The field of robotic manipulation is attempting to expand from structured warehouse and lab environments to in-the-wild applications in human-centric settings. When deployed in novel settings, a robot may see and need to interact with a new object for the first time. Model-based control approaches rely on a priori dynamics models, an assumption that does not extend beyond highly constrained lab settings. Data-driven control approaches can require immense amounts of data and may only perform well on scenarios in their training data distribution. An alternative that boasts the interpretability of model-based approaches without requiring a model or inordinate amounts of data beforehand is to rapidly build a model of a newly encountered object. In this talk, I will present our work on data efficient dynamics model building through contact-rich interactions. The talk will cover learning dynamics models of single and multi-link objects from state information, theoretical insights into why our loss formulation is more data efficient than naïve alternatives, and how we combine dynamics learning with pose estimation from computer vision.