This episode unpacks the critical challenges hindering embodied AI progress—data scarcity and real-world validation—revealing how BitRobot Network leverages crypto incentives to crowdsource data and build foundational robotics models.
The Convergence of AI and Robotics: Defining Embodied AI
- Michael Cho begins by noting a significant shift in investor sentiment towards robotics over the last 1-2 years, moving it from a "dirty vertical" to one attracting substantial funding.
- He attributes much of this excitement to advancements in AI, particularly the Transformer architecture.
- The Transformer architecture is a type of neural network initially used for natural language processing (like in LLMs) but has proven flexible enough to handle various data types (modalities).
- Michael suggests robotics data can be treated as just another modality for these powerful models, creating the field often referred to as embodied or physical AI – the intersection where AI controls physical systems.
Data Scarcity: The First Bottleneck in Embodied AI
- Despite the potential of AI architectures, Michael highlights a fundamental problem: the lack of readily available robotics data.
- Unlike text or images scraped from the internet, data capturing how robots interact with the physical world is scarce.
- "The big challenge for robotics is that this robotics data doesn't really exist in the natural world," Michael states, identifying this as a primary bottleneck that projects like BitRobot aim to solve through novel collection methods.
Real-World Evaluation: The Critical Distinction for Embodied AI
- Michael argues that beyond data collection, an even greater challenge for embodied AI is real-world evaluation, distinguishing it sharply from purely digital AI.
- Digital AI models (like LLMs) can be benchmarked rapidly, often within hours of release, because evaluation occurs entirely in the digital realm.
- However, assessing a robotic AI model requires testing it physically.
- Michael uses the example of Tesla's Full Self-Driving (FSD): even millions of miles without failure don't definitively prove Level 5 autonomy (fully autonomous driving under all conditions) because a critical failure could occur on the very next mile.
- This makes real-world failures, unfortunately, the only truly verifiable signal of a model's limitations, rendering the iteration cycle for embodied AI inherently slow and dependent on large physical fleets for testing.
- Actionable Insight: Investors must recognize that progress metrics and development cycles for embodied AI differ fundamentally from digital AI. Scaled real-world deployment and testing infrastructure are critical indicators of project viability.
The Long Road to Physical AGI
- Reflecting on the challenges, Michael suggests that achieving truly capable, general-purpose robotics will take years, drawing parallels to the slow progress in self-driving cars.
- He views robotics and physical AI as potentially the "last frontier" on the path to Artificial General Intelligence (AGI), asserting that cognitive, digital AI capabilities might surpass human levels much sooner (within 2-3 years), leaving physical interaction as the remaining hurdle.
- Strategic Implication: Embodied AI represents a long-term investment thesis requiring patience and an understanding of the incremental, physically-grounded nature of progress.
Mapping vs. Vision-Based Navigation in Robotics
- Addressing the topic of mapping (like geospatial positioning), Michael offers a nuanced perspective.
- While maps and sensor data like GPS or IMU (Inertial Measurement Unit - a device measuring orientation and acceleration) can aid navigation, he argues that truly advanced autonomous systems, like humans, should primarily rely on real-time vision.
- He considers mapping data helpful but "second order," suggesting that the ultimate goal is for models to navigate based purely on observation, making vision-based approaches potentially more critical for achieving robust, generalizable robotic mobility.
Robotics Data vs. Digital Data: Value and Longevity
- Michael contrasts the nature and value of robotics data with digital data.
- He references projects like Grass and their concept of LCR (Live Context Retrieval - accessing real-time internet information for LLMs), noting that the value of new digital information often decays rapidly (minutes).
- Robotics data, however, retains its value far longer.
- Michael explains, "a crash that you record 10 years ago that data is equally worthwhile today because the physical world doesn't change." This longevity applies to various failure cases (e.g., a robot arm failing to fold clothes).
- He also touches upon the utility of video data (like from YouTube) for training basic robotic models, acknowledging its value but also its limitations due to the "Embodiment Gap" – the difficulty of transferring skills learned from observing one embodiment (like a human) to another (like a robot).
- Investor Insight: Datasets containing real-world robotic interaction data represent potentially durable assets with longer intrinsic value compared to transient digital information streams.
The Power of Cross-Embodiment Learning
- Michael introduces "Cross-Embodiment," a key research concept in embodied AI.
- This refers to the finding that training AI models using data collected from diverse types of robots can lead to better, more generalizable performance, even enabling capabilities the model wasn't explicitly trained for (e.g., using data from robotic dogs and arms to train a drone-flying model).
