Wow, their claims are that ordinary third-person video may be nearly as useful as expensive teleoperated robot data for pretraining. If so this is certainly a holy grail moment.

Plus they're open-weighting their World Action Model (!!)

Ordinarily I'd be very skeptical but Runway is a $5B company, so maybe this is what we've been waiting for re general-purpose robotic pretraining?

Does anyone have experience with Praxis-1 yet?