Location via proxy:   [ UP ]  
[Report a bug]   [Manage cookies]                

Commalds

Download as pdf or txt
Download as pdf or txt
You are on page 1of 8

Learning a Driving Simulator

Eder Santana
University of Florida
eder@comma.ai

George Hotz
comma.ai
george@comma.ai

Abstract
comma.ais approach to Artificial Intelligence for self-driving cars1 is based on
an agent that learns to clone driver behaviors and plans maneuvers by simulating
future events in the road. This paper illustrates one of our research approaches
for driving simulation. One where we learn to simulate. Here we investigate variational autoencoders with classical and learned cost functions using generative
adversarial networks for embedding road frames. Afterwards, we learn a transition model in the embedded space using action conditioned Recurrent Neural
Networks. We show that our approach can keep predicting realistic looking video
for several frames despite the transition model being optimized without a cost
function in the pixel space.

Introduction

Self-driving cars are one of the most promising prospects for near term Artificial Intelligence research. The state-of-the-art of this research leverages large amounts of labeled and contextually rich
data, which abounds in driving. From a complex perception and controls perspective, the technology
to correctly solve driving can potentially be extended to other interesting tasks such as action recognition from videos and planning. An approach for self-driving cars that is economically attractive
while still expanding the AI frontier should be based on vision, which is also the main sensor used
by a human driver.
Vision based controls and reinforcement learning had recent success in the literature [7] [8] [9] [10]
mostly due to (deep, recurrent) neural networks and unbounded access to world or game iteraction.
Such interactions provide the possibility to revisit states with new policies and to simulate future
events for training deep neural network based controllers. For example, Alpha Go [9] used deep
convolutional neural networks to predict the probability of winning the game from a given state
using the outcomes of several games, lots of which were playing but previous iterations of Alpha
Go itself. The Go game engine was also used for simulating possible unfoldings of the game and
Markov Chain Tree Search. For learning to play Torcs [7] or Atari [8] using only the game screen,
it was also necessary to use several hours of game play.
For the controls problems where unbounded iteraction of a learning agent with the world is not easily
affordable there are two possible solutions. The solutions are to either hand code a simulator or to
learn to simulate. Hand coding a simulator involves domain expertise for defining rules of physics
and modelling the randomness of the world. But for most cases, such expertise already encompass
all the information for posing the related controls problem, for example flight simulators [11], robot
walking [12], etc.
Here instead we focus on learning to simulate aspects of the world from examples of a human agent.
We focus on generating video streams of a front facing camera mounted in the car windshield.

This work was done during a Summer 2016 internship at comma.ai. The companion material of this
paper can be found online: http://research.comma.ai
1
https://www.youtube.com/watch?v=KTrgRYa2wbI

Figure 1: 80 160 samples from the driving dataset.

Previous successful approaches for learning to simulate for controls were based on state space representations of the agent physics [13]. Other vision only models focused on low dimensional [15] or
videos with simple textures, such as those of Atari games [14] [16]. Texture rich approaches were
based on passive video prediction for action recognition [17] without the possibility of an external
agent modifying the future of the video and without generating a compact representations of the
data.
This paper contributes to this learned video simulation literature. No graphics engine or world model
assumptions are made. We show that it is possible to train a model to do realistic looking video
predictions, while calculating a low dimensional compact representation and action conditioned
transitions. We leave controls on the simulated world for future work. In the next section we
describe the dataset we used to study video prediction of real world road videos.

Dataset

We are publicly releasing part of the driving data used in this paper. The dataset has the same video
and sensors used in the comma.ai self-driving car test platform.
We mounted a Point Grey camera in the windshield of an Acura ILX 2016, capturing pictures of
the road at 20 Hz. In the released dataset there is a total of 7.25 hours of driving data divided in
11 videos. The released video frames are 160 320 pixels regions from the middle of the captured
screen. Beyond video, the dataset also has several sensors that were measured in different frequencies and interpolated to 100Hz. Example data coming from sensors are the car speed, steering angle,
GPS, gyroscope, IMU, etc. Further detail about the dataset and measurement equipment can be
found online in the companion website.
We also record the time stamps at which these sensors were measured and the time stamps the
camera frames were captured. In our production pipeline the measurement times and simple linear
interpolation are used to sync the sensors and camera frames. We are releasing the raw sensors and
camera frames in HDF5 files for easy to use in machine learning and controls software.
In this paper we focus on the camera frames, steering angle and speed. We preprocessed the camera
frames by downsampling them to 80 160 and renormalizing the pixel values between -1 and 1.
Sample images of this transformed dataset are shown in Fig. 1.
In the next section we define the problem investigated in this paper.
2

