Reproduce “Control of a Quadrotor with Reinforcement Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Jemin Hwangbo, Inkyu Sa, Roland Siegwart, Marco Hutter

A neural network trained entirely in simulation using reinforcement learning achieved robust low-level control of a quadrotor, demonstrating dynamic stabilization from highly unstable initial conditions like upside-down throws. The learned policy evaluates efficiently, running two orders of magnitude faster than a linear Model Predictive Control (MPC) system.

Key insights

  • Reinforcement learning can effectively produce direct state-to-actuator policies for low-level control of complex robotic systems like quadrotors, moving beyond higher-level decision tasks.
  • A novel conservative, deterministic on-policy learning algorithm provides stable convergence for continuous control tasks without requiring a stochastic policy for exploration.
  • Learned policies exhibit significant computational efficiency, allowing for evaluation times two orders of magnitude faster than traditional controllers.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.