Jetson Nano can play Soccer

Hello, members of community.

I will introduce my project using a Jetson Nano board.

I uploaded code and explanation about this in my Github.

This is the project I started to test the Deep Learning method what I’m currently studying in an actual environment. The overall plan of the project is as follows.

  1. Make a Gazebo simulation similar to the soccer field situation by installing the goalpost, changing the floor to lawn.
  2. In this stadium, placing two Jetbot. One is for kicking the footboall and another is for defending it.
  3. Training a Attacker Jetbot and Defender Jetbot by the Deep Reinforcement Learning method in the situation of Multi-Agent such as Self-Play of AlphaGo Zero.
  4. When the training is completed, confirmed the performance using the actual soccer field and Jetbot equipped with the trained model of Gazebo.

No matter how well the simulation is made, there are many cases where the operation of is not good in the actual situation becasue of various physical factor. Thus, I plan to use many Deep Reinforcement Learning methods that have been studied to solve this such as https://github.com/iclavera/learning_to_adapt .

If you have any question about my project, please leave a comment.

Thank you for reading!!

1 Like

Hi Kimbring2,

I am trying to replicate your setup. However I am struggling with the python code for Gazebo Simulator in jetbot_soccer/jetbot/jetbot_control/src/main.py. To me it is not clear if it python2.7 or python3 based. Python3 is not really supported by ROS and for python2.7 there are no current tensorflow installation instructions for the Nano, just TX1 or TX2.

Please can you give me some details on your installation.

Thanks,
Thomas

Hi, Tsteinle

Thank you for your interest in my project.

At first, I use a below python version when I run a code.

Python 2.7.15+ (default, Oct  7 2019, 17:39:04) 
[GCC 7.4.0] on linux2
Type "help", "copyright", "credits" or "license" for more information.

I also check Python 3 was not supported by ROS. Thus, you should use a Python 2.

And concept of my project is transferring trained model from Gazebo simulator to Jetson Nano.
Thus, you do not need a install a Tensorflow directly in Jetson Nano.

Maybe, you can freeze a Tensorflow pretrained model and use it for inferencing at Jetson Nano using TensorRT.
Detailed link : https://github.com/kimbring2/jetbot_soccer/blob/master/frozen_graph.ipynb

Sorry for inconveniences because the project is not yet fully completed. However, I am trying to improve project.

If you have more question, feel free to ask.

Thank you very much.
From Kimbring2

Hi KImbring2,

Ok, got it.
But how did you install tensorflow for python2.7 on your Nano? I could not find any instructions to do so, only for python3.
And which base uSD-Card image did you use?

Thanks,
Thomas

To Tseinle,

I sent a private message. Please check it.

Thank you,
From kimbring2