#43 Chris Paxton & Michael Cho: Robopapers And Future Of Robot Learning

Robots will entertain long before they help at home

The bottleneck in robot learning is the physical world itself. Data is expensive to collect, results are hard to verify, and failures compound once a machine leaves the lab. The first market that works today is entertainment, where a cheap humanoid is enough, and Michael Cho's BitRobot is testing whether crypto incentives can fund the work that homes will eventually require.

Takeaways

• Keeping robots working is the hardest part. Chris Paxton says the challenge is getting robots into the real world and keeping them running, with data a by-product. Early in his career a grad student slammed a hand-tuned robot arm into a wall and broke it.
• Robotics lacks data and real-world evaluation. Michael Cho names two bottlenecks, a lack of robotics data and real-world evaluation. You cannot speed up time or verify a paper's reported success rate, and there are no benchmarks because every robot and every house differs.
• Teleoperation is the gold standard and unscalable. Michael Cho calls teleoperation the best data but costly, because robots break and cannot respawn, and a $10,000 robot will not scale as a collection device. Chris Paxton says the best data is what a robot gathers itself.
• Human video is essential but incomplete. Michael Cho says video gives a broad concept of a task but cannot be the only input, because human motion cannot be recorded at 30 hertz with exact angles and velocities. Millimetres and 50 milliseconds matter.
• Simulation falls short of what robots need. Michael Cho guesses simulation gets robotics 80 to 90 percent of the way, but the bar is 99.9999 percent. He blames the near-infinite search space and handcrafted reward functions rather than compute.
• Entertainment is the first real robot market. Michael Cho argues robots are funny before they are useful, and Unitree's R1 humanoid at $6,000 to $7,000 will put them in bars and discos, which is how the first 100,000 humanoids ship. Chris Paxton notes a $20,000 Unitree robot at Walmart.
• Crypto incentives could fund embodied AI research. Michael Cho's BitRobot uses crypto incentives to gather resources for embodied AI without raising money, a different form of capital formation from the limited liability company. He thinks a few billion dollars could be gathered, enough to accelerate researchers.

What comes next

• Both hosts expect robots to arrive later than people think, and Chris Paxton says doing robotics for so long makes him bearish by default.
• Chris Paxton is coming around on egocentric video, and says reinforcement learning and reasoning for robots excite him most.