@startups 🍿

"co-founder Sergey Levine described the company’s ambition simply: “Think of it like ChatGPT, but for robots.”"

"Co-founder Lachy Groom told TechCrunch the company has no timeline for commercialization....“There’s no limit to how much money we can really put to work,” Groom said. “There’s always more compute you can throw at the problem.”

Yes, we have evidence that if you throw enough money at it, eventually any approach to designing robot behavior in tandem with hardware will enable enough iteration that the robot will work (about $200m of DoD money between 2008 and 2014 got Boston Dynamics off the ground, but it was still unprofitable in 2019). $11b is plenty, given enough time.

Yes, new simulation environments (Omniverse) are revolutionizing controller development by turning "training in simulation" into "a robust ability to perform a function in the world (like "walking on a flat surface") for a specific piece of hardware" (leap of months to minutes of work given years of sunk cost).

But generalizing from
"train a digital twin of my specific quadruped to walk on a flat surface in response to a remote control"
to
"reliable general purpose controller that will do any requested task on my specific hardware"
is a leap of years to minutes from where we are.

And the leap from that to
"general purpose controller that will work on any hardware to autonomously accomplish any task well"
is another leap in kind.

Count me skeptical.

#Robotics