Flexible Drones Trained in Simulation

⊲ arXiv ⊳ Collision-resilient drones can survive hitting a wall yet still fail to regain controlled flight, as the stock PX4 autopilot rejects valid velocity data after high-speed impacts. Researchers at ETH Zurich and the University of Bonn have developed REFLEX, a small neural network running entirely on the flight controller that stabilises the drone using only onboard motion sensing. Trained in NVIDIA Isaac Sim on the aftermath of 47 real collisions, it recovered from 19 of 22 impacts at 5 m/s, against one of 14 for the stock autopilot.
◂ Training on 47 real collisions inside NVIDIA Isaac Sim shows how small real datasets can anchor simulation-based learning for messy physical events. ▸
⊲ Featured Image – ETH Zurich ⊳
⊲ www.simulation.news ⊳