots. Launchpad for robot tokens. DEX for robots to procure resources. This is the dawn of the Fat Robot Thesis.

For context, web2 investors have already fallen in love with robots (and projected that onto others) and put obscene amounts of capital where their mouths are. Michael Dempsey’s article from last year perfectly dissects the core beliefs needed to underwrite this, the incentive structure for why it’s happening at such a rapid pace while ignoring many intermediate steps and jumping from 0 to all-purpose humanoids.

From Michael Dempsey

The graph above explains what assumptions you have to hold to believe that robotics will be a hugely rewarding category for investors and founders. Will the scaling laws (more data, more compute) apply to foundation robotics models? The consensus among researchers (and, by extension, investors, too) is “yes.” If so, it means a lot of capital can be deployed, which robotics has an insatiable demand for. It is a perfect example of exploration to exploitation, where most ideation stops after a consensus has been established.

Large language models are probabilistic. They are designed to predict the next token (which can be a word, part of a word, or another unit) in a sequence given the tokens that came before it. This prediction isn’t a single fixed answer; instead, the model assigns probabilities to a range of possible next tokens. These models generalize well because language and text are explained by grammar and logic.

However, even this oversimplification could invite criticism; for example, Ludwig Wittgenstein's interpretation of language (“the meaning of a word is its use in the language”) would reject a probabilistic model of language, because socially relevant context is key. LLM is much closer to a Platonic ideal of forms that captures a vast array of linguistics and approximates some “ideal” form of language.

Now, take this probabilistic model to physical reality, where randomness and variations are infinite (unlike in language, where they can be systemized at least to some extent). How much data and compute will be required to build a system that can operate in an infinitely random reality, or realities that haven’t even been conceived by humans yet because they haven’t been encountered? It looks to us like every proponent for AGI (embodied or otherwise) always lacks an explanation for how these systems would go from being dummy assistants to explaining the unexplained, or how knowledge is created in the first place.

The attempt to pattern-match and generalize by building robots that are very similar to us (humanoids) is enticing. But often the future is unpredictable and unfolds in ways we couldn’t imagine. If you were designing a perfectly capable human in 2025, would you give him fingers that looked like sausages? Or would you equip him with multifunctional tools instead? Or is that a wrong question to ask in the first place?

The world is always rethinking the shape of everything. Supply chains, manufacturing, national security – all of it hinges on the nexus of robotics bridging the gaps between that future state and what exists today. Robotics will lie at the heart of many of those transformations, and so will crypto, which is the fastest form of money, and the most fluid form of capitalism – the fastest way to lose life savings. Welcome to the Robot-Industrial Complex.

Whether it is data collection, funding, or networking, crypto-powered robotics will be an integral part of the upcoming industrial revolution. In a spirit similar to DeSci, crypto should fund small-scale, localized experiments as opposed to headline-grabbing robots that do backflips.

We need cars, not automated horses.

Without question, we’d find an investor in crypto who’d fund this if it had a token

The Fellowship of the Rover

Our robot journey started in 2023 with Frodobots. A startup of humble beginnings founded by people with no formal training or credentials, yet with a big and passionate heart for robotics. Entirely self-taught in a discipline that genuinely interested them. Michael and his team fit well into what we are looking for.

Many investors, including us, shied away at first. It was a weird combination of gaming, crypto, AI, robots, and DePIN – everything and nothing at the same time, a new category. Frodobots would assemble and sell these sidewalk robots to individuals all around the world, and the owners would then rent these robots to players to zip around the cities, guiding them with a controller while sitting on a sofa at home. The players would participate in a game to complete quests and earn rewards. The more difficult the quests, the longer the routes, the more rewards players would earn.

Two gamers drove a robot into a store and bought milk tea

On our diligence call, we drove our robot into a drunk German guy holding a beer who started dancing in front of us and talking to us through a microphone attached to the robot. It was clear there was something to this idea, despite the myriad of reasons we came up with for why this wouldn’t work, the most obvious being that it would be banned by local authorities.

On the surface, it looked like a playful game where you explored the city with a robot. But the team’s core idea was to capture video data from the journeys (there’d be multiple cameras attached to the robot) and build the most valuable dataset in the world for training sidewalk robots that would be on par or better than possessed by companies like Tesla or Waymo. Eventually, people paid attention, and Frodobots hosted a robotics competition with Google’s DeepMind in Abu Dhabi to crowdsource real-world datasets.

The undertaking of Frodobots sounded as quixotic as Frodo’s quest to destroy the Ring of Sauron. But, every great feat requires naivety and delusion.

