We discovered an emergent property of VLAs like π0/π0.5/π0.6: as we scale up pre-training, the model learns to align human videos and robot data! This gives us a simple way to leverage human videos. Once π0.5 knows how to control robots, it can naturally learn from human video.

Continuous learning is the only defensible moat.

~80% of UMI gloves sit idle at any given time. I dug into two robotics labs using universal manipulation interfaces (Generalist + Sunday) to learn more about how they've operationalized data collection. Everyone focuses on scaling operators, hardware, and cheaper

. Praneet and I taught a robot to play Ms. Pacman Our learnings and takeaways below (with full report/code at the end) 👇 p.s. sound on!

Just returned from my first trip to China, mostly looking at the energy and robotics industries. Fascinating. Random observations, both business and general, below... 1/x

NEW: I tracked all the Chinese lithium battery factories announced or built in the past decade 𝗼𝘂𝘁𝘀𝗶𝗱𝗲 China's borders. The truth is, companies like CATL and BYD are so technologically advanced that there's little international competition. As batteries receive global

Was a great way to repay my grandma for years of cleaning up after me... Go team!! Mehul

If you’re deploying robots in the real world, let’s get in touch

"intention density" is behind the visceral difference between AI outputs that feel beautiful, human, designed vs. uninspired/slop it points at something much more specific than taste: how many distinct, willful decisions went into an output? how much of its structure can be