episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that we have not been able to find the math
https://karpathy.ai/lexicap/0010-large.html#00:27:21.360
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that allows you to leapfrog ahead.
https://karpathy.ai/lexicap/0010-large.html#00:27:23.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And we are kind of making gradual progress
https://karpathy.ai/lexicap/0010-large.html#00:27:25.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
one step at a time, a new experiment here,
https://karpathy.ai/lexicap/0010-large.html#00:27:28.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
a new experiment there that gives us new insights
https://karpathy.ai/lexicap/0010-large.html#00:27:30.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and gradually building up,
https://karpathy.ai/lexicap/0010-large.html#00:27:32.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but not getting to something yet where we're just,
https://karpathy.ai/lexicap/0010-large.html#00:27:34.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
okay, here's an equation that now explains how,
https://karpathy.ai/lexicap/0010-large.html#00:27:36.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you know, that would be,
https://karpathy.ai/lexicap/0010-large.html#00:27:39.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
have been two years of experimentation to get there,
https://karpathy.ai/lexicap/0010-large.html#00:27:40.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but this tells us what the result's going to be.
https://karpathy.ai/lexicap/0010-large.html#00:27:42.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Unfortunately, not so much yet.
https://karpathy.ai/lexicap/0010-large.html#00:27:45.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Not so much yet, but your hope is there.
https://karpathy.ai/lexicap/0010-large.html#00:27:47.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
In trying to teach robots or systems
https://karpathy.ai/lexicap/0010-large.html#00:27:50.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to do everyday tasks or even in simulation,
https://karpathy.ai/lexicap/0010-large.html#00:27:53.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
what do you think you're more excited about?
https://karpathy.ai/lexicap/0010-large.html#00:27:58.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Imitation learning or self play?
https://karpathy.ai/lexicap/0010-large.html#00:28:02.740
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So letting robots learn from humans
https://karpathy.ai/lexicap/0010-large.html#00:28:04.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or letting robots plan their own
https://karpathy.ai/lexicap/0010-large.html#00:28:08.700
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to try to figure out in their own way
https://karpathy.ai/lexicap/0010-large.html#00:28:11.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and eventually play, eventually interact with humans
https://karpathy.ai/lexicap/0010-large.html#00:28:13.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or solve whatever the problem is.
https://karpathy.ai/lexicap/0010-large.html#00:28:18.320
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What's the more exciting to you?
https://karpathy.ai/lexicap/0010-large.html#00:28:20.180
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What's more promising you think as a research direction?
https://karpathy.ai/lexicap/0010-large.html#00:28:21.860
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So when we look at self play,
https://karpathy.ai/lexicap/0010-large.html#00:28:24.660
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
what's so beautiful about it is goes back
https://karpathy.ai/lexicap/0010-large.html#00:28:32.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to kind of the challenges in reinforcement learning.
https://karpathy.ai/lexicap/0010-large.html#00:28:34.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So the challenge of reinforcement learning
https://karpathy.ai/lexicap/0010-large.html#00:28:37.260
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
is getting signal.
https://karpathy.ai/lexicap/0010-large.html#00:28:38.460
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And if you don't never succeed, you don't get any signal.
https://karpathy.ai/lexicap/0010-large.html#00:28:40.580
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
In self play, you're on both sides.
https://karpathy.ai/lexicap/0010-large.html#00:28:43.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So one of you succeeds.
https://karpathy.ai/lexicap/0010-large.html#00:28:46.740
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And the beauty is also one of you fails.
https://karpathy.ai/lexicap/0010-large.html#00:28:48.020
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so you see the contrast.
https://karpathy.ai/lexicap/0010-large.html#00:28:49.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
You see the one version of me that did better
https://karpathy.ai/lexicap/0010-large.html#00:28:51.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
than the other version.
https://karpathy.ai/lexicap/0010-large.html#00:28:53.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So every time you play yourself, you get signal.
https://karpathy.ai/lexicap/0010-large.html#00:28:54.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so whenever you can turn something into self play,
https://karpathy.ai/lexicap/0010-large.html#00:28:57.260
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you're in a beautiful situation
https://karpathy.ai/lexicap/0010-large.html#00:29:00.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where you can naturally learn much more quickly
https://karpathy.ai/lexicap/0010-large.html#00:29:02.080
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
than in most other reinforcement learning environments.
https://karpathy.ai/lexicap/0010-large.html#00:29:04.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So I think if somehow we can turn more
https://karpathy.ai/lexicap/0010-large.html#00:29:07.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
reinforcement learning problems
https://karpathy.ai/lexicap/0010-large.html#00:29:12.460
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
into self play formulations,
https://karpathy.ai/lexicap/0010-large.html#00:29:13.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that would go really, really far.
https://karpathy.ai/lexicap/0010-large.html#00:29:15.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So far, self play has been largely around games
https://karpathy.ai/lexicap/0010-large.html#00:29:17.180
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where there is natural opponents.
https://karpathy.ai/lexicap/0010-large.html#00:29:20.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But if we could do self play for other things,
https://karpathy.ai/lexicap/0010-large.html#00:29:22.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and let's say, I don't know,
https://karpathy.ai/lexicap/0010-large.html#00:29:24.740
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
a robot learns to build a house.
https://karpathy.ai/lexicap/0010-large.html#00:29:25.580
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, that's a pretty advanced thing
https://karpathy.ai/lexicap/0010-large.html#00:29:26.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to try to do for a robot,
https://karpathy.ai/lexicap/0010-large.html#00:29:28.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but maybe it tries to build a hut or something.