Problem definition

Let xt denote the t-th frame of the dataset. Xt = {xtn+1 , xtn+2 , ..., xt } denotes a n frames
long video. These frame are associated with control signals St = {stn+1 , stn+2 , ..., st }, At =
{atn+1 , atn+2 , ..., at } corresponding to the ego speeds and steering angles of the car. Learning
to simulate the road can be defined as estimating the function F : R801603n Rn Rn
R801603 that predicts xt+1 = F (Xt , St , At ).
Note that this definition is high dimensional and the dimensions are highly correlated. In Machine
Learning problems like those are known to converge slowly or underfit the data [26]. For example,
previous work [20] showed that when using convolutional dynamic networks without proper regularization, the model simply learns a transformation close to the identity while still getting low cost
scores due to correlations in time. Previous approaches learned the function F directly, but were
demonstrated only in very simple, artificial videos [14]. Recently, it was shown [20] [17] that it is
possible to generate videos with complex textures, but the referred work did not address the problem
of action conditioned transitions and they do not generate a compact intermediate representation of
the data. In other words, their models are fully convolutional with no downsampling and no low
dimensional hidden codes. We believe that compact intermediate representations are important for
our research because probabilities, filtering, and consequently controls, are ill defined in very high
dimensional dense spaces [18].
As far as we know, this is the first paper that tries to learn video predictions from real world highway
scenes. As such, in this first paper, we decided to learn the function F in a piecewise manner, so
we could debug its parts separately. First we learned an autoencoder to embed the frames xt into a
Gaussian latent space zt R2048 , where the dimensionality 2048 was chosen experimentally and the
Gaussian assumption enforced with variational autoencoding Bayes [1]. This first step simplified the
problem from learning transitions directly in the pixel space to learning in the latent space. Beyond
that, assuming that an autoencoder is correctly learned respecting the latent space Gaussianity, we
can generate realistic looking videos as long as the transition model never leaves the high density
region of the embedding space. This high density region is a hypersphere of radius , which in its
turn is function of the dimensionality of embedding space and variance of Gaussian prior. We go
into the details of the autoencoder and the transition model in the next section.

Driving simulator

Due to the problem complexity we decided to learn video prediction with separable networks. We
leave investigations of the end-to-end approach for future work. The proposed architecture is based
on two models: an autoenecoder for dimensionality reduction and an action conditioned RNN for
learning the transitions. Our complete model is depicted in Fig. 2.
4.1

Autoencoder

To learn the data embedding we chose a model that enforced the probability distribution of the
latent space to be Gaussian. Specially to avoid discontinuous regions of low probability inside a
hypersphere centered in the origin. Such discontinuities could make it harder to learn a continuous
transition model in the latent space. Variational autoencoders [1] and related work [19] [21] had
recent sucess learning generative models with Gaussian priors in the latent and original data space.
Unfortunately, the Gaussian assumption in the original data space does not hold for natural images
and VAE predictions usually look blurry (see Fig. 3). On the other hand, generative adversarial networks (GAN) [22] and related work [2] [3] learn the generative model cost function alongside the
generator. They do so by alternating the training of generative and discriminator networks. The generative model transforms samples from the latent space distribution into data from a desired dataset.
The discriminator network tries to tell samples from the desired dataset from images sampled by
the generator. The generator is trained to fool the discriminator, thus the discriminator can be
considered a learned cost function for the generator.
For our purposes we need to learn not only the generator from the latent to the road images space.
But also an encoder from road images back to the latent space. We need to combine VAE and
GANs. Intuitively, one could simply combine the VAE approach with a learned cost function. In
3

image
database

n-dimensional
Gaussian

random
sampling
z

discriminator

fake/legit

generator

encoder
zt

.
.
.

.
.
.

