Reproduce “MetaTune: Adjoint-based Meta-tuning via Robotic Differentiable Dynamics” on OpenResearch

About this paper

Xiexin Peng, Bingheng Wang, Tao Zhang, Ying Zheng

Researchers at Huazhong University of Science and Technology and National University of Singapore introduce MetaTune, a unified adjoint-based meta-learning framework that jointly and automatically tunes robotic feedback controllers and disturbance observers using differentiable dynamics. This approach achieved superior tracking accuracy, reducing RMSE by 15-20% at higher speeds and up to 40% under strong disturbances in sim-to-sim transfer experiments on a quadrotor.

Key insights

• Adjoint-based gradient computation is essential for computational efficiency and physical consistency in differentiable control, especially for long horizons and complex systems.
• Joint optimization of coupled controller and observer parameters, rather than decoupled tuning, is critical for superior disturbance rejection and stable tracking performance.
• Neural policies, leveraging physics-informed adjoint gradients, can achieve highly adaptive gain scheduling that generalizes well across tasks and environments.

Summary by alphaXiv.