carla-simulator / imitation-learning Goto Github PK

Hi, first thanks a lot for giving the dataset!

A question for dataset target description is:

Noise, Boolean ( If the noise, perturbation, is activated, (Not Used) )

I checked several h5 files, all the noise is 0,
however, there are noises in

Steer Noise, float
Gas Noise, float
Brake Noise, float

So, does that mean:

1. although there are noise, but noise are not input into system?(Noise Boolean are all 0 which means not used)
   or
2. All dataset are collected under noise?(I only checked for several h5 files, their's steer or gas or brake noise always exist)

In Conditonal Imitation Learning Paper, it writes

Our training dataset comprises 2 hours of human driving in Town 1 of which only 10%(roughly 12 minutes) contain demonstrations with injected noise. Collecting training data
with strong injected noise was quite exhausting for the
human driver. However, a relatively small amount of such
data proved very effective in stabilizing the learned policy.

But the dataset you provided is far more than 2 hours. So is this dataset same as the dataset paper used?

Thanks for sharing this, it seems to me this is a trained model for test purpose only using your benchmark system. According to your paper, you are utilizing a custom dataset (around 14 hours of manual and automatic driving in the Town01) with Adam optimizer for training the network. However, I need more detail about the perturbation technique? How can I add the same method for data collection in the CARLA using python interface? Because you're not going to release your dataset and the training code in the future soon, I'm going to build my own dataset, at least before you release the training code. Another question is, what about the second table results? Is the benchmarking system produce same results as the first table? (By tables I mean tables in "CARLA: An Open Urban Driving Simulator") And the last question is, how long the benchmark is? How many hours? I run that recently and it still runs after more than 2 and half hours with 1080 GPU and a good computer :D

In the different branches
for the steer, I think that the nonlinear function should be the tanh([-1,1]) which means the turning rate should be [-180 degrees~180 degrees]
for the brake and the gas, the activation function should be sigmoid ([0,1])
in the model, all the activation function is Relu
Is the output converted by the CARLA control?
I haven't run the code in the CARLA yet because now I'm using the Ubuntu 14.04

Hello,

I have a question about the loss used to train the model.
In your paper, the action is apparently two dimensional (steer and acceleration), whereas it seems to be 3-dimensional in the code (steer, acceleration and brake). Which one is the correct one ? And what is the loss for the action ? the squared error between groundtruth and computed action ? Or is there a weigh to give more importance to some aspect of the action (steer for instance) (the lambda in the 6th equation in the paper) ?

Moreover, is the loss for the actions summed with the loss of the speed branch during training ? And if yes, what is the weight of each loss (as the loss for the speed will be much bigger than the loss for the actions).

Thank you in advance

Hello,

I had a question regarding the present checkpoint in Conditional Imitation Learning. When I run run_CIL.py with the available checkpoint, the agent moves straight with no turns or dynamic objects around.

Is the checkpoint only for the task mentioned as 'Straight' ? If so, how are we to change the available training code for implementing the other tasks like 'One turn' or 'Navigation Dynamic' ?

It'd be of great help to receive some help.
Thanks!

when i run "python run_CIL.py",some problem happen

ERROR: 'CameraRGB'
Traceback (most recent call last):
  File "run_CIL.py", line 64, in <module>
    results = corl.benchmark_agent(agent, client)
  File "/home/dingjiangang/carla_env/local/lib/python2.7/site-packages/carla/benchmarks/benchmark.py", line 199, in benchmark_agent
    + '.' + str(end_point))
  File "/home/dingjiangang/carla_env/local/lib/python2.7/site-packages/carla/benchmarks/benchmark.py", line 108, in run_navigation_episode
    control = agent.run_step(measurements, sensor_data, target)
  File "/home/dingjiangang/imitation-learning/agents/imitation/imitation_learning.py", line 95, in run_step
    control = self._compute_action(sensor_data['CameraRGB'].data,
KeyError: 'CameraRGB'

hi,
I found that the value of the throttle in the dataset is basically 0.5 or 1, but the speed is basically uniform. Why? Assuming the throttle is always 0.5, the speed should always increase. I'm so confused.
Looking forward to your reply~

Hello.
I've read the paper and I've got a few questions about the Imitation Learning experiment.

