In such scenarios, model-based controllers would greatly benefit from having accurate and robust dynamics models. Whilst there has been numerous research efforts in learning such models, it is often the case that these frameworks are mainly tested in continuous motion, yielding poor results when deployed on real systems exposed to non-smooth frictional discontinuities.
Inspired by a recent promising direction in machine learning, our work focuses on a newly developed methodology of learning dynamics models undergoing external impact. In this talk, I will present our current and ongoing work in this domain, with our contribution on using physics-guided neural networks for augmenting data-driven deep models with physical consistency.
Title: Decentralized Ability-Aware Adaptive Control for Multi-Robot Collaborative Manipulation