episode
stringlengths
45
100
text
stringlengths
1
528
timestamp_link
stringlengths
56
56
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
let me know.
https://karpathy.ai/lexicap/0010-large.html#00:01:00.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
In terms of getting a robot to beat him at tennis,
https://karpathy.ai/lexicap/0010-large.html#00:01:01.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it's kind of an interesting question
https://karpathy.ai/lexicap/0010-large.html#00:01:07.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
because for a lot of the challenges we think about in AI,
https://karpathy.ai/lexicap/0010-large.html#00:01:08.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
the software is really the missing piece,
https://karpathy.ai/lexicap/0010-large.html#00:01:14.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but for something like this,
https://karpathy.ai/lexicap/0010-large.html#00:01:16.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
the hardware is nowhere near either.
https://karpathy.ai/lexicap/0010-large.html#00:01:18.620
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
To really have a robot that can physically run around,
https://karpathy.ai/lexicap/0010-large.html#00:01:22.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
the Boston Dynamics robots are starting to get there,
https://karpathy.ai/lexicap/0010-large.html#00:01:26.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but still not really human level ability to run around
https://karpathy.ai/lexicap/0010-large.html#00:01:28.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and then swing a racket.
https://karpathy.ai/lexicap/0010-large.html#00:01:33.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So you think that's a hardware problem?
https://karpathy.ai/lexicap/0010-large.html#00:01:36.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I don't think it's a hardware problem only.
https://karpathy.ai/lexicap/0010-large.html#00:01:38.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think it's a hardware and a software problem.
https://karpathy.ai/lexicap/0010-large.html#00:01:39.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think it's both.
https://karpathy.ai/lexicap/0010-large.html#00:01:41.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
And I think they'll have independent progress.
https://karpathy.ai/lexicap/0010-large.html#00:01:43.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So I'd say the hardware maybe in 10, 15 years.
https://karpathy.ai/lexicap/0010-large.html#00:01:45.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
On clay, not grass.
https://karpathy.ai/lexicap/0010-large.html#00:01:51.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, grass is probably harder.
https://karpathy.ai/lexicap/0010-large.html#00:01:52.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
With the sliding?
https://karpathy.ai/lexicap/0010-large.html#00:01:53.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Yeah.
https://karpathy.ai/lexicap/0010-large.html#00:01:54.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
With the clay, I'm not sure what's harder, grass or clay.
https://karpathy.ai/lexicap/0010-large.html#00:01:55.420
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
The clay involves sliding,
https://karpathy.ai/lexicap/0010-large.html#00:01:58.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which might be harder to master actually, yeah.
https://karpathy.ai/lexicap/0010-large.html#00:02:01.600
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
But you're not limited to a bipedal.
https://karpathy.ai/lexicap/0010-large.html#00:02:06.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, I'm sure there's no...
https://karpathy.ai/lexicap/0010-large.html#00:02:08.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Well, if we can build a machine,
https://karpathy.ai/lexicap/0010-large.html#00:02:09.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it's a whole different question, of course.
https://karpathy.ai/lexicap/0010-large.html#00:02:11.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
If you can say, okay, this robot can be on wheels,
https://karpathy.ai/lexicap/0010-large.html#00:02:13.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it can move around on wheels and can be designed differently,
https://karpathy.ai/lexicap/0010-large.html#00:02:16.300
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
then I think that can be done sooner probably
https://karpathy.ai/lexicap/0010-large.html#00:02:19.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
than a full humanoid type of setup.
https://karpathy.ai/lexicap/0010-large.html#00:02:23.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What do you think of swing a racket?
https://karpathy.ai/lexicap/0010-large.html#00:02:26.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So you've worked at basic manipulation.
https://karpathy.ai/lexicap/0010-large.html#00:02:27.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
How hard do you think is the task of swinging a racket
https://karpathy.ai/lexicap/0010-large.html#00:02:31.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
would be able to hit a nice backhand or a forehand?
https://karpathy.ai/lexicap/0010-large.html#00:02:34.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Let's say we just set up stationary,
https://karpathy.ai/lexicap/0010-large.html#00:02:39.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
a nice robot arm, let's say, a standard industrial arm,
https://karpathy.ai/lexicap/0010-large.html#00:02:42.720
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and it can watch the ball come and then swing the racket.
https://karpathy.ai/lexicap/0010-large.html#00:02:46.580
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It's a good question.
https://karpathy.ai/lexicap/0010-large.html#00:02:50.700
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I'm not sure it would be super hard to do.
https://karpathy.ai/lexicap/0010-large.html#00:02:51.540
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, I'm sure it would require a lot,
https://karpathy.ai/lexicap/0010-large.html#00:02:56.200
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
if we do it with reinforcement learning,
https://karpathy.ai/lexicap/0010-large.html#00:02:58.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it would require a lot of trial and error.
https://karpathy.ai/lexicap/0010-large.html#00:03:00.000
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It's not gonna swing it right the first time around,
https://karpathy.ai/lexicap/0010-large.html#00:03:01.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but yeah, I don't see why I couldn't
https://karpathy.ai/lexicap/0010-large.html#00:03:03.380
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
swing it the right way.
https://karpathy.ai/lexicap/0010-large.html#00:03:07.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think it's learnable.
https://karpathy.ai/lexicap/0010-large.html#00:03:09.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think if you set up a ball machine,
https://karpathy.ai/lexicap/0010-large.html#00:03:10.340
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
let's say on one side,
https://karpathy.ai/lexicap/0010-large.html#00:03:12.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and then a robot with a tennis racket on the other side,
https://karpathy.ai/lexicap/0010-large.html#00:03:13.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think it's learnable
https://karpathy.ai/lexicap/0010-large.html#00:03:17.780
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and maybe a little bit of pre training and simulation.