~
zt+1

autoencoder with
generative adversarial cost

.
.
.

transition model
~
zt+n

Figure 2: Driving simulator model: an autoencoder trained with generative adversarial costs coupled
with a recurrent neural network transition model
the literature Donahue et. al [23] proposed bidirectional GANs that learns both generation and encoding as bijective transformations. Lamb et. al. [24] proposed discriminative generative networks
that uses distances between features of a previously trained classifier as part of the cost function.
Finally, Larsen et. al. [25] proposed to train a VAE alongside a GAN, where the encoder optimizes
both the Gaussian prior in the latent space and the similarity between features extracted by the GAN
discriminator network. The generator receives as inputs random samples from the latent space distribution and outputs of the encoder network. The generator is optimized to fool the discriminator
and minimize the similarity between original and decoded images. The discriminator is still only
trained to tell fake from legit images.
We used Larsen et. al. [25] approach to train our autoencoder. In Fig. 2 we show the schematic
diagram of this model as part of our architecture. According to [25] the encoder (Enc), generator
(Gen) and discriminator (Dis) networks are optimized to minimize the following cost function:
l
L = Lprior + LDis
llike + LGAN .

(1)

In (1) Lprior = DKL (q(z|x)||p(z)) is the Kullback-Liebler divergence between the distribution of
the encoder outputs, q(z|x), and a prior distribution, p(z). This is the same VAE regularizer. Here
the prior distribution is p(z) is a Gaussian N (0, 1) and we use the reparametrization trick [1] to
optimize that regularizer. Thus, during training we have z = +  and at test time z = , where
and are outputs of the encoder network and  is a Gaussian random vector with the same number
of dimensions as and .
l
LDis
llike is an error calculated using Disl , the hidden activations of the l-th layer of the discriminator
network, Dis. The activations are calculated using a legit image x and its corresponding encodeddecoded version
Gen(Dis(x)).
Assuming yl = Disl (x) and yl = Disl (Gen(Enc(x))), we have


2
l
LDis
=
E
(y

)
.
During
training, Dis is considered constant with respect to this cost to
l
l
llike
avoid trivial solutions.

Finally, LGAN is the generative adversarial network cost [22]. That cost function represents the
game between Gen and Dis. When training Dis, both Enc and Gen are kept fixed and we have
LDis
GAN = log(Dis(x)) + log(1 Dis(Gen(u))) + log(1 Dis(Gen(Enc(x))),

(2)

where u N (0, 1) is another random variable. The first r.h.t of (2) represents the log-likelihood
of Dis recognizing legit samples and the other two terms represents its log-likelihood to tell fake
samples generated from both random vectors u or codes z = Enc(x). When training Gen, we keep
4

both Dis and Enc fixed and we have


LGen
GAN = log(Dis(Gen(u))) + log(Dis(Gen(Enc(x))),

(3)

which accounts for Gen being able to fool Dis. Following [25], the gradient of LGAN w.r.t Enc is
considered zero during training.
We trained this autoencoder for 200 epochs. Each epoch consisted of 10000 gradient updates with
a batch size of 64. Batches were sampled randomly from driving data as described in the previous
section. We used Adam for optimization [4]. The autoencoder architecture followed Radford et. al.
[3], with the generator made of 4 deconvolutional layers each one followed by batch normalization
and leaky-ReLU activations. The discriminator and encoder consisted of convolutional layers where
each but the first layer was followed by batch normalization. The activation function was ReLU.
Disl is the output of the 3rd convolutional layer of the decoder, before batch normalization and
ReLU are applied. The output size of the discriminator network is 1 and its cost function was binary
cross-entropy. The output size of the encoder network is 2048. This compact representation is
almost 16 times smallet than the original data dimensionality. For network sizes and further details,
please check Fig. 2 and the companion code of this paper. Sample encoded-decoded and target
images are shown in Fig. 3.
After training the autoencoder, we fix all its weights and use Enc as preprocessing step for training
the transition model. We describe the transition model in the next section.
4.2

Transition model

After training the autoencoder described above we obtain the transformed dataset applying Enc :
xt 7 zt . We train RN N : zt , ht , ct 7 zt+1 to represent the transitions in the code space:
ht+1 = tanh (W ht + V zt + U ct ) ,
zt+1 = Aht+1

(4)

where W, V, U, A are trainable weights, ht is the hidden state of the RNN and ct are the concatenated
control speed and angle signal. We leave LSTMs, GRU and multiplicative iterations between ct and
zt for future investigations. The cost function for optimizing the trainable weights is simply the
mean square error:


LRN N = E (zt+1 zt+1 )2 .

