Reproduce “Agile Fall Recovery for Quadrotors with Bidirectional Thrust via Reinforcement Learning” on OpenResearch

About this paper

Anke Zhao, Yuhang Zhong, Kenghou Hoi, Junyu Mou, Junjie Wang, Lijie Wang, Jialiang Hou, Fei Gao

An RL-based framework enables bidirectional-thrust quadrotors to autonomously recover from arbitrary ground-resting attitudes to a stable hover, relying on minimal, noisy onboard sensors and achieving zero-shot sim-to-real transfer with 10/10 successful recoveries in normal indoor tests. The method generalizes to outdoor and disturbed conditions, outperforming a proportional-controller baseline.

Key insights

• An asymmetric actor-critic design with a recurrent actor is critical for learning complex, contact-rich maneuvers from noisy, partially observable sensor inputs.
• The use of a robust low-level INDI controller significantly improves disturbance rejection and reduces the sim-to-real gap, simplifying the high-level policy's learning task.
• High-fidelity sensor and motor modeling, combined with extensive domain randomization in simulation, enables successful zero-shot transfer of a fall recovery policy to real quadrotors using minimal sensors.

Summary by alphaXiv.