I'm not a deep learning expert, so some questions may sound silly.

1. When executing the navigation tasks, after training, do the trained system receives new images from the camera installed on the car? Do you still have the images recorded during the execution of the expriment?

2. What camera mode was used by the forward-facing camera? Also, what there any kind of post-processing applied to the camera (i.e. ground truth, semantic segmentation, or just scene final)?

Hello,
The run_cil.py script cannot connect to the server. When trying to establish a connection, I get a connection refused or connection timed out errors. I know this is not a firewall or network reachability problem as I can run the manual_driving.py script and can control the vehicle remotely.

Find the output generated by running run_CIL.py below. I'd appreciate if any potential solution is shared.

 === Conv 0   :   5 2 32
Dropout 0
Tensor("Network/conv_block0/relu1:0", shape=(?, 42, 98, 32), dtype=float32)
 === Conv 1   :   3 1 32
Dropout 1
Tensor("Network/conv_block1/relu2:0", shape=(?, 40, 96, 32), dtype=float32)
 === Conv 2   :   3 2 64
Dropout 2
Tensor("Network/conv_block2/relu3:0", shape=(?, 19, 47, 64), dtype=float32)
 === Conv 3   :   3 1 64
Dropout 3
Tensor("Network/conv_block3/relu4:0", shape=(?, 17, 45, 64), dtype=float32)
 === Conv 4   :   3 2 128
Dropout 4
Tensor("Network/conv_block4/relu5:0", shape=(?, 8, 22, 128), dtype=float32)
 === Conv 5   :   3 1 128
Dropout 5
Tensor("Network/conv_block5/relu6:0", shape=(?, 6, 20, 128), dtype=float32)
 === Conv 6   :   3 1 256
Dropout 6
Tensor("Network/conv_block6/relu7:0", shape=(?, 4, 18, 256), dtype=float32)
 === Conv 7   :   3 1 256
Dropout 7
Tensor("Network/conv_block7/relu8:0", shape=(?, 2, 16, 256), dtype=float32)
Tensor("Network/reshape:0", shape=(?, 8192), dtype=float32)
 === FC 0   :   512
Dropout 8
Tensor("Network/fc1/relu9:0", shape=(?, 512), dtype=float32)
 === FC 1   :   512
Dropout 9
 === FC 2   :   128
Dropout 10
 === FC 3   :   128
Dropout 11
 === FC 4   :   512
Dropout 12
 === FC 5   :   256
Dropout 13
 === FC 6   :   256
Dropout 14
Tensor("Network/Branch_0/fc7/relu15:0", shape=(?, 256), dtype=float32)
 === FC 8   :   256
Dropout 15
 === FC 9   :   256
Dropout 16
Tensor("Network/Branch_1/fc10/relu17:0", shape=(?, 256), dtype=float32)
 === FC 11   :   256
Dropout 17
 === FC 12   :   256
Dropout 18
Tensor("Network/Branch_2/fc13/relu19:0", shape=(?, 256), dtype=float32)
 === FC 14   :   256
Dropout 19
 === FC 15   :   256
Dropout 20
Tensor("Network/Branch_3/fc16/relu21:0", shape=(?, 256), dtype=float32)
 === FC 17   :   256
Dropout 21
 === FC 18   :   256
Dropout 22
Tensor("Network/Branch_4/fc19/relu23:0", shape=(?, 256), dtype=float32)
Restoring from  D:\auto_dev\CARLA\CARLA_0.8.2\PythonClient\agents\imitation/model/model.ckpt-450000

