Safety Guarantees in Imitation Learning
To design an online filter for the control inputs to avoid unsafe states, which improves iteratively
Guide: Prof. Florian Shkurti, University of Toronto.
- Conducted extensive literature survey on Lipchitz Neural Networks, the use of Normalizing Flows to learn stochastic differential equations and design of NN controllers for linear systems using Linear Matrix Inequalities
- Examined subspace projection techniques of learning dynamics to impose stability architecturally
- Recreated the results of in-Distribution Barrier Function (iDBF) on a custom toy problem
- Improved the safety framework of iDBF by generating contrastive distribution using Normalizing Flows
The barrier we were able to learn based on modifications to the in-Distribution Barrier Function