Biologically-Inspired Learning for Humanoid Robots

Lecturer (assistant)
Number0000002585
Type
Duration4 SWS
TermSommersemester 2019
Language of instructionEnglish
Position within curriculaSee TUMonline
DatesSee TUMonline

Dates

Admission information

See TUMonline
Note: -

Objectives

"After this course, students are capable of: - Using the robot operating system (ROS) with the NAO robot. - Understand the main biological mechanisms responsible for learning. - Implement and evaluate biologically-inspired algorithms for sensory-motor mappings and reinforcement learning."

Description

"1) Introduction - Motivation - Human brain research - Human brain inspiration in motor control 2) Learning - Why humanoids need to learn? - What humanoids need to learn? - Learning algorithms -- Supervised learning -- Unsupervised learning -- Reinforcement learning - Learning by self-exploration - Learning by demonstration 3) The cerebellum - Facts - Anatomy - Neural circuitry - Effects of cerebellar disease 4) Computational model of the cerebellum - Associative memory - Cerebellar model articulation controller (CMAC) 5) The basal ganglia - Anatomy and major components - Projections from and to other brain regions - Direct and indirect pathway - Basal ganglia loops 6) Reinforcement learning (RL) - Characteristics - Reward - Agent and environment - Major components of a RL agent - Temporal difference learning 7) Self-organizing maps (SOMs) 8) The central pattern generator (CPG) - Biological approach - Computational model - Multilayered CPG"

Prerequisites

C/C++ programming skills

Teaching and learning methods

"Following teaching methods are used: - Lectures that provide the necessary theory for the tutorials. - Tutorials with laboratory assignments which ensure that major parts of the taught content, e.g. learning algorithms, are realized and tested on the robots."

Examination

"The final grade is determined by the laboratory/tutorial assignments (40%), paper presentations (20%), and the oral exam (40%). Laboratory/tutorial assignments evaluate the students ability to implement and test the approaches and methods for robot learning presented in the lectures. The students have to show their results on the robots by demonstrating specific robot capabilities. Paper presentations verify whether the students can describe biologically-inspired methods and algorithms in a concise manner. The oral exam evaluates the students understanding of the biological principles responsible for learning. We offer 2 blocks of exercises, which students have to solve it individually. At the end of the module each student presents a relevant research paper from the related literature in the field. The final grade is composed as follows: A) final oral exam: 30% B) laboratory assignments : 30% C) individual laboratory assignments : 30% D) paper presentation: 10% "

Recommended literature

R. S. Sutton and A. G. Barto: "Reinforcement learning: An introduction", MIT Press, 1998.

Links