D:\Program Files\Python\Python36\lib\site-packages\h5py\__init__.py:36: FutureWa
rning: Conversion of the second argument of issubdtype from `float` to `np.float
ing` is deprecated. In future, it will be treated as `np.float64 == np.dtype(flo
at).type`.
  from ._conv import register_converters as _register_converters
INFO: listening to server localhost:2000
INFO:tensorflow:Restoring parameters from D:\auto_dev\CARLA\CARLA_0.8.2\PythonCl
ient\agents\imitation/model/model.ckpt-450000
INFO: Restoring parameters from D:\auto_dev\CARLA\CARLA_0.8.2\PythonClient\agent
s\imitation/model/model.ckpt-450000
INFO: START
ERROR: (localhost:2000) failed to read data: timed out
ERROR: (localhost:2000) failed to read data: [WinError 10054] An existing connec
tion was forcibly closed by the remote host
INFO: START
ERROR: (localhost:2000) failed to read data: timed out
ERROR: (localhost:2000) failed to read data: [WinError 10054] An existing connec
tion was forcibly closed by the remote host
INFO: START
ERROR: (localhost:2000) failed to read data: timed out
ERROR: (localhost:2000) failed to read data: [WinError 10054] An existing connec
tion was forcibly closed by the remote host
INFO: START
ERROR: (localhost:2000) failed to read data: timed out
ERROR: (localhost:2000) failed to read data: [WinError 10054] An existing connec
tion was forcibly closed by the remote host
INFO: START
ERROR: (localhost:2000) failed to read data: timed out
ERROR: (localhost:2000) failed to read data: [WinError 10054] An existing connec
tion was forcibly closed by the remote host
INFO: START
ERROR: (localhost:2000) failed to read data: timed out
ERROR: (localhost:2000) failed to read data: [WinError 10054] An existing connec
tion was forcibly closed by the remote host

Hello @felipecode @juaxix @nsubiron @dosovits @marcgpuig ,

I'm currently working on training the model you provided in paper in order to reproduce the result. However, I found the loss is hard to decrease. It always around certain level, up and down again and again. I have tried different lr and optimizer but seems not work.

I check the label, AKA the first three columns in "target" dataset (steer, acc, brake) . I found the steer angle is too small most of time. Maybe that is the reason for hard training since the loss calculated every time is also small and the corresponding gradient is small too?

I want to know if you have done any pre-processing on the label? or is there any tricks for training?

Very thankful for your work and hope you can help me with these questions.

@felipecode
Hi there,
I have checked the .h5 dataset and found that the shape of "target" part is 200 * 28.
I know the meaning of the num of cols "28". But I don't know the meaning of the rows number "200". Why are there 200 points for every image? what does it represent？

Look forward to your reply~ Thanks a lot in advance!

The README mentions that in the provided dataset, the 'targets' key has 27 measurements.
But when I printed out the size of 'targerts' it showed that it has 28 measurements .

Here is the code:

datasetDirTrain = '/mnt/AshishMehta/Downloads/CARLAImitationLearningDataset/AgentHuman/SeqTrain/'
datasetDirVal = '/mnt/AshishMehta/Downloads/CARLAImitationLearningDataset/AgentHuman/SeqVal/'

import glob
import h5py
datasetFilesTrain = glob.glob(datasetDirTrain+'*.h5')
datasetFilesVal = glob.glob(datasetDirVal+'*.h5')

dataset = h5py.File(datasetFilesTrain[120], 'r')

print(list(dataset.keys()))
print(dataset['rgb'].shape)
print(dataset['targets'].shape)

ind = 120

print(dataset['targets'][ind])
dataset.close()

The output is as follows :

['depth', 'labels', 'rgb', 'targets']
(200, 88, 200, 3)
(200, 28)
[-2.0544434e-01 6.4567566e-01 0.0000000e+00 0.0000000e+00
0.0000000e+00 -1.9859870e-01 6.4567566e-01 0.0000000e+00
-3.1488324e+02 1.5389464e+04 2.5427959e+01 9.4275820e+06
2.1926109e+05 2.3462818e+06 0.0000000e+00 6.0470324e-02
1.1559283e+01 1.5627192e+00 -4.1669765e+00 8.0673536e+07
6.6544400e+05 -3.6079049e-02 9.9933618e-01 5.0120126e-03
2.0000000e+00 0.0000000e+00 0.0000000e+00 0.0000000e+00]

Hello,
Can the (annoyingly) requirements for CUDA 9,0 be reduced? I've only managed to get access to CUDA 6.5.14.

