experiment_details.md

Experiment Details

Notations

Agent / Model

It is the model that controls the vehicle. It takes in relevant input (eg. camera images) and outputs the vehicle control commands (steering, throttle, brake). The Imitation Learning model is an example an agent.

Scenario / Episode

It is where an agent drives a vehicle from one point to another, or until a timeout flag is triggered.

Attack / Adversary

In the context of this project, it is drawing black lines on the road within Carla. Eg. `../adversary/adversary_Town01_nemesisA.png`

Baseline

It is a scenario without any attack.

Hijack

It is defined as the ability to modify the route taken by a vehicle by using physical adversaries.

Scenario setup

Different tasks and defined in carla/driving_benchmark/adversary_suite.py as follows:

def _poses_town01_nemesis(self):
    """
    Each matrix is a new task. We have all the four tasks
    """

    def _poses_straight():
        return [[31, 36], [144, 146]]

    def _poses_turn_right():
        return [[42, 47], [79, 103]]

    def _poses_turn_left():
        return [[70, 66], [85, 98]]

    def _poses_straight_intersection():
        return [[100, 120], [119, 99]]

    def _poses_right_intersection():
        return [[107, 99], [119, 96]]

    def _poses_left_intersection():
        return [[100, 96], [107, 120]]

    return [_poses_straight(),
            _poses_turn_right(),
            _poses_turn_left(),
            _poses_straight_intersection(),
            _poses_right_intersection(),
            _poses_left_intersection()]

The numbers for each pose correspond to various starting and ending positions on the map. For example, if we look at three poses at the intersection, we have 6 paths that correspond to the following:

Hijack Poses