Neural-Agent Dogfighting Flight Simulator

• 

  Title Screen

• 

  Two Plans dogfighting from a fixed perspective

• 

  Two Plans dogfighting from their first person perspective

Inspiration

I recently watched OpenAI's paper on multi-agent hide and seek players. Most of the algorithms and techniques used are very well documented in the paper, and were able to be somewhat nicely ported over to Unity. Instead of using OpenAI's Gym directly, we opted for using Unitys integration with OpenAI, the ML-Agents tool.

What it does

There are two modes, 1v1, and BotVsBot. The 1v1 mode is simply a dogfighting flight simulator for a splitscreen of two people. The BotVsBot mod allows for the ml-brains to take control of the sticks and fly the aircraft. They're given rewards for flying straight, fighting, hitting, and killing. They are punished for getting killed, hitting the ground, and fleeing the arena.

How we built it

We build it using Unity and ML-Agents

Challenges we ran into

Aerodynamics for sure. It was difficult to make a somewhat accurate flight simulator in a decent amount of time.

Accomplishments that we're proud of

At the time of writing this, the agents are just now learning to fly the airplanes and not crash them into the ground

What we learned

ML-Agnet framework, flight simulation code

What's next for Neural-Agent Dogfighting Flight Simulator

Let it train for longer and watch!

Built With

• c#
• learning
• machine
• ml-agent
• networks
• neural
• python
• unity