Throw‑flipping an object to a precise landing pose is very challenging, even for humans, due to parasitic rotation: to throw farther, the arm inevitably rotates faster, causing excessive object spin.
In this work, we present a learning‑from‑experience framework that combines impulse‑momentum control with physics‑based data assimilation.
The impulse‑momentum principle provides extra control authority to decouple translational and rotational motion, while physics‑informed data assimilation reduces sample complexity by 40% and enables knowledge transfer across objects with different centers of mass.
Reference: Yang Liu, Bruno Da Costa, and Aude Billard, Learning to Throw‑Flip, IROS 2025.
Paper: https://arxiv.org/abs/2510.10357
