Learning-Based Robotic 1000-2M26LBR

Introduction to Robot Learning; From classical control and motion planning to end-to-end learned policies; where learning helps and where it does not; overview of the current research and industrial ecosystem.
Imitation Learning and Vision-Language-Action models; autoregressive tokenisation vs. diffusion and flow matching action experts
Systems and deployment; inference latency and real-time constraints; safety layers around learned policies; hardware considerations (grippers, cameras, calibration).
Evaluations and data collection for robotics
Offline RL techniques; RECAP
Simulation and sim-to-real transfer; modern simulators and GPU-accelerated training;
Modern robotics failure cases - what we cannot currently do?
RL for locomotion and navigation; what is RL good and bad at?; difficulties of designing reward functions; problems with input modalities;
World Action Models

Course coordinators

Marek Cygan
Mateusz Wyszyński

Requirements

Deep neural networks
Robot control

Prerequisites (description)

- Solid grounding in machine learning and deep learning, including transformer architectures. - Basic knowledge of reinforcement learning (MDPs, policy gradient methods, Q-learning). - Proficiency in Python; comfort with Linux and git. - Helpful but not required: experience with ROS2, a robotics simulator (e.g., MuJoCo, Isaac), or prior hands-on hardware experience.

Assessment criteria

In total, you can get from 0 to 100 points. 51 points are needed to pass the course.

The final grade is based on:
Project: 0 - 40 pts
Exam: 0 - 60 pts

Grades:
51-60: (3) dst
61-70: (3.5) dst+
71-80: (4) db
81-90: (4.5) db+
91-100: (5) bdb

Additional criteria for PhD students: PhD students are additionally required to:

- Include in the project a clearly identified research contribution beyond reproducing existing work (e.g., a novel method variant, a systematic ablation, or a new evaluation protocol), together with a concise related-work survey positioning the contribution.
- Prepare the final report at the quality level of a workshop paper submission.
- Deliver one seminar-style talk during the semester presenting a recent research paper or their own ongoing work.

Bibliography

K. M. Lynch, F. C. Park, Modern Robotics: Mechanics, Planning, and Control, Cambridge University Press, 2017.
R. S. Sutton, A. G. Barto, Reinforcement Learning: An Introduction, 2nd ed., MIT Press, 2018.
P. Holderrieth, E. Erives, An Introduction to Flow Matching and Diffusion Models, lecture notes for MIT Class 6.S184, 2025. arXiv:2506.02070.
Y. Lipman et al., Flow Matching Guide and Code, 2024. arXiv:2412.06264.

Additional information

Information on level of this course, year of study and semester when the course unit is delivered, types and amount of class hours - can be found in course structure diagrams of apropriate study programmes. This course is related to the following study programmes: