*This was a HYBRID Event with in-person attendance in Levine 512 and Virtual attendance…
In contrast to traditional control systems where detailed dynamics models are constructed from a mix of physical understanding and empirical data, machine learning for intuitive physics, reinforcement learning, and robotics often takes a hands off approach treating the dynamics as a black box with little to no assumed structure. We show how desirable high level properties like symmetries, energy and momentum conservation, and other constraints can be reintroduced into these models to improve generalization. These high level attributes represent prior knowledge about the underlying physics of the system in the Bayesian sense, and can even be incorporated in a way that does not limit the flexibility of the model.