Jing-Yan Liao

Physical Intelligence turns terabytes (TBs) of robotics sensor and video data into high-value training datasets. Before Ray on Anyscale, scaling multimodal data pipelines across diverse hardware slowed progress and demanded deep infrastructure expertise, pulling developers away from differentiated research work. With Anyscale, Physical Intelligence enables more than a dozen of its developers to run experiments concurrently and deploy workloads seamlessly across a multi-cloud environment, accelerating iteration and keeping teams focused on their mission to build advanced AI for robotics. Read the full story: https://lnkd.in/g29uj4vi

29

Fengqing Jiang (left) and Yichen Feng (right), graduate students in the lab of UW ECE Professor Radha Poovendran, have received a Best Paper Award at Agents4Science2025 — a virtual academic conference where artificial intelligence (AI) agents are the primary authors and reviewers for submitted research papers. Jiang and Feng received the award for their paper, "BadScientist: Can a Research Agent Write Convincing but Unsound Papers that Fool LLM Reviewers?" This paper showed that a fabrication-oriented research agent can generate persuasive, yet unsound, papers accepted by current large language model (LLM)-based reviewers, revealing a critical vulnerability in the AI agent-assisted scientific research paradigm and underscoring the need for stronger safeguards. The Agents4Science2025 award recognized this paper's impact on research integrity and UW ECE's leadership in responsible AI development. Jiang and Feng's paper was also ranked in AI World as the second most notable AI research paper for the week of October 26, 2025. Jiang has received Honorable Mention for Best Paper at the 2025 International Conference on Learning Representation (ICLR 2025) Bi-Alignment Workshop for a paper describing "Safechain," a work that inspects the safety of large language models with long-chain-of-thought reasoning capabilities. In addition to working in Poovendran's lab, Feng is a student in the UW ECE Professional Master's Program. Poovendran is the UW lead for the National Science Foundation's AI Institute for Agent-based Cyber Threat and Intelligence Operation, or ACTION Institute. Jiang is the UW representative of the Student Advisory Committee at the ACTION Institute, which provided support for the research described above. Learn more: https://lnkd.in/eWn56qrr #AI #artificialintelligence #LLMs

58

4 Comments

Thrilled to share that our paper “Reproducibility in the Control of Autonomous Mobility-on-Demand Systems” has been accepted to IEEE Transactions on Robotics (T-RO)! 🎉 This work was led by my Ph.D. students Xinling Li and Meshal Alharbi (proud of them!!!), and brings together a big group of colleagues and friends who have been spearheading work in this area since the early days. It’s special to see so many of these threads come together in one place. Daniele Gammelli, James Harrison, Filipe Rodrigues, Maximilian Schiffer, Marco Pavone, Emilio Frazzoli, Jinhua Zhao We focus on autonomous mobility-on-demand (AMoD) systems (i.e., fleets of self-driving vehicles operating at city scale) and tackle a simple but crucial question: 👉 How can we make results in this area truly reproducible, transparent, and comparable? This is not just an academic concern. As AI-driven systems increasingly make decisions about transportation, energy, logistics, and other society-critical infrastructure, we need to be able to trust, audit, and reproduce the algorithms that control them. In the paper, we: - Propose a system-level reproducibility framework for AMoD, from modeling and control to simulation and evaluation. - Identify common sources of irreproducibility in assumptions, datasets, simulators, and algorithms. - Offer practical guidelines and a reproducibility checklist for researchers and practitioners. Our hope is that this paper can support a more transparent, rigorous research culture for future urban mobility and large-scale autonomous systems. Looking forward to sharing the camera-ready version soon and continuing the conversation with the community! https://lnkd.in/er8MxC9s

150

3 Comments

I returned to the #UniversityOfMinnesota on Thursday to talk about Spectral Properties of Graph Neural Networks (GNNs). The message here is that graph convolutions have trouble handling high frequencies, defined as signals that change quickly over the graph. This observation has implications on the stability of GNN outputs to graph deformations and the transferability of GNNs across scales. Stability and transferability properties of GNNs are analogous to stability and transferability properties of CNNs. This is because GNNs and CNNs are algebraically equivalent objects. (See references below). A recording of this talk delivered at the #SimonsFoundation: https://lnkd.in/e4Rxw-gX My graph neural networks course at #Penn: https://gnn.seas.upenn.edu Stability properties of GNNs: https://lnkd.in/en-UH8jj Transferability properties of GNNs: algehttps://lnkd.in/efrnDJFM Algebraic neural networks: manifold https://lnkd.in/eDpxp-tj Manifold convolutions: https://lnkd.in/eraqxeDP Learning by transference: https://lnkd.in/etaPVQVS Learnable perception action communication loops: chttps://lnkd.in/ejZ5qKkC Covariance Neural Networks: https://lnkd.in/ee2CBweA For those of you that have never lived in #Minnesota, it is always a treat to return to the best state in the union.

