Reproduce “Learning Perception-Aware Agile Flight in Cluttered Environments” on OpenResearch

About this paper

Yunlong Song, Kexin Shi, Robert Penicka, Davide Scaramuzza

Researchers at the University of Zurich and ETH Zurich developed a learning-based framework that enables quadrotors to execute perception-aware agile flight in cluttered environments, achieving minimum-time navigation with a vision-based policy that integrates camera orientation into the reward function. The system demonstrated 100% success rates in challenging simulations and a low computational latency of 1.41 ms, validated through hardware-in-the-loop flights reaching speeds of 54.36 km/h.

Key insights

• Integrating explicit perception awareness into reinforcement learning reward functions significantly improves the robustness and success rates of agile flight policies in complex, cluttered environments.
• Distilling knowledge from a privileged state-based teacher to a vision-based student policy is an effective method for achieving high-performance, low-latency, end-to-end control for agile quadrotor navigation.
• Deep learning policies can achieve substantially lower computational latency (approximately 1.41 ms) compared to traditional modular pipelines and other learning-based methods, enabling higher control frequencies crucial for agile, high-speed flight.

Summary by alphaXiv.