IN2107 Variational Inference in Robotics

Lecturer (assistant)
  • Patrick van der Smagt [L]
  • Furkan Erdem
  • Abdalla Swikir
Number0000001967
TypeSeminar
Duration2 SWS
TermSommersemester 2023
Language of instructionEnglish
Position within curriculaSee TUMonline

Admission information

Objectives

At the end of the course, students are expected to ● have an understanding of machine learning methods for control, in particular latent-variable models; ● be able to comprehend, critically analyze, and implement state-of-the-art research publications related to these topics.

Description

The seminar focuses on variational inference methods that are applicable to robotic perception and control problems. In recent years, applying machine learning to sequential data of systems has developed as a promising path to the application of control methods to very complex systems, such as robotic systems perceived through noisy sensors. One promising direction is that of variational inference, where high-dimensional and hence intractable problems of Bayesian inference can be approximated through stochastic optimization techniques. Here, the member of a family of distributions closest to the true posterior distribution is determined. This approximate distribution can then be used in place of the true one, yielding solutions to learning, inference or control problems such as system identification, state estimation, manipulation, localization or navigation. In this course, students will familiarize themselves with the methodology of variational inference in a deep learning framework. For that, fundamental knowledge such as machine learning, control, and robotics will be refreshed. Typical problems of robotics will be phrased as probabilistic inference and solutions based on recent developments will be studied and discussed. Examples are Dreamer, Learning To Fly and variational inference-based SLAM: - https://danijar.com/project/dreamer/ - https://argmax.ai/blog/drone/ - https://argmax.ai/blog/vislam/

Prerequisites

● Math ○ Basic linear algebra, probability theory, calculus ● Machine Learning ○ Neural networks ○ Gradient-based optimization ○ Python auto-differentiation packages like TensorFlow, PyTorch, or Jax. ● Control ○ Markov Decision Processes, Bellman’s Equation (optional) ○ PID control (optional) ● Robotics ○ Forward kinematics and dynamics (optional) ○ Simultaneous localization and mapping, SLAM (optional) ○ Navigation (optional)

Teaching and learning methods

The course will start with four sessions that introduce the core concepts to understand current research trends. After that, the seminar will pause for four weeks to let the first students prepare their presentations. After that, students will present on a weekly basis. We will provide a selection of novel and exciting research publications. These will cover a broad range of topics and difficulty levels, which should guarantee a good choice for every participant. Students will pick one of these papers. Their task is to read, understand, and analyze the paper. The analysis can include a potential (re-)implementation, reproduction of (partial) results, interesting questions not tackled in the paper, a comparison to related papers. Finally, the student will present the paper and findings to fellow classmates. The students are encouraged to extend their presentations with showcases of their own programming.

Examination

Evaluation will be based on the live presentation during the seminar. The presentation is supposed to take 30 minutes with 10 minutes of questions by the other students and the instructors. Criteria are - Understanding of the material presented, (50%) - Quality of the slides, (15%) - Quality of the presentation, (15%) - Ability to engage with the audience during the questions, e.g. whether questions can be answered or the student clearly knows about the limits of their knowledge. (20%) The course cannot be retaken.

Recommended literature

● Simo Sarkka, Bayesian Filtering and Smoothing ○ https://users.aalto.fi/~ssarkka/pub/cup_book_online_20131111.pdf ● Kevin Murphy: Probabilistic Machine Learning: An Introduction ○ https://probml.github.io/pml-book/book1.html ● Ian Goodfellow and Yoshua Bengio and Aaron Courville: Deep Learning ○ https://www.deeplearningbook.org/ ● Richard S. Sutton and Andrew G. Barto: Reinforcement Learning: An Introduction ○ http://incompleteideas.net/book/the-book.html ● Steve Brunton: Control Bootcamp ○ https://www.youtube.com/watch?v=Pi7l8mMjYVE ○ Based on chapters 1-3 from ■ Thomas Duriez Steven L. Brunton Bernd R. Noack: Machine Learning Control – Taming Nonlinear Dynamics and Turbulence (http://faculty.washington.edu/sbrunton/mlcbook/)

Links