Robot Learning

Relations:

In order to register for the course B3B33VIR, the student must have registered for the required number of courses in the group BEZBM no later than in the same semester.

It is not possible to register for the course B3B33VIR if the student is concurrently registered for or has already completed the course B3B33UROB (mutually exclusive courses).

It is not possible to register for the course B3B33VIR if the student is concurrently registered for or has previously completed the course B3B33UROB (mutually exclusive courses).

The requirement for course B3B33VIR can be fulfilled by substitution with the course B3B33UROB.

Course guarantor:

Lecturer:

Tutor:

Supervisor:

Department of Cybernetics

Synopsis:

The course teaches application of machine learning methods and optimization on well-known robotic problems, such as semantic segmenation from RGB-D data or reactive motion control. The core of the course represents teaching of deep learning methods.

Stidents will use basic knowledge from optimization and linear algebra such as robut solving of overdetermined systems of (non)linear (non)homogenous equations or gradient minimization methods. The labs are divided into two parts, in the first one, the students will solve basic tasks in PyTorch, in the second one, individual semestral work.

Requirements:

Syllabus of lectures:

1. Overview and lecture outline.

2. Regression ML/MAP

3. Classification ML/MAP

4. Neural networks, backpropagation

5. Convolution leyer, backpropagation

6. Normalization leyer (BatchNorm, InstanceNorm, ...) a backpropagation

7. Training I (SGD, momentum and their convergence ratio)

8. Training II (Nester gradient, Adam optimizer, activation function impact on optimization problems)

9. Architectures of deep neural networks I: detection (yolo), segmentation (DeepLab), classification (ResNet)

10. Architectures of deep neural networks II: pose regression, spatial transformer nets.

11. Generative Adversarial Networks, Cascaded Refinement Networks, Style Transfer Networks

12. Reinforcement learning in robotics (policy gradient, imitation learning, actor-critic, aplications)

13. Learning from weak annotations (weak-supervision, self-supervision)

14. Presentation of semestral works

Syllabus of tutorials:

In the first half of labs, the students will solve basic tasks in PyTorch, in the second one, the students will work on individual semestral works.

Study Objective:

The course teaches application of machine learning methods and optimization on well-known robotic problems, such as semantic segmenation from camera and deep images or reactive robot control. The core of the course represents teaching of deep CNN application methods.

Study materials:

Ian Goodfellow and Yoshua Bengio and Aaron Courville, Deep learning, MIT press, 2016 http://www.deeplearningbook.org

Note:

Further information:

https://cw.fel.cvut.cz/wiki/courses/b3b33vir/start

No time-table has been prepared for this course

The course is a part of the following study plans:

• Cybernetics and Robotics 2016 (compulsory elective course)