When an engine fails at 35,000 ft, our RL-based co-pilot generates candidate recovery trajectories, scores each on safety, ethics, and survivability, then presents the pilot with a ranked, fully explainable decision. Not a black box. Every recommendation comes with a reason.
PPO agents trained on custom flight-dynamics environments with reward decomposition for explainability
6-DOF aerodynamic model with configurable mechanical failure injection and terrain awareness
Avionics-grade visualization prototype with Leaflet maps, Canvas PFD, and real-time telemetry rendering
Reward decomposition and saliency mapping to translate RL decisions into human-readable rationale
PyTorch, Stable-Baselines3, Gymnasium for environment design, training, and inference
Designed with EASA AI Roadmap 2.0 and FAA ethical AI guidelines as a certification target
Making reinforcement learning transparent enough for safety-critical aviation decisions.
Replay Simulation