Kind regards,

Hi,
I have the following questions. Really appreciate if anyone can answer..!!

• Firstly I understand the control commands are like one hot encoded inputs for--> High level command, int ( 2 Follow lane, 3 Left, 4 Right, 5 Straight). How are these generated - carla outputs these automatically? or are they looking at the steer values to output these ?
• speed prediction auxiliary output - really cant get my head around why this is helping longitudinal performance?
• Training: do we run the backprop only on the active branch for each frame ?

Hi,

Is it possible to train end-to-end network on carla dataset. I mean, what is mentioned in paper is conditional imitation, but I don't want to use conditional command. I just want to use steer+gas+brake for my network to learn and predict on basis of image. I made one such network which tends to drive straight with some deflection towards right or left (random behaviour) and gets collided or overlaps lane. So I just wanted to ask if anyone has made or is it possible to have such trained model running on carla?

I would like to know what does the targets represent, please.
I found that they have the shape of ?,28.
So can I know what it means?

Hi. Has the training code been released yet?

Hey,

are there any intentions to release the source code of the training?

Cheers

Hi.
When both server and client are on the same computer, what address should I use to connect to the server (e.g. 'localhost', '127.0.0.1', or the computer's name)? I can confirm the server is set up and running correctly, as identified by TCPview, but the only message returned by the experiment client are 'connection timed out', or 'failed to read from server'). Client and server ports match.

Hello,
What's the total file size of the dataset? Due to the lack of storage on my current hard disk, I couldn't download the full dataset. I know that it is larger than 10GB, but how much exactly? I can then free the required space up.

i use your source training code to training from scratch, but i found when the loss decreased to 2, it start to fluctuate between 3 and 7.so it that reflects that the training is overfitting?
and are there any training tricks? Hoping for your answer.THX

I've been about this every way I can think of and am unable to find a solution.

I can run the manual and example clients on the carla-server with no problem.

When I try to run 'python run_CIL.py' I am returned with this:

:~/Desktop/Codevilla/imitation-learning$ python run_CIL.py
Traceback (most recent call last):
File "run_CIL.py", line 5, in
from carla.benchmarks.corl_2017 import CoRL2017
ModuleNotFoundError: No module named 'carla'

With sudo it kicks up this:

:~/Desktop/Codevilla/imitation-learning$ python run_CIL.py
Traceback (most recent call last):
File "run_CIL.py", line 5, in
from carla.benchmarks.corl_2017 import CoRL2017
ModuleNotFoundError: No module named 'carla'

Any ideas?

Thanks.

Hello,
A few questions about this experiment:

1. Is the image dataset used in any way?
2. Is the training code made available in this experiment?

Hi guys,

Currently, I am testing the released model. However, I found my result is not as good as the result of this paper.
For example, the result of 'new town' with 'straight' is 97%, but my test result is 71%. I checked each episode and found in many cases that the car stops at the start point and doesn't move at all.
I didn't change any of the released code and my Carla version is 0.7.1.

Any ideas why could this happen? BTW, my command to launch Carla is:
./CarlaUE4.sh /Game/Maps/Town01 -carla-server -benchmark -fps=10 -windowed -ResX=800 -ResY=600
Best,
Tairui

Hey,

just one question regarding the Loss function (equation 6 in your paper).
What values did you use for the Lambda?

Thanks in advance & cheers

Hello! In the paper (Section 4, Subsection D) Data augmentation), you mention Data Augmentation with randomly sampled magnitudes.
Unfortunately, random magnitudes of the transformations introduces quite a lot of hyperparameters to tune when trying to reproduce the paper.

Could you please supply the extends of the magnitudes of the augmentations you used in the paper as a baseline for our reproduction work?

I'd be specifically interested in the Minima/Maxima of the distributions concerning:

• Contrast
• Brightness
• Salt and Pepper magnitude
• Gaussian Noise Sigma

Als well as the Kernel Size of the gaussian Blur.

