Reinforcement Learning for Robotics (Lecture with Project)

Lecturer	Alejandro Agostini
Allocation to curriculum	See TUMonline
Offered in	Wintersemester 2020/21
Semester weekly hours	4
Scheduled dates	See TUMonline
Contact	Alejandro Agostini (alejandro.agostini@tum.de)

Content

The course will cover the following topics:

1. Introduction to reinforcement learning (RL): Markov decision process, dynamic programming, Q-learning, SARSA, Actor-Critic, policy-based RL, value-based RL.

2. Reinforcement learning in continuous state-action spaces. Function approximation problem.

3. Reinforcement learning for robotics: mission and problems. Optimal control. Biased sampling, risk of damage, ware-out problem.

4. Model-free reinforcement learning (GMMRL, PI2).

5. Model-based reinforcement learning (PILCO, PI-REM).

6. Approaches combining nonlinear optimal control (ILQR, MPC) and reinforcement learning.

7. Introduction to deep reinforcement learning (end-to-end approaches).

Previous Knowledge Expected

Fundamentals of Linear Algebra, Probability and Statistics, Programming skills in Matlab/SImulink

Objective

At the end of this course, students are able to:

- Implement machine learning algorithms for robots and autonomous systems.

- Evaluate the performance of reinforcement learning algorithms.