Ep. 21: BitRobot - Creating the World's Open Robotics Lab

BitRobot bets cheap robots and crypto can feed embodied AI

BitRobot is building an open robotics lab on the bet that embodied AI is bottlenecked by data. Crypto incentives let contributors worldwide fund and operate cheap robots, gathering real-world data through subnets that pursue different research bets in parallel. The network has produced the largest dataset of its kind and a humanoid fighting league spun out of its studio.

Takeaways

• Cheap robot data rivals data from costly robots. Michael Cho and his brothers built a dumb sidewalk robot during the pandemic, hacking on Saturdays for six months. His hypothesis was that data from a sub-$1,000 machine is as valuable as data from a $10,000 robot, because embodied AI is bottlenecked by data.
• Crypto incentives can crowdsource robot data at scale. Helium sold hundreds of thousands of units at about $500 each, crowdsourcing hardware cost and global deployment. Michael Cho was skeptical of crypto until he saw that, and now argues permissionless incentivization lets contributors worldwide fund and operate robots instead of one company paying upfront.
• BitRobot subnets gather teleoperation and video data. Amanda says the network collects teleoperation and video data from real robots and runs simulations. Michael Cho says robot diversity matters as much as environment diversity. Subnet one is a cloud-based game where players drive real robots over the internet.
• Third-party subnets extend BitRobot's data network. Virtuals launched Seesaw a month earlier to collect egocentric video shot on iPhones, and it has received almost 300,000 tasks. Amanda says BitRobot wants robotics startups and manufacturers to launch their own subnets, so the lab grows beyond what one team can build.
• Humans still beat AI in Earth Rover. The Earth Rover challenge pits top human gamers against AI models from university labs including UT Austin and Seoul National University. DeepMind and Meta AI researchers are co-organizers. Michael Cho says the humans still win, and published academic papers are already using the dataset.
• Reinforcement learning is moving into real robots. Michael Cho cites Physical Intelligence's work as an example. He argues the current collect-train-deploy-retrain paradigm is not how animals or humans learn, which is by learning as they go. He also points to egocentric video boosted by Meta's ARIA glasses and fast-improving humanoids.

UFB turns humanoid combat into a media business

1. BitRobot incubated UFB, a humanoid fighting league that has sold out shows in San Francisco and Los Angeles.
2. The New York Times and the Washington Post have covered the fights, and UFB is now working with creators such as Wiz Khalifa, who are piloting the bots.