RTDynamics AI vs Rule-Based Dogfight in MAK VR-Forces. Video now available.

Dogfight between a traditional rule-based AI pilot and a neural network-controlled fighter trained with reinforcement learning (PPO algorithm).

The neural network was trained using a custom Gymnasium-compatible environment built with RTDynamics products:

• FixedWingLib CGF : Physics based flight dynamics model + virtual pilot
• CML : Air combat maneuvers
• EWAWS : Physics based sensors, weapons and countermeasures

The neural network acts as the tactical decision-maker. It observes the enemy aircraft and any incoming missiles, then selects the appropriate maneuver (pure pursuit, lift vector turn etc.) along with maneuver parameters like load factor and speed. It also decides when to launch missiles for maximum effectiveness

The trained neural network is exported in ONNX format. A dogfight controller, which supplies inputs to the network and interprets its outputs, is integrated into VR-Forces via a plugin. This plugin loads the exported ONNX file at runtime.

---

Interested in AI topics? Join the RTDynamics AI Newsletter here