ACASXu Closed Loop Simulation Falsification Benchmark

Closed-loop simulation code using ACAS Xu neural networks for collision avoidance.

This repo contains different versions of in-plane flight with ACAS Xu for collision avoidance. Generally, the networks are activated by the ownship every 2 seconds to choose a command. The intruder also can adjust its command during flight.

Quick Start

These instructions are for Ubuntu 20.04, although they may work on other systems. Before the code can run, you need to install the python package requirements.

If you don't have pip3, first install that:

sudo apt update
sudo apt upgrade
sudo apt install python3-pip
pip3 install --upgrade pip

Next, install the required packages:

pip3 install -r requirements.txt

Next, you should be able to run the code. Change the directory to acasxu_dubins and run python3 acasxu_dubins.py or python3 parallel_acasxu_dubins.py (this version should use all cores on your machine so it's faster for long runs).

The expected output is as follows, along with a visualization of the closest-encounter out of all 10000 simulations:

0.........1.........2.........3.........4.........5.........6.........7.........8.........9.........
Did 10000 sims in 3.5 secs (0.354ms per sim)

Seed 671 has min_dist 4254.5ft
plotting state State(v_own: 148, v_int: 579, rho: 50848.08155907395)

Features:

• Nice visualization capability (and mp4 export).
• Initial state, chosen randomly, can also be rejected if ownship neural network command is not clear-of-conflict.
• Simulations stop once the distance between airfraft is increasing. This means that simulations may take different amounts of time to run.
• Includes random initial state as well as random intruder commands. If only straight-line intruder commands are desired, intruder_cmd_list can be set to all 0's (clear-of-conflict, fly straight).
• Goal: minimize distance between the two aircraft. Under 500 ft would be a violation. If intruder is faster the ownship, the property should always be possible to violate (assuming intruder can maneuver). Interuder velocity can therefore be used to tune the difficulty of the benchmark.

Usage:

Both acasxu_dubins.py and parallel_acasxu_dubins.py have the same flags as below:

usage: acasxu_dubins.py [-h] [--save-mp4]

Run ACASXU Dublins model simulator.

optional arguments:
    -h, --help  show this help message and exit
    --save-mp4  Save plotted mp4 files to disk.
    --intruder-turn  Toggles boolean flag to allow intruder to perform commands other than flying straight.
    --fixed-seed FIXED_SEED
                          Simulates the parameters generated by provided seed

TOOD list:

• add different plant models
• randomize ownship and intruder velocities within valid neural network range (parameter uncertainty)
• add command-line arguments for different versions of the benchmark
• add replay capability and save mp4 from command line
• add command-line usage to the readme
• find head-on example with same velocity and mirroring commands for paper.
• add different "desired" ownship commands, rather than straight
• add intruder (and maybe ownship) commands to be anything between -3 and 3 degrees per second, rather than just in [-3, -1.5, 0, 1.5, 3.0]
• add input quantization version (see https://arxiv.org/abs/2108.07961 for example values)