Projects Neural Network Control
How to Train Your Quadrotor
Reinforcement-learning flight controllers that are smooth, responsive, and survive the move from simulation to hardware.
RE+AL is a framework for training reinforcement-learning attitude controllers for quadrotors. It extends the earlier Neuroflight work with more realistic training inputs and a multiplicative reward design, producing smoother motor commands that transfer more reliably from simulation to real hardware. In flight tests, the trained agents tracked better than a tuned PID controller while drawing less current.