liampaull/dt-exercises

mooc-exercises

NOTE: clone using git clone --recursive ..., otherwise the submodules will not be imported

Local development workflow with the simulator

Quick start

in the root of this repo, run make start. The first time you run it, it will download the images, it might take some time. Once it starts, you should see a noVNC link that looks like this:

novnc_1               | noVNC HTML client started:
novnc_1               | 	=> connect via http://172.20.0.2:6901/?password=...

Click on the link (the one with http) and you should access the graphical user interface of the container. The password is quackquack. This is useful to run ros tools. Click on Application in the top left, and click terminal emulator. You can then try rqt_image_view and select the available topic in the dropdown list. to view what your robot sees. Other useful tools are rostopic list, rosnode info, rviz and others.

You should also be able to see in the terminal where you ran make that a Jupyter notebook server was started:

lanefollow_1          |     To access the notebook, open this file in a browser:
lanefollow_1          |      http://127.0.0.1:8888/?token=594c8735d4e95585d7f3e91d57fe24c99afefde142bea2b6

Use this link to access the notebook server (make sure you use the URL that starts with http://127.0.0.1). In this notebook, you can click on new and select terminal. You will then have access to a shell inside the container (alternatively, you can connect to the container by using docker exec -it $CONTAINERID /bin/bash from a terminal). This shell also have access to ros, so it can be used to start your agent (roslaunch my_package my_launch.launch). We have access to everything in the catkin/ folder, but the downside is that we do not have a GUI here.

You can now use your favorite editor on your computer and the changes will be reflected in this container. Try to make your robot complete a lap !

How does it work ?

when you run make start, the following container are started by docker-compose:

• a container for roscore. We could remove this at some point but I found it easier to use this way for development
• a container running lanefollow, which should be the agent from local development workflow. It is basically what we had a udem last year: https://github.com/duckietown-udem/udem-fall19-public/blob/master/DockerfileNB
• novnc container
• simulator: the goal was to use the same container that is used in the challenge server, so it is this image: https://github.com/duckietown/challenge-aido_LF-simulator-gym/tree/mooc-exercises
• middleware: In order to be able to use the challenge server's simulator image as-is, I re-used the same code to interface it. The middleware is used to handle the fifos connections and is really close to this: https://github.com/duckietown/challenge-aido_LF/tree/daffy-aido4/experiment_manager, though it was modified to remove the need for the scenario maker and use parameters from the param file.

Challenge submission

Since the older version of 1_develop and 3_submit have been merged, it should be easy to submit our agent to the AIDO challenge server. In order to submit your code, you can either modify the lf_slim launch file to be your launch file, or provide another launch file. To do so, modify the Dockerfile in the root by appending your launchfile name to the command:

CMD ["python", "solution.py"] would become

CMD ["python", "solution.py", "my_lf.launch"] (This will most likely be made easier through the param file, users should not modify the dockerfile)

Description of files

• catkin_ws/
  • Contains the user's packages. Has dt-core, dt-car-interface and dt-ros-commons as submodules. (This is why the repo should be clone using git clone --recursive git@github.com:duckietown/mooc-exercises.git). This folder is mounted into the notebook/agent docker image in the local development workflow.
• exercises/
  • (Experimental, not implemented) This was intended to host the different exercises
• docker-compose
  • This is the file that is used to start all the containers required for the local development workflow
• Dockerfile
  • Needed for dts challenge submit and dts challenge evaluate. It basically packages the submission into a docker image
• Makefile
  • Makes it easier to clean and start the workflow properly.
• submission.yaml
  • required for challenge submission
• lf_slim.launs
  • default launch file for challenge submission
• setup/
  • contains all the required configuration files for local development workflow

Configuration files:

• fifos-connector/

  • This is the middleware image. Currently used locally, but is should be removed and an image from dockerhub will be used instead.

• notebooks/

  • Those are the exercises from last year, imported for convenience

• utils/

  • common code for the notebook exercises from last year

• DockerfileNB

  • Dockerfile for the Notebook/agent image.

• launch_car_interface.sh

  • convenience script to launch the car interface automatically (could be moved inside start.sh)

• start.sh

  • Entrypoint of the notebook/agent container. This will automate all the stuff required:
    • run catkin build if required (since catkin is mounted it is not always required)
    • start the car interface
    • start the notebook
    • start the ros bridge (solution.py)

• requirement*

  • pip requirement files

• rosagent.py

  • ros node to interface the agent. This node does not spin() or loop, the event handling is done in solution.py.

• solution.py

  • This file is used both by the notebook/agent image and when we submit our agent to the AIDO server. It import rosagent and forward all the comunication to the FIFOS. Everything that communicates with the simulator goes through this (Send commands, receive observation, etc). The line wrap_direct(agent) is blocking and will act as a event handler instead of rospy.spin().

• parameters.env

  • Contains all the parameters for the containers. Can be used to tune the simulator.

TODOS

• One recurrent issue is the framerate. The number of renders was decreased to improve speed, but having more than 15 fps would be great.
• The notebook container has not been updated since last year, everything should be moved to python3 once ente is merge into daffy
• At some point we might want to move away from docker-compose and makefile. Everything will be migrated to python and dts mooc ...
• The workflow to make a submission should be made as easy as possible, i.e. we should not have to edit the dockerfile. It might be possible to use parameters.env
• Some of the knobs for the simulator are either not implemented, not working properly or not tested. We should have every param that we need working in the param file
• All the images that we need should be moved to the duckietown dockerhub and added to CI. They should be made as dt-projects
• It might be a better experience to have noVNC and notebook/agent container in the same container. We should also make use of the dashboard for everything that we can
• We should reuse as many docker layers as possible. base images have been changed recently and making sure we reuse the new base images when possible would save time and space.
• Move as much stuff as we can into setup and out of root of the repo to make it easier to see what is actually useful to student.