Syllabus
Course Syllabus
The course is organized around four questions: How should a robot represent uncertain state? How should it make sequential decisions? How can it learn behavior from data and experience? How do these ideas appear in modern robot-learning systems?
The first three modules develop the mathematical foundations in depth. The final module deliberately connects those foundations to contemporary robot learning rather than treating recent methods as a separate vocabulary.
PDF copy of the syllabus PDF
Module 1: Estimation and Reasoning Under Uncertainty (~11 lectures)
- Background on probability
- Topics: Markov chains, Hidden Markov Models, Kalman Filter, Extended and Unscented Kalman Filter, particle filters, occupancy grids, rigid transformations
- Core question: What should the robot believe about the state of the world, and how should that belief change as new observations arrive?
Module 2: Sequential Decision-Making, Control, and Planning (~6 lectures)
- Background on linear control and dynamic programming
- Topics: Markov Decision Processes, Bellman equation, Value and Policy Iteration, Linear Quadratic Regulator, Linear Quadratic Gaussian control, Iterative LQR, sampling-based motion planning
- Core question: Given a model of the world, how should a robot choose actions over time?
Module 3: Learning to Act (~7 lectures)
- Background on optimization and deep learning
- Topics: Imitation Learning, Policy Gradient methods, Q-Learning, Inverse RL, Model-Based RL, Offline RL, Deep RL
- Core question: When the model or the desired behavior is not known explicitly, how can a robot learn a useful policy from demonstrations, interaction, or previously collected data?
Module 4: From Robot Learning to Modern Embodied AI (~1-2 lectures)
- Topics: generalization and Sim2Real, meta-learning/adaptation, generative robot policies, robot foundation models, vision-language-action models, and world models
- Core question: Which parts of the classical robot-learning stack are being replaced, which are being learned, and which remain fundamentally the same?
- This module is intended as a conceptual bridge. The emphasis is on understanding how the foundations from Modules 1–3 reappear in current systems, not on surveying every recent model.
Tentative schedule
| Lecture | Topic | Notes |
|---|---|---|
| 1 | Introduction | HW 0 out (not graded) |
| 2 | Background on probability | HW 1 out |
| 3 | Markov Chains | |
| 4 | Hidden Markov Models I | |
| 5 | Hidden Markov Models II | |
| 6 | Kalman Filter | HW 1 due |
| 7 | Extended Kalman Filter | |
| 8 | Unscented Kalman Filter | HW 2 out |
| 9 | Particle Filter | |
| 10 | Rigid Transforms, Quaternions | |
| 11 | Occupancy Grids | Summary on Lec 4-10 |
| 12 | Dynamic Programming, Bellman Equation | HW 2 Due |
| 13 | Value Iteration | |
| 14 | Policy iteration | HW 3 out |
| 15 | Background on Linear Control, LQR | |
| 16 | LQG, Iterated LQR | |
| 17 | Midterm | |
| 18 | Sampling Based Motion Planning | HW 3 Due |
| 19 | Optimization, Imitation Learning | |
| 20 | Policy Gradient | |
| 21 | Tabular Q-Learning | |
| 22 | Continuous Q-Learning | HW 4 out |
| 23 | Inverse RL, Model-based RL | |
| 24 | Offline RL | |
| 25 | Deep RL | HW 4 due |
| 26 | Closing topics |