Lecture Note: Probabilities, Energy, Boltzmann & Partition Function
pdf Probabilities & Energy.
This course is usually a lab course, working directly with Panda robots. This year we'll have to do it completely in simulation.
I will polish our code at github as the basis. In the past (at U Stuttgart), students mostly used the python bindings to code their systems. But the underlying library is C++ and also provides a directly C++ interface very similar to the python interface.
Usually we used this code to control the robots directly -- this year we will use a physics simulator plugin (nvidia PhysX or bullet, swappable) to model the environment, while the robot is kinematically modeled.
The goal challenge will be for your to code robot(s) to literally build things (towers, high constructions, whatever) from pieces in the environment. This will require you to integrate perception and robot manipulation with long term planning -- which is also a goal of research of our lab (See https://ipvs.informatik.uni-stuttgart.de/mlr/lgp/ for some older videos -- there are more results in the pipeline.)
To work towards such a final project, the course will alternate between two sessions each weak: One introduction/tutorial session, where I'll introduce the concepts and explain the next exercise problems, then help you with first steps with the code. And a second session where you work on your exercises, can ask me or the tutor for help, or present solutions to your exercises. The basic four steps in the course are 1) learn to generate/design robot motion using inverse kinematics or path optimization, 2) learn to extract objects from images (plain OpenCV for a start), 3) combine perception with motion generation by getting object representations into your world configuration representation on the fly, 4) work towards sequential manipulation (mostly pick-and-place) to build statically stable constructions.
The course is very practical, as you can see. But I strongly believe that this practical experience makes you better realize what are the fundamental, also theoretical open challenges in robotic AI.
Prerequisites: Knowledge basic robotics (kinematics) and good training in 3D geometry and linear algebra are necessary. Good coding skills in Python or C++. And Ubuntu 18.04 machine! (Or really good knowledge in docker or VMs -- but I will not provide direct support for these.)
With this course I intend to use selected state-of-the-art publications from the strongest conferences in robotics and AI/ML for teaching. However, unlike in a seminar, this is not just about a student presenting a paper. Instead, I would like to take a scientific paper as a starting point and from that go backwards and have tutorials and lectures that introduce the background and used methods until the whole class really understands that paper.
I have never given such a course, but want to try it to better connect research with teaching, and avoid the often cursory discussions in seminars.
You will have to read papers, prepare mini-lectures on some of the background topics, and work on exercises. Details are yet open.
Prerequisites: Advanced knowledge in robotics, potentially including reinforcement learning or computer vision.
pdf Probabilities & Energy.
Die gängigen Erklärungen zu “Was ist Informatik?” – etwa von der Gesellschaft für Infomatik, der TU Dresden, oder auf Wikipedia – machen es einem schwer, sic...