(5)

It is easy to see that (5) is optimal because we impose a Gaussian constraint Lprior in the distribution
of the codes z when training the autoencoder. In other words, MSE equals up to a constant factor
the log-likelihood of a normally distributed random variable. Given a predicted code zt+1 we can
estimate future frames as x
t+1 = Gen(
zt+1 ).
We trained a transition model with sequences of length of 15 frames. We used teacher forcing in
the first 5 frames and fed the outputs back as new inputs in the remaining 10 frames. In other
words, z1 , ..., z5 were calculated using Enc(xt ) and fed as input. We fed the outputs z6 , ..., z15
back as input. In the RNN literature feeding outputs back as input is informally referred as RNN
hallucination. It is important to note that we consider the gradient of the outputs equal to zero when
they are fed back as inputs to avoid trivial solutions.

Results

We spent most of the effort of this research investigating autoencoding architectures that could
preserve the road texture. As mentioned before we studied different cost functions. Although results
were similar in terms of MSE, a learned cost function using generative adversarial networks had the
most visually appealing results. In Fig. 3 we show decoded images with models trained using two
different cost functions. As expected, MSE based neural networks generate blurred images. The
main issues for our purposes is that blurring connects the lane marks into a single long lane. Also
5

(b)

(a)

Figure 3: Samples using similar fully convolutional autoencoders. Odd columns show decoded
images, even columns show target images. Models were trained using (a) generative adversarial
networks cost function (b) mean square error. Both models have MSE in the order of 102 and
PSNR in the order of 10.

Figure 4: Samples generated by letting the transition model hallucinate and decoding z using Gen.
Note that Gen was not optimized to make these samples realistic, which support our assumption
that the transition model did not leave the code space.

blurring reconstruction does not preserve leading car edges. This is not desired for several reasons,
the main one is that it gets rid of one of the main clues used for visual odometry and leading car
distance estimation. On the other hand MSE learns to generate curved lane markings faster than
the adversarial based model. It is possible that learning to encode pixels with the car steering angle
information should avoid this problem. We leave this investigation for future work.
Once we obtained a good autoencoder, we trained the transition model. Results of predicted frames
are shown in Fig. 4 . We trained the transition model in 5Hz videos. The learned transition model
keeps the road structure even after 100 frames. When sampling from the transition model with
different seed frames we observed simulations of common driving events such as passing lanes,
approaching leading cars and leading cars moving away. The model failed to simulate curves. When
we initialized the transition model with frames of driving on a curve, the transition model quickly
straightens the lanes and goes back to simulate going forward. Nevertheless, it is promissing that we
could learn transitions for video generation without explicitly optimizing a cost function in the pixel
space. Also, we believe that even more realistic simulations will be possible with more powerful
transition models (deep RNNs, LSTM, GRU) and contextual encoding (sensor fusion of pixels plus
steering angle and speed). The dataset we are releasing with this paper has all the sensors necessary
to experiment with this approach.
6

Conclusions

This paper presented comma.ai initial research on learning a driving simulator. We investigated
video prediction models based on autoencoders and RNNs. Instead of learning everything end-toend here we first trained the autoencoder with generative adversarial network based cost functions
to generate realistic looking images of the road. We followed up training an RNN transition model
in the embedded space. Results from both autoencoder and transition model look realistic but more
research is needed to successfully simulate all the relevant events for driving. To stimulate further
research on this subject we are also releasing a driving dataset with video and several sensors such
as car speed, steering angle, etc. We also release the code to train the networks we investigated here.

