Meets :          Tuesdays and Thursdays  2-3:30pm
Instructor :  Souradeep Dutta   (souradeep [at] ece.ubc.ca )
Office Hours : Wednesdays 11am-1pm , KAIS 4034

Prerequisites :

1. A graduate level machine learning course 

2. Undergraduate level exposure to robotics

3. Undergraduate level exposure to programming

Grading :  30% on course project , 30% on assignments,  20% on class participation, 20% on paper presentation

Topics Overview : Fundamental ideas on reinforcement learning and behavior cloning will be discussed. The aim is to understand the theoretical reasons why reinforcement learning is highly sample inefficient. Even though it can sometimes be really effective.  Behavior cloning has its benefits in terms of samples required, however can have generalization issues. Especially when learning from a few demonstrations. 

Next, we take a deep dive into the different frameworks of the imitation learning approaches, and the specific problems they solve. And understand their short comings. There has been a renewed interest in this line of development recently after the success of GPT.  Since much of RL is not practical for real robotic systems owing to its risks and cost. This course will provide a fresh set of algorithmic tools to the students beyond the ones offered by a standard machine learning course.