https://karpathy.ai/lexicap/0010-large.html#00:29:29.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
If that can be done through self play,
https://karpathy.ai/lexicap/0010-large.html#00:29:31.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it would learn a lot more quickly
https://karpathy.ai/lexicap/0010-large.html#00:29:34.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
if somebody can figure that out.
https://karpathy.ai/lexicap/0010-large.html#00:29:35.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I think that would be something
https://karpathy.ai/lexicap/0010-large.html#00:29:36.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where it goes closer to kind of the mathematical leapfrogging
https://karpathy.ai/lexicap/0010-large.html#00:29:37.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where somebody figures out a formalism to say,
https://karpathy.ai/lexicap/0010-large.html#00:29:41.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
okay, any RL problem by playing this and this idea,
https://karpathy.ai/lexicap/0010-large.html#00:29:43.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you can turn it into a self play problem
https://karpathy.ai/lexicap/0010-large.html#00:29:47.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where you get signal a lot more easily.
https://karpathy.ai/lexicap/0010-large.html#00:29:48.700
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Reality is, many problems we don't know
https://karpathy.ai/lexicap/0010-large.html#00:29:50.740
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
how to turn into self play.
https://karpathy.ai/lexicap/0010-large.html#00:29:52.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so either we need to provide detailed reward.
https://karpathy.ai/lexicap/0010-large.html#00:29:53.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That doesn't just reward for achieving a goal,
https://karpathy.ai/lexicap/0010-large.html#00:29:56.980
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but rewards for making progress,
https://karpathy.ai/lexicap/0010-large.html#00:29:58.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and that becomes time consuming.
https://karpathy.ai/lexicap/0010-large.html#00:30:00.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And once you're starting to do that,
https://karpathy.ai/lexicap/0010-large.html#00:30:02.660
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
let's say you want a robot to do something,
https://karpathy.ai/lexicap/0010-large.html#00:30:03.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you need to give all this detailed reward.
https://karpathy.ai/lexicap/0010-large.html#00:30:05.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Well, why not just give a demonstration?
https://karpathy.ai/lexicap/0010-large.html#00:30:07.180
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Because why not just show the robot?
https://karpathy.ai/lexicap/0010-large.html#00:30:09.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And now the question is, how do you show the robot?
https://karpathy.ai/lexicap/0010-large.html#00:30:11.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
One way to show is to tally operate the robot,
https://karpathy.ai/lexicap/0010-large.html#00:30:14.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and then the robot really experiences things.
https://karpathy.ai/lexicap/0010-large.html#00:30:16.620
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And that's nice, because that's really high signal
https://karpathy.ai/lexicap/0010-large.html#00:30:19.020
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
to noise ratio data, and we've done a lot of that.
https://karpathy.ai/lexicap/0010-large.html#00:30:21.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And you teach your robot skills in just 10 minutes,
https://karpathy.ai/lexicap/0010-large.html#00:30:23.060
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
you can teach your robot a new basic skill,
https://karpathy.ai/lexicap/0010-large.html#00:30:26.020
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
like okay, pick up the bottle, place it somewhere else.
https://karpathy.ai/lexicap/0010-large.html#00:30:27.860
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That's a skill, no matter where the bottle starts,
https://karpathy.ai/lexicap/0010-large.html#00:30:30.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
maybe it always goes onto a target or something.
https://karpathy.ai/lexicap/0010-large.html#00:30:32.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
That's fairly easy to teach your robot with tally up.
https://karpathy.ai/lexicap/0010-large.html#00:30:34.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Now, what's even more interesting
https://karpathy.ai/lexicap/0010-large.html#00:30:38.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
if you can now teach your robot
https://karpathy.ai/lexicap/0010-large.html#00:30:40.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
through third person learning,
https://karpathy.ai/lexicap/0010-large.html#00:30:41.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
where the robot watches you do something
https://karpathy.ai/lexicap/0010-large.html#00:30:43.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and doesn't experience it, but just kind of watches you.
https://karpathy.ai/lexicap/0010-large.html#00:30:45.700
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It doesn't experience it, but just watches it
https://karpathy.ai/lexicap/0010-large.html#00:30:48.500
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and says, okay, well, if you're showing me that,
https://karpathy.ai/lexicap/0010-large.html#00:30:49.820
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that means I should be doing this.
https://karpathy.ai/lexicap/0010-large.html#00:30:52.180
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I'm not gonna be using your hand,
https://karpathy.ai/lexicap/0010-large.html#00:30:53.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
because I don't get to control your hand,
https://karpathy.ai/lexicap/0010-large.html#00:30:55.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but I'm gonna use my hand, I do that mapping.
https://karpathy.ai/lexicap/0010-large.html#00:30:57.100
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And so that's where I think one of the big breakthroughs
https://karpathy.ai/lexicap/0010-large.html#00:30:59.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
has happened this year.
https://karpathy.ai/lexicap/0010-large.html#00:31:02.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
This was led by Chelsea Finn here.
https://karpathy.ai/lexicap/0010-large.html#00:31:03.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It's almost like learning a machine translation
https://karpathy.ai/lexicap/0010-large.html#00:31:06.460
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
for demonstrations, where you have a human demonstration,
https://karpathy.ai/lexicap/0010-large.html#00:31:08.280