
PPO quad racing in our sim: 5.7 s laps, shaky take-off
Trained a quad to race a 90 m figure-eight gate track with PPO in our sim (Unreal Engine 5 + JSBSim). 96 drones in parallel, 100M steps on a laptop. The policy commands body rates, sees simulated sensors through an estimator and never gets ground truth.
Median lap is 5.7 s. A time-optimal planner (TOGT) does 4.72 s on the same quad, so it's about 20% off.
The weak spot is the take-off: only about half of the ground starts finish a lap. Anyone solved that with a curriculum, or by giving ground starts more weight?
Built on: Swift (Nature 2023) for the reward and mid-air starts
https://www.nature.com/articles/s41586-023-06419-4
TOGT https://github.com/FSC-Lab/TOGT-Planner