Adversarial Driving

This package provides MDP models for safety validation of autonomous vehicles. It is built on top of AutomotiveSimulator.jl.

The actions of the AdversarialDrivingMDP represent disturbances to adversarial agents on the road. The reward is designed to encourage critical scenarios for the ego vehicle.

For a quick start, look at the two scenarios in examples/. The output of those examples are shown in the following gifs.

Nominal Behavior	Failure Example

Installation

Install with import Pkg; Pkg.add(url="https://github.com/sisl/AdversarialDriving.jl")

Usage

The AdversarialDrivingMDP is very versatile in its construction and supports the following arguments

• sut::Agent - The system under test (SUT) or ego-vehicle. This is the entity that the safety validation is being performed upon.
• adversaries::Array{Agent} - The list of adversaries on the road with the SUT.
• road::Roadway - The roadway defined in AutonomotiveSimulator.jl.
• dt::Float64 - Simulation timestep.
• other_agents::Array{Agent} - Other agents on the road that are not being controlled as adversaries and are not the SUT. Default: Agent[].
• γ - Discount factor on future rewards. Default: 1.
• ast_reward - Whether or not to use the AST reward, defined here. Default: false.
• no_collision_penalty - Penalty (in AST reward) for not finding a collision. Default: 1e3.
• scale_reward - Whether or not to scale the reward so it is in the range [-1,1]. Default: true.
• end_of_road - Define an early end of the road. Default: Inf.

Maintained by Anthony Corso (acorso@stanford.edu)