Cessna 172 RL Flight Control

Reinforcement Learning for Longitudinal Flight Control of a Cessna 172 Skyhawk

A curriculum-trained PPO agent evaluated against a classical PID controller across turbulence, wind shear, step-climb, and degraded control authority scenarios using JSBSim.

Overview

This repository contains the MATLAB source code, trained agent, and evaluation results for an undergraduate dissertation investigating whether a Proximal Policy Optimisation (PPO) agent can match or exceed the performance of a hand-tuned PID controller for altitude and speed tracking on a Cessna 172 Skyhawk in the JSBSim flight dynamics simulator.

Key Findings

Scenario	PID	PPO Agent
Moderate turbulence (Sev 1–3)	Comparable	Comparable
Severe turbulence (Sev 5)	Higher RMS	Lower RMS, smoother
Step climb (CG-varied MC)	Precise overshoot, fast settle	Sluggish, steady-state error
Wind shear (adverse MC)	Maintained stall margin	Dropped below stall in some runs
Elevator icing (40% authority)	Divergent — permanent altitude loss	Stable — maintained target

Prerequisites

• MATLAB R2024a or later
• Reinforcement Learning Toolbox
• Python 3.12 (for JSBSim bridge)
• JSBSim — see docs/setup.md for installation

Repository Structure

cessna172-rl-control/
├── environment/            # RL environment class (JSBSim wrapper)
├── training/               # Curriculum learning scripts (Phases 1–4 + finetune)
├── models/                 # Trained PPO agent (.mat)
├── evaluation/             # Single-run evaluation scenarios
├── monte_carlo/            # Monte Carlo and parametric sweep scripts
├── visualisation/          # FlightGear 3D demo
├── results/                # Figures (300 DPI) and raw data (.mat)
└── docs/                   # Setup and installation guide

Evaluation Scripts → Dissertation Sections

Script	Description
evaluation/step_climb.m	500ft step-climb transient response
evaluation/turbulence_sweep.m	Severity 1–5 sweep with 10-run MC
evaluation/severe_turbulence.m	Severity 5 detailed time-series
evaluation/wind_shear.m	Microburst wind shear penetration
evaluation/elevator_icing.m	Elevator icing + downdraft (key finding)
monte_carlo/step_climb_mc.m	Parametric MC — CG/payload variation
monte_carlo/severe_turbulence_mc.m	Stochastic MC — 10 seed realisations
monte_carlo/wind_shear_mc.m	Adverse MC — randomised onset/severity
monte_carlo/icing_authority_sweep.m	Elevator Authority sweep — 80% down to 40%

Quick Start

% 1. Run the most important scenario (elevator icing)
run('evaluation/elevator_icing.m')

% 2. Run the parametric authority sweep
run('monte_carlo/icing_authority_sweep.m')

% 3. Run the turbulence severity sweep
run('evaluation/turbulence_sweep.m')

Training

The agent was trained using a 4-phase curriculum:

Phase	Ailerons	Throttle	Turbulence	Episodes
1	0%	Fixed 80%	Severity 1	1200
2	50%	Fixed 80%	Severity 1–2	3500
3	100%	Fixed 80%	Severity 1–3	5000
4	100%	100% (learned)	Severity 1–4	5460

Followed by conservative fine-tuning with a reduced learning rate.

Licence

This project was completed as an undergraduate dissertation at Loughborough University. All rights reserved.