Fangyu Sun

Zero-shot sim-to-real transfer for quadrotor control is now systematically achievable, thanks to a unified framework that integrates differentiable physics learning with training, validation, and real-world deployment.

Ditch the point-mass abstraction: this RL policy directly controls quadrotor body rates from depth images, achieving unprecedented speed and stability in complex environments.

Vector fields can guide differentiable policy learning to achieve agile drone racing, enabling faster convergence and better sim-to-real transfer.