10-701 Student Presentation - Robust Iterative Policy Search for Aerobatic Helicopter Flight

Jiaji Zhou and Kumar Shaurya Shankar

In many real robotics examples where dynamics and sensor data are noisy, the robustness of control algorithm is essential. Additionally, most physical systems are high-dimensional and it is highly unlikely to have data for all parts of the state-space. Since it is unlikely for any formulated model to completely capture all of the underlying system dynamics, learning control algorithms should exhibit some degree of robustness to undermodeling.

In this research project we intend to explore a hybrid approach to solving a control problem through a reinforcement learning approach; specifically, we propose to combine Iterative Linear Quadratic Regulator methods with Policy Search by Dynamic Programming.