The Fellowship of the $299 Earth Rover

Data, Evaluation, and Incentivized Failures

Last year, Frodobots began working with top robotics researchers from around the world. What began as a zany quest-based real-world game opened up a wide window of opportunity. Perhaps appearing unserious is key to actually discovering a unique truth.

The insight here is that the success of robotics in autonomous driving, as demonstrated by Tesla and Waymo, cannot be easily replicated in other robotic embodiments. Why? Because there isn’t a global fleet of devices (like Tesla cars) that can be used to collect data at scale…yet.

Not all data is created equally. In robotics, there are several main methods to obtain it:

1. Synthetic data (artificially generated data that mimics real-world data, created using computer simulations);
2. Internet video data (video content sourced from online platforms such as YouTube);
3. Real-world data (data collected directly from the robot’s environment using onboard sensors).

The data is plentiful in the 1st and 2nd methods, yet the most valuable is the 3rd. Every model has to be ultimately tested in a real-world setting, but the high barriers of capital and labor intensity have prevented this in many robotic fields. Consider that every autonomous car still needs a human, the driver! So the robot-to-human ratio is 1:1, even for Tesla.

For efficiency purposes, researchers have mostly worked with synthetic data and rarely attempted to run it in the real world. As we’ve discussed earlier, the edge cases in the physical environment are infinite and cannot be all simulated in a lab.

There is a need for a solution to run many experiments in different fields of robotics while distributing the capital and labor costs across many different entities and individuals. There are so many robotics enthusiasts in the world who tinker around with robots irrespective of any financial reward.

To put it loosely, Frodobot’s evolutionary step – BitRobot – is a combination of Bitcoin (compute), Bittensor (research), Helium (DePIN), and Axie Infinity (human labor). Different subnets will be dedicated to different robotic experiments (e.g., sidewalk robots, surgical robot arms, human-teleoperated cooking, etc.), and the robot teleoperators performing the work will be benchmarked and rewarded with BitRobot tokens.

There will be opportunities for a broad group of contributors. You can imagine that a subnet will require embodiment hardware, compute networks, and human teleoperators to manage the embodiments, among other things.

We should point out that BitRobot’s evaluation of existing (synthetic) data will be even more important than crowdfunding new data from teleoperators. Private AI labs, universities, and startups lack the physical hardware to run real-world experiments. We’ve heard examples of robotics teams that have only worked with simulated data and never put it to test on robots in the last 2 years. But now, they will be able to obtain this hardware from BitRobot and perform the evaluation of their data created in a lab.

As Michael from Frodobot’s has pointed out to us: the first team that achieves AGI (whatever that means) is the one that achieves most failures in the physical world (crashes or other failures). The company with the most physical failures in its autonomous system today is Tesla. Just like every plane crash makes the next flight safer, every robot failure brings us closer to embodied intelligence. That’s why we should strive to have more experiments and more failures – not less.

From Degens to Degen Defense (DeDef)

Crypto is best at creating new markets. BitRobot’s token incentives will be used to bootstrap and get RWAs into the wild at much higher rates. Imagine the variety of robot-based games and activities people will be able to do. What is the value of novel forms of interactions and new roles for people to play in a world that values entertainment and indulgence above all else? Perhaps the value of BitRobot will not be the data (people in crypto really have an unmatched desire to collect data) but an amusement?

Imagine a network of agents (à la Truth Terminals) on Tesla wheels selling AI-generated fortunes or memecoin predictions, teleoperated by random people on the internet. The possibilities explode with a network to easily facilitate these experiments. It sounds far-fetched, but BitRobots might equally become a dystopian social experiment as much as a solution to a hardware/resource allocation problem.

If you’re a casual agent speculator or robotics enthusiast, keep your eye out for robots.fun and Small Autonomous Motherfucker (SAM). The former will launch an agentic launchpad for robots where embodied agents will launch their tokens, challenge each other to a deathmatch, and rent robotic hardware from humans. All of it will underpinned by the platform token SAM.

If there is a wave of successful experimentation with BitRobot, we would love to see attempts at developing Decentralized Security ideas that we wrote about previously. Perhaps BitRobot’s subnets could also be used to crowdfund robotic transformations of smaller-scale manufacturing operations as part of the new administration’s national security agenda.

Reach out to us on X if you want to discuss it.

Disclaimer: Zee Prime holds a position in Small Autonomous Motherfucker ($SAM) and has invested in Frodobots.