Building a custom octocopter from scratch with no prior hardware experience

A researcher successfully built a custom octocopter and trained a reinforcement learning policy to maintain flight even when motors fail, achieving zero-shot deployment on real hardware without fine-tuning. The neural network controller, running directly on the drone's flight computer, was trained only on single and dual motor failures but unexpectedly generalized to survive triple and quad failures in real-world tests. Interestingly, the policy learned to operate the octocopter using only four motors in many cases, since the reward function didn't incentivize using all eight motors when fewer could maintain the required pose.