We recommend the use of Anaconda to create a virtual environment for the project.

Run the following commands to install the Dukietown simulator:

# create a clean environment with python 3.8
conda create --name duckietown python=3.8 
conda activate duckietown

# git clone our repo
git clone https://github.com/AdaCompNUS/CS4278-5478-Project-Materials.git
cd CS4278-5478-Project-Materials

# install
pip install -e gym-duckietown

To test the installation, you may manually control the robot:

cd gym-duckietown
python manual_control.py --map-name map1

We will evaluate your system on 5 maps together with start and goal locations to compute the average accumulated reward. A primary component of your grade is how fast the robot finds the goal, and the reward achieved. In particular, take note below:

• Oscillating in place may generate a high score in the simulator, but will be penalized in our evaluation. This is a flaw in the Duckietown simulator.

To initialize the gym-duckietown environment, there are 7 arguments:

• --map-name: the name of the map
• --seed: random seed of the environment
• --start-tile: the starting tile. E.g. 1,13
• --goal-tile: the goal tile. E.g. 3,3
• --max-steps: the maximum run step. The default value is 1500. Do not change this default value when you generate the control files for submission.
• --control_path: the control file to execute.
• --manual: whether to use manual operation. When set to true, the robot can be controlled with arrow keys. Otherwise, the specified control file will be executed.

Generate the control files for submission, similar to what you have done in Lab 1. A sample file for the python environment is available here together with sample control file. They illustrate how to add arguments and output your controls to a file. To try our simple policy,

python run.py --map-name map4_0 --seed 2 --start-tile 1,13 --goal-tile 3,3 --manual False

You should be able to see the robot moving.

The robot is only allowed to take in the following information as input:

• the starting tile
• the goal tile
• a map image
• the goal images
• image observation after each step
• coarse robot position after each step

Test cases for the two submissions are in the testcases folder. For an example control file, see map4_0_seed2_start_1,13_goal_3,3.txt.