00:01The dream of a robot foundation model makes possible that you can just walk up to a robot and get
00:09it to do any task almost immediately and generalize that behavior.
00:14Our new model, Gen 1.5, is an immediate learning generalist. It's a one-shot learner.
00:21It can learn and generalize new tasks in seconds, prompted in context.
00:25It can compose prompts together to learn longer horizon tasks.
00:29It can be prompted from a simulator and transfer those behaviors into the real world.
00:35And in some cases, it can just watch humans and mimic them immediately on the spot.
00:41For few-shot learning, it can learn new tasks with as few as 1 to 10 gradient steps on 1
00:47to 5 minutes of data.
00:49And it's showing physical generalization, including improvising how to use new tools to accomplish tasks.
00:57And it does all of these things out of the box.
01:00The fastest way it learns is with zero training on a new task and just a few seconds of demonstration
01:07data put into the model's context.
01:09We show the robot what to do, and it generalizes it.
01:14We call this physical prompting.
01:16We've done variations of in-context prompting for about a year now, but what we're seeing for the first time
01:23is a very generalized in-context learning capability.
01:26These tasks are simple today, and while the success rates are not super high, we've never seen a model that
01:34can do this before.
01:35It's not something we explicitly trained for, but has emerged from our new training recipe.
01:42Beyond its context, it can also learn with extremely few training steps, and among the new levels of physical generalization
01:51we are seeing is novel tool use improvisation.
01:55In one example, we taught it to use a brush to sweep a block into a bowl.
02:00And when we gave it a banana instead, it could figure out how to turn the banana into an impromptu
02:08brush.
02:09And when we gave it a dustpan, it used its other hand to nudge the block onto the dustpan and
02:16then raise the block up and dumped it into the bowl.
02:20It could also sweep multiple objects and figure out how to switch hands ambidextrously.
02:26It's overall showing more improvisational intelligence than previous models.
02:32It had a Lego brick stuck on one of its hands, and it used its other hand to remove it.
02:37We had taught it to put blocks into a bowl, but then we covered the bowl with a piece of
02:42paper, and it removed the piece of paper to accomplish its task.
02:46We taught it to open a jar lid with one hand, and then it just did it with two hands.
02:52And when we gave it different bottles and cups, it figured out how to open those too.
02:58We're also starting to see human-to-robot in context learning emerge, where a person can just show the robot
03:07what to do with their own human hands, and the robot mimics it on the spot with its hands.
03:13We are still in the very early days, but we're tremendously excited about the potential of this new frontier of
03:23intelligence in robot foundation models.
03:26We're still in the very early days, and we're still in the very early days.
03:27We're still in the very early days, and we're still in the very early days.
03:28We're still in the very early days, and we're still in the very early days.
03:29We're still in the very early days, and we're still in the very early days.
03:32We're still in the very early days, and we're still in the very early days.
03:32We're still in the very early days, and we're still in the very early days.
Comentarios