Course:
Course Program:
Instructor: Emre Ugur (contact)
Textbook: No textbook, but please see the list of readings below.
Lectures: Monday 12:00-15:00
Location: Computer Engineering Dept, A5
Mailing-list: Send email if not automatically registered.
Note: Machine Learning and Robotics background is desired but not a must. Send cconsent request.
Assessment:
- Homeworks (Coding: GMM, HMM, Q-Learning, Policy gradient, Clustering+SVM)
- Final Exam
- Project
Aim:
This course is about (i) "general" approaches that aim development of robot intelligence, and (ii) "more focused" advanced learning methods that endow robots with a particular set of sensorimotor skills. To get a grasp of general approaches, we study the general framework of developmental robotics, active learning and intrinsic motivation, bottom-up skill development and symbol acquisition. These approaches are formulated mostly in an interdisciplinary manner, in relation to the findings from infant development, human information processing, experimental and ecological psychology. For the latter one, we will focus on particular methods that are effective in learning manipulation skills and sensorimotor representations such as learning by demonstration, grasp learning, and probabilistic modeling. Rather than detailed analysis of the Machine Learning methods, we will focus on their exploitation for different robot learning problems.
Schedule:
| Course Overview, Reinforcement Learning, | |
| Reinforcement Learning Cont'd, | |
| Developmental Robotics, | |
| Affordance Learning, | |
| Intrinsic Motivation, | |
| Data for HW2, | |
| IM cont'd, signal->symbol, | |
| Learning from Demonstration, DMP, | |
| DMP cont'd, probabilistic methods in LfD, |
Expected Outcome:
- Breath: Through introductory material presented by the instructor, and paper presentations by students: An overview of the state-of-the-art methods in robot learning, particularly in developmental robotics, learning by demonstration, policy search methods and reinforcement learning, probabilistic modeling of sensorimotor experience, grasp synthesis algorithms.
- Depth: Through implementing and extending many of machine learning methods in robotic problems and one of the advanced methods/papers in the final project, an in-depth knowledge on the corresponding method and the robot learning task.
Final Project: Students might suggest implementing a completely novel approach or choose replicating and extending one high-impact paper. A final project report written in conference/workshop paper format is expected. The topics are not limited to the ones discussed in the lectures as long as they are related to robot learning.