Here's a little more, and no cables this time ;) It's all real-time speed, zero CGI, fully learned in Simulation & zero-shot transferred to real. Besides the fact that it's fun, we had to make significant improvements to our robot model in Sim, domain randomization and other techniques which will directly transfer to more practical situations as well (robust walking, and agile full-body control in general). We also learned quite a few things on our hardware and tuned our power profile. I hope it is getting pretty clear that Tesla is much more than an amazing car (!) company: it's the place to be for transformational real-world AI.

More coming shortly! The team has been working very hard in the background. Entirely trained in simulation with RL. Many optimizations and fixes have been put in place in our sim-to-real training code. The cable is there in case of a fall (not actually holding the bot) since this is an early result which will get more stable very quickly.

This weekend, I drove to Napa and back - close to 200 miles - on FSD. I think I had max a couple interventions and they were both comfort-related, meaning not necessary per se. My car took me through my neighborhood’s winding, hilly streets, got on the highway, navigated San Francisco’s weekend road mess, crossed the Bay Bridge, stopped at a supercharger, and then headed straight to my destination in Napa. The way back was just as clean, this time over the Golden Gate Bridge. I had relatives with me and they couldn’t believe it; they said it drove smoother than any of us (thanks, Dad! … ). Most people just haven’t realized yet that this technology actually exists now, and you can get it on your own Tesla car. The moment it goes entirely unsupervised will be wild. Massive engineering push for a decade by the team. @aelluswamy