Using Physics as Prior in Imitation Learning.
Abstract
System identification, a methodology to obtain a dynamic model has been studied extensively in the classical robotics field. Approaches to system identification can be naively divided into two branches: Whitebox models and Blackbox models. In this work, we attempt to learn a model that brings advantages of both branches called Deep Lagrangian Networks through human demonstration to obtain the dynamic model of a humanoid. This model will inherently abide Lagrangian Mechanics and thus obey the laws of physics. We then discuss ideas on how this model can be used in a model-based reinforcement learning setting to generate a physically plausible policy in a sample-efficient way.
Contribution
- Explore if a physically plausible dynamic model can be learned from demonstration using a Deep Lagrangian Network.
- Discuss if the learned model can be used in model-based reinforcement learning to generate a good policy in a sample efficient manner.