We are back for another round with the 2nd Annual Northeastern Systems and Control Workshop (NESCW), which will be held @CUSEAS @Columbia on May 3rd, 2025! If you are a systems and controls researcher in the northeast, you are invited! @CSSIEEE @IFAC_Control @l4dc_conf

🚨🚨🚨 New paper led by @ult_flymachine and @LarsLindemann2, with my student Eliot Shekhtman, on applying conformal prediction in interactive environments! Check out details in the quote tweeted thread below! 👇

While this work focuses on interaction-induced distribution shifts, we believe that it provides interesting ideas to address other control problems where distribution shifts arise and exchangeability assumptions are violated, prohibiting the use of vanilla statistical tools💡

This adjustment is performed based on a policy-to-trajectory sensitivity analysis, resulting in a safe, episodic planner. We further conduct a contraction analysis of the system providing conditions under which both the CP results and the policy updates are guaranteed to converge

We realize this via adversarially robust CP where we perform a regular CP step in each episode using observed data under the current policy, but then transfer safety guarantees across policy updates by analytically adjusting the CP result to account for distribution shifts.

In this paper, we address this “chicken-and-egg” problem and propose an iterative framework that robustly maintains safety guarantees across policy updates by quantifying the potential impact of a planned policy update on the environment's behavior.

A lot of recent works have used conformal prediction (CP) to generate distribution-free safety guarantees using observed data. However, CP's assumption on data exchangeability is violated in interactive settings, resulting in the loss of any safety guarantees

This creates an intricate coupling and interaction-induced distribution shifts where changing the control policy changes the behavior of uncontrollable agents, and vice versa, thereby invalidating safety guarantees in existing work.

For instance, the control of a self-driving vehicle among pedestrians and human-controlled vehicles. In this setting, the behavior of uncontrollable agents are unknown and interactive, that is, they may react to the behavior of the autonomous agent.

Not only Omid, but also Eliot Shekhtman and @NikolaiMatni played a central role in getting this paper together ☺️ Specifically, our paper deals with the "chicken-and-egg" problem that arises when designing control policies for autonomous agents in interactive environments 🐔🥚

We learned acceleration algorithms for fast parametric convex optimization. Only 10 training instances used for each example and robustness is guaranteed with PEP! Joint work w/ Jinho Bok, @NikolaiMatni, @pappasg69 Paper: https://t.co/yeVf2921dT Code: https://t.co/JgyMqBEZC2

I cannot believe that’s how we lost.

Excited to announce our NeurIPS ’25 tutorial: Foundations of Imitation Learning: From Language Modeling to Continuous Control With Adam Block & Max Simchowitz (@max_simchowitz)

Hurry up and register, early bird deadline is tomorrow!

Going to @IEEECDC2025, in Rio? 🌴☀️ Don’t miss our workshop on Control Architectures Theory (CAT) – Dec 9, right before the main conf! With an amazing speaker lineup + organizers Aaron Ames, @NikolaiMatni & me. 👉 https://t.co/0ntebFSLiz #CDC2025 #Autonomy #ControlArchitectures

Celebrating the successful PhD defense of David Snyder from our group at @Princeton! David is off to @Penn as a postdoc next, working with @NikolaiMatni and @pappasg69.

And a dedicated invited session has also been accepted, organised with @pappasg69 , Aaron Ames, and @NikolaiMatni 🚀

Excited to share that after last year's success, the second edition of our Workshop on Control Architectures Theory (CAT), co-organized with @NikolaiMatni and Aaron Ames has been accepted to @IEEECDC2025. Stay tuned for more info, Rio here we come! 🌏

I’ll be presenting our paper “On The Concurrence of Layer-wise Preconditioning Methods and Provable Feature Learning” at ICML during the Tuesday 11am poster session! DL opt is seeing a renaissance 🦾; what can we say from a NN feature learning perspective? 1/8

🧵generative models are sweet, but navigating existing repositories can be overwhelming, particularly when starting a new research project so i built jax-interpolants, a clean & flexible implementation of the stochastic interpolant framework in jax https://t.co/GVaL3iPmFk