Data collection is a major obstacle to applying deep learning methods to robotics. In this thesis, we propose to perform a quantitative study of the use of synthetic data from 3D software and physics simulators to train machine learning models that can later be deployed on physical systems.
Safe Model Predictive Control (Safe MPC) aims to ensure that a physical system’s safety constraints are satisfied with high probability. Our research is on extending constrained MPC methods to cope with probabilistic safety constraints. We further research modeling uncertainty of dynamics to ensure safe exploration when combined with safety constraints learned in simulation, and learning powerful data-efficient surrogate models for complex dynamics.