- Michael explains the intuition: "all these different robotic types they ultimately do stuff in the physical world on earth where gravity is the same everywhere." This suggests that data diversity might be as crucial as data volume.
- Strategic Implication: Networks or projects capable of sourcing and integrating data from a wide variety of robot forms (humanoids, drones, arms, rovers) may hold a competitive advantage in building truly foundational robotics models.
Introducing BitRobot Network: From Frotobots to Foundational Models
- Michael details the evolution from Frotobots, initially focused on crowdsourcing data for sidewalk navigation using crypto incentives, to the broader vision of BitRobot Network.
- This shift was driven by interactions with leading AI researchers (from DeepMind, UC Berkeley, Stanford) who faced significant bottlenecks: lack of data, insufficient compute power, and critically, no access to physical robot fleets for real-world model evaluation.
- He shares an anecdote about a top Korean self-driving research team that hadn't tested models in the real world for two years before collaborating with Frotobots.
- The Earth Rover Challenge, a competition organized with DeepMind pitting AI against human gamers (partnered with YGG), further highlighted these needs, requiring Michael to source H100 GPUs (high-performance computing chips essential for AI training) from crypto compute providers Ionet and Exobits.
- This experience solidified the vision for BitRobot as an ecosystem using crypto incentives across multiple "subnets" (specialized networks within the larger network) to address data collection, AI model training, and potentially evaluation, aiming to create foundational models applicable across various robot types.
- This vision led to a collaboration and forthcoming whitepaper co-authored with Jonathan Victor (formerly Filecoin ecosystem lead) and Juan Benet (creator of IPFS - InterPlanetary File System, a peer-to-peer data storage protocol, and founder of Protocol Labs).
Inspiration and Building on Giants: BitTensor, DePIN, and Helium
- Michael explicitly acknowledges the influence of other successful crypto projects on the BitRobot concept.
- He credits BitTensor for demonstrating the viability of a network composed of many specialized subnets working together.
- He also cites Helium and the broader DePIN (Decentralized Physical Infrastructure Networks – crypto-incentivized networks building real-world infrastructure) space as crucial precedents.
- "I think makers we we all just built on you know the shoulders of previous giants," Michael remarks, indicating that BitRobot's ambitious model is conceivable now because these earlier projects paved the way.
Experimenting with Memecoins: The SAM Strategy
- Michael discusses the unconventional strategy of launching SAM, a "sister memecoin," before BitRobot's official token generation event (TGE - the moment a project's primary crypto token becomes publicly available).
- Despite his self-professed initial skepticism about crypto ("I thought all crypto are scams until I discovered Helium"), he now views memecoins as a potential "financial manifestation of a cult or a community."
- The SAM experiment aims to rapidly build community, generate attention, and gather feedback in a low-stakes way, leveraging the fun and speculative side of crypto to ultimately benefit the serious, long-term goals of BitRobot.
- He acknowledges the experimental nature and the need to avoid diluting the core project's focus.
- Strategic Consideration: This highlights an emerging trend where serious infrastructure projects experiment with memecoin mechanics for early community bootstrapping and marketing, a potentially valuable tactic for investors and researchers to observe.
The Long-Term Vision and Current Robotics Applications
- Michael reiterates the long-term nature of the robotics challenge, estimating 5-10 years, potentially longer for truly generalizable humanoid robots capable of replacing human labor.
- Given this timeline, he positions current robotics applications, like the $150 Frotobot vehicle, primarily in the realm of entertainment and gaming.
- In these contexts, robot failures are acceptable and even expected, fitting the "frivolous" nature of memecoins and providing a low-risk environment for current technology.
- Investor Note: Near-term traction for robotics projects may come from entertainment or niche applications before widespread industrial or consumer utility is achieved.
Conclusion: Key Takeaways for Crypto AI Investors and Researchers
- This discussion underscores that embodied AI development is uniquely constrained by physical data acquisition and real-world testing.
- BitRobot Network proposes a crypto-native ecosystem to overcome these hurdles.
- Investors and researchers should monitor BitRobot's ability to incentivize diverse, high-quality data collection and facilitate physical model evaluation.
Where to Learn More
- Twitter: Follow @Frotobots and the new @BitRobotNetwork handle.
- Michael Cho: Follow Michael on Twitter @MIC_OLC.
- Community: Join the Discord (accessible via Twitter links), which hosts web2 gamers, researchers, and increasingly, crypto community members.
- Hardware: The $150 robot is available for purchase (details via project links), currently positioned as a real-life Mario Kart-style RC experience.
