Maybe time for the two communities to learn from each other. The paper: arxiv.org/abs/2508.07400 . As a bonus, they received a CDC 2025 outstanding student paper award. Congrats Mohamad and Alperen!
IRL with time-varying rewards = identification of linear time-varying systems
IRL with switching rewards = identification of switched linear systems
Uncertainty in the optimal policy = process noise in sys id
Uncertainty in model = measurement noise in sys id (hence hard)
My students Mohamad L. Shehab and Alperen Tercan presented a fun paper at CDC last week. Key takeaway: Inverse reinforcement learning and linear system identification problems are more or less the same problem:
Maybe it is time to revise the three little pigs storyβ¦
4) Finally, I had distal radioulnar joint instability (hence the extended casting) and learned that hand surgeons and control theorists share one principle: we respect the unstable.
3) Full body manipulation & contact planning: we use our hands and contact forces to navigate the world all the time. Imagine getting in and out of an economy seat on a plane; or jumping over to sit on a table. When your wrist is not as mobile or weak to bear weight, these tasks become quite hard.
2) Jaw joints (think about Boston dynamic spot or any two finger manipulator) are indeed very useful, for instance for providing support for opening packages.
1) Bimanual manipulation is very powerful. While I initially thought not using my left arm/hand wonβt be a big deal, I quickly realized for many tasks, we use both hands. Trying to do them with one is hard and even impossible at times.
How breaking my wrist and being in a long arm cast for an extended period of time made me appreciate certain robotics problems?
If I recall it right, they used to support a on-demand print option for each book. Not sure if this is something the author can choose. With a few clicks, I found this: services.publishing.umich.edu/Books/Electr... though it seems some recent ones are offered as ebooks only.
This is mainly for ECE textbooks but we have the following: fet.engin.umich.edu
βOptimizing Data for Decision-Makingβ by Asu Ozdaglar (MIT)
βFoundation Models for Autonomous Vehiclesβ by Marco Pavone (Stanford and NVIDIA)
Second (and last) day of L4DC also featured three great keynotes:
βYou Trained with Offline RL---But How to Tune It?β by Nan Jiang (University of Illinois)
βSome variants of gradient dominance conditions motivated by LQR direct policy optimizationβ by Eduardo Sontag (Northeastern University)
βFederated Reinforcement Learning: Statistical and Communication Trade-offsβ by Yuejie Chi (Carnegie Mellon University)
We had three wonderful keynotes today at L4DCβ25:
"If Gen AI is the answer, what is the question? Some thoughts on the theory of generation" by Ambuj Tewari (University of Michigan)
Official start of L4DCβ25
Time for exciting afternoon tutorials
"Capabilities of Large Language Models (LLMs) in Control Engineering" by Bin Hu, Geir Dullerud, Peter Seiler, and Huan Zhang
"The Scenario Approach: Data Science for Decision and Control" by Marco Claudio Campi and Simone Garatti
L4DC kicked off with the morning tutorials
"Higher-Order Learning Dynamics in Games" by Panayotis Mertikopoulos, Lacra Pavel, Jeff S. Shamma
"Neural Networks in the Loop: a Tutorial on Robust Machine Learning for Control" by Ian Manchester, Peter Seiler, Bin Hu
let's keep having fun (although the ride is a bit bumpy these days)
i am happy to announce that, effective august 25, the number of women full professors in my department will increase by 50%. i am afraid some of the external letter writers will still be calling me a "rising star" in my retirement memoir. jokes aside, thanks everyone for the support.
L4DCβ25 early bird registration deadline is on May 2!
L4DCβ25 (7th Annual Learning for Dynamics and Control Conference) program is shaping up nicely. Check out the program, including exciting papers, keynotes, and tutorials: sites.google.com/umich.edu/l4...
Registration is now open! Come join us in Ann Arbor on June 4-6!
BREAKING: People are being suspended on X in Turkey for posting videos of these protests against ErdoΔanβs corrupt and repressive regime.
Keep sharing everywhere.
I am definitely going to the grocery store more often and walking home from there π
No surprise these programs graduated several future Turing Award winners. Congrats to Andrew Barto, the newest among this group.
They had classes in automata theory, AI, info theory, algorithms, optimization, control theory, and even social decision making. It had a perfect Cyber-Physical Systems curriculum before CPS was a thing (well cybernetics was a thing): drive.google.com/file/d/12SEj...
A little bit of interesting history: in late 60s, early 70s, two PhD programs were established at the University of Michigan: computing and communication sciences (CCS) in LSA and computer, information, and control engineering (CICE) in CoE.
I was taking uber from campus to home the other day and it was overly expensive. If instead I put the grocery store almost across my place as the destination, the price dropped to a half. Was it a discount for the grocery store or overcharging me since it learned my commute patterns?
A great workshop for senior PhD
students and postdocs interested in pursuing academic positions: nextprof.engin.umich.edu/nextprof-nex...
It is also somewhat challenging, since, as opposed to simple closed form solutions as in OLS where tools from statistical learning theory have been used recently, these control design algorithms are in the form of semi-definite programs. It is fun to think about sequences of random SDPs :-)