Thank you in advance.

Hi, felipecode,

Could you tell me whether speed in the provided dataset is measured in km/h, or m/s ?

THanks a lot!!

Hi! I have three questions regarding how to use this repository.

1. With the command I start the the agent, (python run_CIL.py), does it load a trained model and drive the car or does it start training?
2. Does the files in the Model folder represent the pre-trained model on the 24GB dataset or not? If so, how do I start the agent by using the trained model?
3. If it the model is not trained, how do we pass the HDF5(24GB dataset) into the imitation_learning/imitation_learning_network class and train it?

Thanks alot!

Just trying to run a model trained on this dataset with CARLA for benchmarking, and realized that the images in it seems to be mean subtracted, but I can't seem to find the actual mean anywhere in your code, am I missing it?

Any information you could let me know on the preprocessing of the images would be helpful. Sorry if I've missed it somewhere!

Hi,
Apologies if this is not the right channel...

When I run the CIL experiment, will I be able to see the car driving in the virtual streets automatically (based on the trained model)? I mean, can I see it as if I were playing the client in via pygame? I've tried hard to get it running, but all those graphical requirements kind of killed me.

A quick question about the dataset:

Can you please clarify the target key # 24:
High level command, int ( 2 Follow lane, 3 Left, 4 Right, 5 Straight)???

When I open a sample data, I see only a real value for this key. Should not be a vector of 5 real number based on its definition?

`Steer, float
Gas, float
Brake, float
Hand Brake, boolean
Reverse Gear, boolean
Steer Noise, float
Gas Noise, float
Brake Noise, float
Position X, float
Position Y, float
Speed, float
Collision Other, float
Collision Pedestrian, float
Collision Car, float
Opposite Lane Inter, float
Sidewalk Intersect, float
Acceleration X,float
Acceleration Y, float
Acceleration Z, float
Platform time, float
Game Time, float
Orientation X, float
Orientation Y, float
Orientation Z, float
High level command, int ( 2 Follow lane, 3 Left, 4 Right, 5 Straight)
Noise, Boolean ( If the noise, perturbation, is activated, (Not Used) )
Camera (Which camera was used)
Angle (The yaw angle for this camera)`

Hi, first thanks a lot for giving the dataset!
Please, how can i identify if a given image captured from center, left or right camera?

Basically run:

$ python run_CIL.py

Ihave a carla server running .

Cannot assign a device for operation 'Placeholder': Operation was explicitly assigned to /device:GPU:0 but available devices are [ /job:localhost/replica:0/task:0/device:CPU:0 ]. Make sure the device specification refers to a valid device.

python run_CIL.py --help
/home/danfouer/.conda/envs/tensorflow/lib/python2.7/site-packages/h5py/init.py:36: FutureWarning: Conversion of the second argument of issubdtype from float to np.floating is deprecated. In future, it will be treated as np.float64 == np.dtype(float).type.
from ._conv import register_converters as _register_converters
usage: run_CIL.py [-h] [-v] [--host H] [-p P] [-c C] [-n T] [--avoid-stopping]
[--continue-experiment]

optional arguments:
-h, --help show this help message and exit
-v, --verbose print debug information
--host H IP of the host server (default: localhost)
-p P, --port P TCP port to listen to (default: 2000)
-c C, --city-name C The town that is going to be used on benchmark(needs
to match active town in server, options: Town01 or
Town02)
-n T, --log_name T The name of the log file to be created by the scripts
--avoid-stopping Uses the speed prediction branch to avoid unwanted
agent stops
--continue-experiment
If you want to continue the experiment with the given
log name

python2.7 and python3.5 got the same result

Hello,
Would there be any easy way to get the experiment running on CUDA 6.5? I stumbled onto the fact it requires CUDA 9.0. This is unfair as graphics cards which such support are the most expensive ones.

I can confirm the following problems still exist in the current version of the code. Is there anything that I missed?