References
[1] Diederik P Kingma and Max Welling, Auto-encoding variational bayes, arXiv preprint
arXiv:1312.6114, 2013.
[2] Emily L Denton, Soumith Chintala, Rob Fergus, et al., Deep generative image models using
laplacian pyramid of adversarial networks, in Advances in Neural Information Processing
Systems, 2015.
[3] Radford, Alec, Luke Metz, and Soumith Chintala. Unsupervised representation learning with
deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434, 2015.
[4] Diederik Kingma and Jimmy Ba, Adam: A method for stochastic optimization. arXiv
preprint arXiv:1412.6980, 2014.
[5] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil
Ozair, Aaron Courville, and Yoshua Bengio, Generative adversarial nets, in Advances in
Neural Information Processing Systems, 2014.
[6] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly, and Ian Goodfellow, Adversarial autoencoders, arXiv preprint arXiv:1511.05644, 2015.
[7] Jan Koutnk, Giuseppe Cuccu, Jurgen Schmidhuber, and Faustino Gomez, Evolving largescale neural networks for vision-based reinforcement learning, Proceedings of the 15th annual
conference on Genetic and evolutionary computation, 2013.
[8] Volodymyr Mnih, Koray Kavukcuoglu, David Silver, et al., Human-level control through
deep reinforcement learning, Nature, 2015.
[9] David Silver, Aja Huang, Chris Maddison, et al., Mastering the game of Go with deep neural
networks and tree search, Nature, 2016.
[10] Sergey Levine, Peter Pastor, Alex Krizhevsky, and Deirdre Quillen, Learning Hand-Eye
Coordination for Robotic Grasping with Deep Learning and Large-Scale Data Collection,
arXiv preprint arXiv:1603.02199, 2016.
[11] Brian L Stevens, Frank L Lewis and Eric N Johnson, Aircraft Control and Simulation: Dynamics, Controls Design, and Autonomous Systems, John Wiley & Sons, 2015.
[12] Eric R Westervelt, Jessy W Grizzle, Christine Chevallereau, et al., Feedback control of dynamic bipedal robot locomotion, CRC press, 2007.
[13] HJ Kim, Michael I Jordan, Shankar Sastry, Andrew Y Ng, Autonomous helicopter flight via
reinforcement learning, Advances in neural information processing systems, 2003.
[14] Junhyuk Oh, Xiaoxiao Guo, Honglak Lee, et al., Action-conditional video prediction using
deep networks in atari games, Advances in Neural Information Processing Systems, 2015.
[15] Manuel Watter, Jost Springenberg, Joschka Boedecker and Martin Riedmiller, Embed to
control: A locally linear latent dynamics model for control from raw images, Advances in
Neural Information Processing Systems, 2015.
[16] Jurgen Schmidhuber, On learning to think: Algorithmic information theory for novel combinations of reinforcement learning controllers and recurrent neural world models, arXiv
preprint arXiv:1511.09249, 2015.
[17] Michael Mathieu, Camille Couprie and Yann LeCun, Deep multi-scale video prediction beyond mean square error, arXiv preprint arXiv:1511.05440, 2015.
7

[18] Ramon van Handel, Probability in high dimension, DTIC Document, 2014.
[19] Eder Santana, Matthew Emigh and Jose C Principe, Information Theoretic-Learning AutoEncoder, arXiv preprint arXiv:1603.06653, 2016.
[20] Eder Santana, Matthew Emigh and Jose C Principe, Exploiting Spatio-Temporal Dynamics
for Deep Predictive Coding, Under Review, 2016.
[21] Alireza Makhzani, Jonathon Shlens, Navdeep Jaitly and Ian Goodfellow, Adversarial autoencoders, arXiv preprint arXiv:1511.05644, 2015.
[22] Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, et al., Generative adversarial nets, Advances in Neural Information Processing Systems, 2014.
[23] Jeff Donahue, Philipp Krahenbuhl and Trevor Darrell, Adversarial Feature Learning, arXiv
preprint arXiv:1605.09782, 2016.
[24] Alex Lamb, Vincent Dumoulin Vincent and Aaron Courville, Discriminative Regularization
for Generative Models, arXiv preprint arXiv:1602.03220, 2016.
[25] Anders Boesen Lindbo Larsen, Sren Kaae Snderby, Hugo Larochelle and Ole Winther, Autoencoding beyond pixels using a learned similarity metric, arXiv preprint arXiv:1512.09300,
2015.
[26] Jose C Principe, Neil R Euliano, W Cur Lefebvre, Neural and adaptive systems: fundamentals
through simulations with CD-ROM John Wiley & Sons, Inc., 1999.

You might also like