Reinforcement Learning for Drone Racing
My colleague and I developed two different approaches for a drone racing challenge, where the drone needs to go through the gates and avoid obstacles. He focused on automatic and smart ways to generate a drone trajectory, where I tried to improve the performance under unknown dynamics or noise using a shared control approach. I trained an RL agent using the progress reward for path planning and used the convex combination of the precalculated trajectory and RL agent.
This project explores robust drone racing under uncertain dynamics by combining trajectory priors with reinforcement learning. The figures below summarize the approach and results.
The drone trajectory from different perspectives under modelling mismatch and measurement noise.