File "run_CIL.py", line 5, in <module>
    from carla.driving_benchmark.experiment_suite import CoRL2017
ImportError: No module named experiment_suite

Fixed by changing experiment_suite to experiment_suites.

File "run_CIL.py", line 7, in <module>
  from agents.imitation.imitation_learning import ImitationLearning
  File "./CARLA_0.8.2/PythonClient/agents/imitation/imitation_learning.py", line 12, in <module>
    from carla.benchmarks.agent import Agent
ImportError: No module named benchmarks.agent

Fixed by changing carla.benchmarks.agent to carla.agent

Traceback (most recent call last):
  File "run_CIL.py", line 65, in <module>
    agent = ImitationLearning(args.city_name, args.avoid_stopping)
  File "./CARLA_0.8.2/PythonClient/agents/imitation/imitation_learning.py", line 21, in __init__
    Agent.__init__(self, city_name)
TypeError: __init__() takes exactly 1 argument (2 given)

Fixed by changing Agent.init(self, city_name) to Agent.init(self)

P.S: I am not sure if these changes can raise any issue.

Hi. What directory should I unpack the dataset to?

Hello,

In the paper you report 88% completion rate. However when I run the trained model from this repo only one of the 25 straight runs pass (4%). When I double the timeout it goes up to 47%. Do you use a custom benchmark to achieve the 88% result reported?

The command used is:

python3 run_CIL.py --host <ip to remote machine>

I have a TitanX on the machine that I'm running this on.

Hello, @felipecode ,
I have some questions about the "branch" structure. It's amazing as well as a little confusing :)
I check the your code of imitation-learning network and found that you set five branches which include a speed branch for auxiliary task.
(1)Does the other four branches(except for speed branch) implicitly match the four high level command (follow lane, left, right, straight)?
(2)And the items in each branch --> (steer,gas,brake), Do they just match the output actions?
(3) I saw there are 28 cols in target data. So can I say that the first three cols --> (steer,gas,brake) are labels for training? These three cols should be considered as labels,not measurement inputs, am I right?
Actually, I noticed that you just set speed as the only measurement input while ignoring the other data in "target"(total 28 cols). Why don't use the other data?

Maybe I asked too much~
Hope you can help me with these questions. Thank you very much!

Hello,

I am facing a problem when I launch the imitation learning with the new version of Carla, it seems not to work.
The error I get is the following :

Traceback (most recent call last):
File "run_CIL.py", line 72, in <module>
args.host, args.port)
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/driving_benchmark/driving_benchmark.py", line 294, in run_driving_benchmark
benchmark_summary = benchmark.benchmark_agent(experiment_suite, agent, client)  
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/driving_benchmark/driving_benchmark.py", line 129, in benchmark_agent
+ '.' + str(end_index))
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/driving_benchmark/driving_benchmark.py", line 206, in _run_navigation_episode
measurements, sensor_data = client.read_data()
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/client.py", line 127, in read_data
return pb_message, dict(x for x in self._read_sensor_data())
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/client.py", line 127, in <genexpr>
return pb_message, dict(x for x in self._read_sensor_data())
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/client.py", line 175, in _read_sensor_data
yield self._parse_sensor_data(data)
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/client.py", line 180, in _parse_sensor_data
return parser.name, parser.parse_raw_data(data[4:])
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/client.py", line 196, in parse_image
return sensor.Image(frame_number, width, height, image_type, fov, data[24:])
File "/home/florence/tensorflow1.5-3.5/lib/python3.5/site-packages/carla/sensor.py", line 149, in __init__
assert len(raw_data) == 4 * width * height
AssertionError

This error is weird as the assertion is the same in the previous version of carla...

Thank you in advance

Florence

Thanks for sharing the great work.

I have a question abut the dataset. I understand that each .h5 consists of 200 images, but my question how these .h5 have been collected? Particularly, I mean these files starts from data_03663.h5 and ends by data_06951.h5. Dose all of these files have been collected sequentially? Does it mean these files are part of sequence or episodes?

Thanks for clarification.
Best.

First, I would like to know how did you generate the labels for control branches? and any plans to share source code that you used for training?

