Shuo Yang

I gave an in-depth tutorial on "Formal Verification and Control with Conformal Prediction" at KTH today 🚀 Since I got positive feedback, I wanted to share the presentation, in the hope that others can also benefit from it 🙂 Find the recording here: youtu.be/kfPBjaMCXmM?si…

It enables faster training without sacrificing performance, and makes online training of overparameterized policies practical and efficient. 🔗 Project page: apollo-lab-yale.github.io/25-RSS-DRIFT-w… Check it out — happy to connect & discuss!

Excited to share that our work DRIFT has been accepted to #RSS2025! Inspired by PEFT, DRIFT dynamically adjusts trainable ranks via SVD under the Interactive IL framework, balancing overparameterization and intrinsic low rank structure. 📜 Paper: arxiv.org/abs/2502.03822

What a memorable way to wrap up my PhD: I had the privilege of presenting my work on safe human-centered robotics at the GRASP SFI seminar series @PennEngineers Huge thanks to the @GRASPlab for inviting me and to everyone who attended! youtube.com/watch?v=50ZId5…

We did a big update of our survey "Formal Verification and Control with Conformal Prediction" (arxiv.org/pdf/2409.00536). Among other things, we now provide a detailed comparison of conformal prediction with UQ alternatives. Big thank to the reviewers for really good suggestions!

@xiatao_sun @djrakita10 please see more simulation and real-world experimental results here: arxiv.org/pdf/2502.03822

excited to share our #RSS2025 work on online diffusion policy learning, a collaboration with @xiatao_sun (lead), Yinxing Chen, Francis Fan, Yiyan (Edgar) Liang, and @djrakita10

@xiatao_sun @djrakita10 We then propose an interactive IL algorithm (called DRIFT-DAgger) that combines offline bootstrapping with efficient online adaptation (w/ rank modulation and scheduling), enabling effective integration of expert feedback during the novice policy training

@xiatao_sun @djrakita10 rank modulation: we leverage Singular Value Decomposition to decompose weight matrices into components designated as either frozen or trainable ranks; we use rank scheduler (e.g., linear, cosine, sigmoid, exponential, etc) to coordinate the dynamic adjustment of trainable ranks

@xiatao_sun @djrakita10 to enable online adaption of diffusion policies, we propose two novel components: rank modulation and rank scheduler that exploit the intrinsic low-rank structure of overparameterized models, balancing training efficiency and model performance

Check out the new work of my student @ZhaoYiqi87035 on distributionally robust predictive runtime verification (aka failure prediction/detection) for multi-agent systems 🚀🚀🚀 Joint work with TRINA and @bardhhoxha

We are excited to share our latest paper on predictive runtime verification of stochastic multiagent systems against spatio-temporal logic specifications: arxiv.org/abs/2504.02964.

Our 2025 RSS workshop on "Statistical Uncertainty Quantification in the Era of AI-Enabled Robots" got accepted. We have an amazing lineup of tentative speakers, see sites.google.com/view/rss2025-w… 🚀 The workshop will be held at USC on June 25th (and we guarantee excellent weather 🏖️🌴)

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

My colleague Jyo Deshmukh and I gave a short tutorial on "Formal Verification and Control with Conformal Prediction: Practical Safety Guarantees for Autonomous Systems" at the NSF CPS PI meeting. We were asked for slides and made them available here: tinyurl.com/rcx37pj6 🙂

🔐 Can we ensure AI safety through the lens of safe control? 💥 We model the multi-turn LLM conversation as a neural dialogue dynamical system and introduce a Neural Barrier Function (NBF) to safely steer LLMs against multi-turn jailbreaks. arxiv: arxiv.org/abs/2503.00187

Paper: arxiv.org/pdf/2412.01656 Code: github.com/shuoyang2000/S… Webpage: sites.google.com/view/stlgame Huge thanks to all collaborators! Hongrui Zheng @billy_bz, Cristian-Ioan Vasile, George J. Pappas @pappasg69, Rahul Mangharam

In the drones experiment, we consider the STL spec where the ego drone should: 1) eventually visit the goal, 2) follow altitude rules, 3) avoid the unsafe set, and d) always keep a safe distance with opponent drone. The fictitious self-play can guide us to approx-Nash policy 6/n

In our new @l4dc_conf work, we address the problem of robust and safe policy synthesis under complex tasks in adversarial settings against unknown dynamic agents. We propose a game-theoretic framework to ensure the task is satisfied as much as possible even in the worst case 1/n