This is a research-flavored project focused on developing and characterizing a (possibly) novel approach for efficiently adapating world models.
| Week | Slides | Topics | Links |
|---|---|---|---|
| 1 (9/27) | Link | Motivation, Foundations, Setup | --- |
| 2 (10/4) | Link | MPC, Regressions, and Adaptation | --- |
Activate your virtual env first, then run:
python -m webenvs/ drone and pendulum environments; batched true dynamics
models/world_model.py pretrained MLP and checkpoint loader
models/pretrained/ drone.pt and pendulum.pt
control/ CEM optimizer and MPC controllers
adaptation/ BLR stubs and AdaptiveWorldModel residual composition
web/ Flask server and canvas interface
It would be ridiculous and hypocritical for me to require manually coding all, or even the majority, of this project. However, it is important to note that delegating all implementation to an AI tool is known to have a clear negative impact on learning. Also, people have diverse opinions and habits regarding AI usage for code and research, so asserting any strict AI policy is probably a hopeless endeavor.
As a result, I will provide only recommendations for how you should use AI to contribute to this project:
-
Be comfortable with hand-deriving the math. There is (as of writing this) not a burdensome amount of math to know, and most of it is on the critical path for building a good intuition.
-
Understand at a conceptual level the requirements, function, and effects of each subsystem. Knowing exact module implementations and syntax is generally unnecessary, but being able to understand the purpose of large blocks of code is.
-
Treat your agents as pretty accurate explainers and decent writers, not full engineers. Most consumer agents are good at understanding the function of code and writing functional code. However, the decision-making quality for problems that do not already have exact or analogous answers already published is significantly worse than most competent technical people.
Jeffrey Lu - lujeff [at] umich [dot] edu
This project is supported by compute resources provided by MIDAS, U-M ARC High Performance Computing.
This project is run through the Michigan Data Science Team in the Fall 2026 term.