Drone Feedback Controller
Adaptive Q-learning reinforcement learning agent with a nested PID actuation layer for UAV position stabilisation and waypoint navigation.
Problem to result
Who had the problem, what I built, the hardest part, and what it achieved.
The coursework drone had to reach each waypoint within 10 seconds and stop on it; my first controller overshot by 6.8%, against a target of under 1%.
A drone controller that learns which of 16 tuning profiles to use as it closes in on each waypoint.
Removing the overshoot without breaking the time limit. The integral term wound up during the climb and pushed the drone past the target, and the learning agent kept picking fast profiles near it. I zeroed the integral term on approach, stopped exploration within 0.5 m, made overshoot costly in the reward, and blocked the slow profiles far from the target so it still arrives in time.
Overshoot cut from 6.8% to 0.20% / 0.29% / 0.37% (X / Y / Z); arrival in 8.6 s against a 10 s limit; control loop at 1000 Hz.
Adaptive RL Control Simulation
Real-time Q-learning simulation with PID gain scheduling.
Project Walkthroughs
Click any video to play.
3:15Drone Week-1 trail
Initial drone feedback control session and flight testing.
LiveDrone Week-2 trail
UAV adaptive trajectory tracking and simulation demonstration.
LiveDrone Week-3 trail
Trajectory tracking and stabilisation testing during week 3.
LiveTechnical Evaluvation
UAV adaptive position stabilisation and robust waypoint tracking tests.
Problems & Fixes
What went wrong while building it, and how each problem was solved.
6.8% XY and 3.5% Z overshoot
Cause: Too little damping and too much speed on the approach
Raised kd_xy by more than 250% and lowered kp. Reduced the speed caps on the mid-range profiles. Added an overshoot penalty to the RL reward.
GitHub
GitHub repositories for this project.
Drone Feedback Controller Repository
Access the complete source code on GitHub.