https://karpathy.ai/lexicap/0010-large.html#00:03:20.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Yeah, I think that's feasible.
https://karpathy.ai/lexicap/0010-large.html#00:03:22.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I think the swing the racket is feasible.
https://karpathy.ai/lexicap/0010-large.html#00:03:25.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
It'd be very interesting to see how much precision
https://karpathy.ai/lexicap/0010-large.html#00:03:27.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it can get.
https://karpathy.ai/lexicap/0010-large.html#00:03:28.900
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Cause I mean, that's where, I mean,
https://karpathy.ai/lexicap/0010-large.html#00:03:31.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
some of the human players can hit it on the lines,
https://karpathy.ai/lexicap/0010-large.html#00:03:35.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which is very high precision.
https://karpathy.ai/lexicap/0010-large.html#00:03:37.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
With spin, the spin is an interesting,
https://karpathy.ai/lexicap/0010-large.html#00:03:39.240
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
whether RL can learn to put a spin on the ball.
https://karpathy.ai/lexicap/0010-large.html#00:03:42.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Well, you got me interested.
https://karpathy.ai/lexicap/0010-large.html#00:03:45.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Maybe someday we'll set this up.
https://karpathy.ai/lexicap/0010-large.html#00:03:46.880
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Sure, you got me intrigued.
https://karpathy.ai/lexicap/0010-large.html#00:03:48.400
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Your answer is basically, okay,
https://karpathy.ai/lexicap/0010-large.html#00:03:51.120
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
for this problem, it sounds fascinating,
https://karpathy.ai/lexicap/0010-large.html#00:03:52.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but for the general problem of a tennis player,
https://karpathy.ai/lexicap/0010-large.html#00:03:54.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
we might be a little bit farther away.
https://karpathy.ai/lexicap/0010-large.html#00:03:56.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
What's the most impressive thing you've seen a robot do
https://karpathy.ai/lexicap/0010-large.html#00:03:58.560
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
in the physical world?
https://karpathy.ai/lexicap/0010-large.html#00:04:01.260
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So physically for me,
https://karpathy.ai/lexicap/0010-large.html#00:04:04.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
it's the Boston Dynamics videos.
https://karpathy.ai/lexicap/0010-large.html#00:04:06.480
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Always just bring home and just super impressed.
https://karpathy.ai/lexicap/0010-large.html#00:04:10.920
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Recently, the robot running up the stairs,
https://karpathy.ai/lexicap/0010-large.html#00:04:15.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
doing the parkour type thing.
https://karpathy.ai/lexicap/0010-large.html#00:04:17.700
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I mean, yes, we don't know what's underneath.
https://karpathy.ai/lexicap/0010-large.html#00:04:19.440
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
They don't really write a lot of detail,
https://karpathy.ai/lexicap/0010-large.html#00:04:22.280
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
but even if it's hard coded underneath,
https://karpathy.ai/lexicap/0010-large.html#00:04:23.940
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which it might or might not be just the physical abilities
https://karpathy.ai/lexicap/0010-large.html#00:04:27.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
of doing that parkour, that's a very impressive.
https://karpathy.ai/lexicap/0010-large.html#00:04:29.800
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So have you met Spot Mini
https://karpathy.ai/lexicap/0010-large.html#00:04:32.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
or any of those robots in person?
https://karpathy.ai/lexicap/0010-large.html#00:04:34.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Met Spot Mini last year in April at the Mars event
https://karpathy.ai/lexicap/0010-large.html#00:04:36.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that Jeff Bezos organizes.
https://karpathy.ai/lexicap/0010-large.html#00:04:41.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
They brought it out there
https://karpathy.ai/lexicap/0010-large.html#00:04:42.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and it was nicely following around Jeff.
https://karpathy.ai/lexicap/0010-large.html#00:04:44.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
When Jeff left the room, they had it follow him along,
https://karpathy.ai/lexicap/0010-large.html#00:04:47.760
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
which is pretty impressive.
https://karpathy.ai/lexicap/0010-large.html#00:04:50.640
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
So I think there's some confidence to know
https://karpathy.ai/lexicap/0010-large.html#00:04:52.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
that there's no learning going on in those robots.
https://karpathy.ai/lexicap/0010-large.html#00:04:55.680
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
The psychology of it, so while knowing that,
https://karpathy.ai/lexicap/0010-large.html#00:04:58.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
while knowing there's not,
https://karpathy.ai/lexicap/0010-large.html#00:05:00.160
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
if there's any learning going on, it's very limited.
https://karpathy.ai/lexicap/0010-large.html#00:05:01.140
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
I met Spot Mini earlier this year
https://karpathy.ai/lexicap/0010-large.html#00:05:04.040
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
and knowing everything that's going on,
https://karpathy.ai/lexicap/0010-large.html#00:05:06.840
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
having one on one interaction,
https://karpathy.ai/lexicap/0010-large.html#00:05:09.520
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
so I got to spend some time alone and there's immediately
https://karpathy.ai/lexicap/0010-large.html#00:05:11.000
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
a deep connection on the psychological level.
https://karpathy.ai/lexicap/0010-large.html#00:05:15.960
Pieter Abbeel: Deep Reinforcement Learning | Lex Fridman Podcast #10
Even though you know the fundamentals, how it works,
https://karpathy.ai/lexicap/0010-large.html#00:05:18.640