# How BitRobot Crowdsources Real-World Data for Embodied AI, with Jonathan Victor
BitRobot crowdsources real-world robotics data through a network of contributors, addressing the scarcity of embodied AI training data by gamifying robot control and deploying wearable devices like RoboCap.
Video
Video available
Modern AI relies on vast datasets, but real-world robotics requires physical interaction data that cannot be sourced online. BitRobot, founded by Jonathan Victor, tackles this bottleneck by turning robot control into a game where users drive sidewalk robots in scavenger hunts, generating navigation data from human-like interactions. The company open-sourced FrodoBots-2K, a 2,000-hour urban navigation dataset, which expanded public research resources from 60 hours to 30 times that volume overnight, used by teams at DeepMind, Meta, and UC Berkeley.
Victor argues that traditional teleoperation factories are slow and inefficient, producing only a few dozen usable demonstrations daily. Instead, BitRobot operates as a decentralized market where contributors collect diverse robotics data—urban navigation, manipulation, dexterity—through methods like teleoperation, simulation, and egocentric video. The RoboCap, a six-camera wearable priced around $1,000, enables first-person data capture in real-world settings like bakeries or factories, with contributions scored, anonymized, and paid retroactively based on data novelty.
The network prioritizes rare, high-entropy data over repetitive tasks, ensuring contributors are rewarded for capturing unique interactions. BitRobot uses Solana for transparent accounting and payments, leveraging its low-cost infrastructure and community tools. Victor emphasizes that tokens alone do not sustain the network; commercial contracts with labs are secured first to ensure demand before scaling supply, positioning BitRobot to own the strategic infrastructure of embodied AI data.
The debate centers on whether distributed, messy data collection offers a structural advantage over centralized, controlled datasets. BitRobot’s model bets on diversity and the long tail of real-world interactions as superior training material for embodied AI, arguing that a network open to all contributors will outperform isolated teleoperation farms.