Thanks for this wonderful project!

I am curious whether there would be an update to run_CIL.py incoming so that one can reproduce the result in paper from scratch.

I can see the network files have been updated, and I guess we still need a hook to use these models.

Currently the conditional imitation learning agent
just works on GPU. It shouldn't be hard to also make
it it possible to run on CPU, since the fps requirements
aren't so big. Refers to #23

Hey I am trying to implement the paper, but my results are not good.

1. I noticed that the most of the dataset had a very low steering value even for left or right commands, so was there any data balancing technique used on the training dataset? if yes can you give information about it.

2. I read in other issues that I need to mask the branches so that you back-propagate only on one branch? I am not sure how to do that, can you give more information on how you did it or some code?

Thanks for your help.

Hello,
Quick question: on average, how long does it take to re-run this experiment in full? I can see that the dataset has 24GB instead of the 14GB stated on the main page.

Regards,
Eduardo

Hi, I want to ask how you crop the image to generate this SeqTrain dataset ?

As I am trying to train a model on your dataset and test it on carla, I want to keep consistent with the image pre-processing steps. Thanks.

First, thanks for releasing your dataset!

I'm trying to load it using h5py, and I've found that one file in the training set (data_06790.h5) consistently causes an error of the sort:

IOError: Unable to open file (Unable to find a valid file signature)

Removing this episode from the dataset, everything else loads. I've downloaded the data twice now, and both downloads have the same MD5 hash, so I don't think it's corruption from the download:

8b61e53ec55e1f79004d79fb61ab8d6c CORL2017ImitationLearningData.tar.gz

Hi,
I am running your pre-trained model which is uploaded in this repository. When I am running the dynamic-navigation environment benchmark, the car is not accounting for the traffic light and if a car stops due to traffic light, it just crashes with the car. Is there any thing we can do to attend to any car stopping in front of the agent or to attend to the traffic light? I think all this agent is doing is obeying the commands it is getting from the Planner. Any help with be greatly appreciated.

Hey!

When I run the 'run_CIL.py' of Carla_0.8.2 in Town01, it works good. However when change to Town02, something strange happen:

1. The first episode runs smoothly, but the car turns two times in this episode. The start and end pose is (38, 34). According to CoRL2017 these pose should be a straight line. So what's wrong with the setting?
2. When first episode finish and comes to the second one (start and end pose is (4, 2)), the run_CIL will put the following error and stop running:

Traceback (most recent call last):
File "run_CIL.py", line 72, in
args.host, args.port)
File "/CARLA_0.8.2/PythonClient/carla/driving_benchmark/driving_benchmark.py", line 294, in run_driving_benchmark
benchmark_summary = benchmark.benchmark_agent(experiment_suite, agent, client)
File "/CARLA_0.8.2/PythonClient/carla/driving_benchmark/driving_benchmark.py", line 121, in benchmark_agent
self._get_shortest_path(positions[start_index], positions[end_index]))
File "/CARLA_0.8.2/PythonClient/carla/driving_benchmark/driving_benchmark.py", line 182, in _get_shortest_path
end_point.orientation.x, end_point.orientation.y, end_point.orientation.z])
File "/CARLA_0.8.2/PythonClient/carla/planner/planner.py", line 114, in get_shortest_path_distance
track_target, target_ori)
File "/CARLA_0.8.2/PythonClient/carla/planner/city_track.py", line 91, in compute_route
route = a_star.solve()
File "/CARLA_0.8.2/PythonClient/carla/planner/astar.py", line 142, in solve
return self.get_path()
File "/CARLA_0.8.2/PythonClient/carla/planner/astar.py", line 111, in get_path
path.append((cell.x, cell.y))
AttributeError: 'NoneType' object has no attribute 'x'

So is there are somthing wrong in the code?

Thank you very much!
Dai Qi

Hi, thanks for your great work!
Currently, I am going to collect more data for training. However, I have no idea how to achieve some of the measurement in your 'target list', such as 'control signal'. I checked this page. I didn't find anything like 'control signal' or 'noise'. How can I get those values?