143

1 Comment

🚀 Code Release: Phys-Instruct is now open-source! Following our recent paper, "Ultra Fast PDE Solving via Physics Guided Few-step Diffusion", we are excited to announce that our official implementation is now publicly available! We built this repository not just to share our method, but to serve as a comprehensive toolkit for diffusion-based PDE solving. It provides a unified framework for benchmarking, experimentation, and reducing the barrier to entry for physics-guided generative AI. 🔗 Links: 💻 Code: https://lnkd.in/g-3himQC 📄 Paper: https://lnkd.in/gWBQfsHe 📂 Pretrained Models: https://lnkd.in/gVJdxhDc 👥 Authors: Xiangrui Kong, Yueqi Wang, Haoyang Zheng, 罗维俭 (Weijian Luo), and Guang Lin ✨ What’s included in the release? ✅ Unified Framework: A single codebase supporting 5 different baselines and approaches, including Standard Diffusion (EDM), Physics-Informed Training (PIDM), and Inference-Time Guidance (DiffusionPDE, CoCoGen). ✅ Multi-PDE Support: validated on 5 benchmark PDEs: Darcy Flow, Burgers’, Helmholtz, Poisson, and Navier-Stokes. ✅ Production-Ready: Includes single & multi-GPU distributed training support, step-by-step instructions, and data generation scripts. ✅ Pretrained Models. 💡 Why does this matter? This release allows researchers and practitioners to: 🔹 Reproduce results end-to-end with ease. 🔹 Benchmark different physics-guided diffusion strategies under one roof. 🔹 Deploy few-step diffusion solvers as fast, reusable priors for downstream tasks. We hope this codebase accelerates research on diffusion models, scientific machine learning, and fast PDE solvers. Feel free to reach out if you have questions, ideas, or would like to collaborate! #AI #ScientificComputing #DiffusionModels #PhysicsInformedML #PDE #MachineLearning #GenerativeAI #OpenSource #Research #Python #Pytorch #AIforScience #ComputationalPhysics

44

I’ve had the privilege of working with Professor Krishnamachari over the past few years, and I’m incredibly grateful for his mentorship and guidance throughout my research at USC. This project was a deep dive into how GPT-5 learns, fails, and improves through structured feedback. If you’re interested in LLMs, programming languages, or inference-time adaptation, I’d love for you to give it a read!

11

2 Comments

Pleased to share a trio of recent studies led by my PhD student Hyeonbin Moon, showcasing how physics-informed machine learning (PIML) enables fast and data-efficient inference of material properties. These works demonstrate the potential of PIML to transform material modeling—making it more scalable, generalizable, and robust, even under data scarcity. [1] Flexoelectricity meets AI We present a PINN-based framework for solving both forward and inverse flexoelectric problems, offering a unified, mesh-compatible approach for modeling high-order electromechanical systems. *arxiv: https://lnkd.in/gn3DR9us [2] Generalizable thermoelectric property inference with neural operators By combining PINNs with neural operators, we enable label-free and accurate identification of temperature-dependent thermoelectric properties from sparse data, supporting scalable and high-throughput materials screening. *arxiv: https://lnkd.in/gX9wbxMk [3] Discovering hyperelastic laws from minimal data This work introduces a PINN-based method for discovering hyperelastic constitutive models using only displacement and reaction force data from a single test—robust even in noisy, limited-data settings. *arxiv: https://lnkd.in/g7WdNiGE

57

2 Comments

I am excited to share that our paper, “Broken in Transit: Detecting Type Confusion in ROS 2 Deserialization via Fuzzing,” was accepted and presented at the IEEE/SICE International Symposium on System Integration (SII 2026) in Cancun, Mexico. This work provides the first empirical demonstration of the real-world impact of type confusion in ROS 2, identifying the serialization–deserialization boundary as a previously uncharacterized attack surface in modern robotic systems. Why this matters ROS 2 powers autonomous vehicles, drones, industrial robots, and research platforms. If the middleware misinterprets message types during deserialization, it can lead to denial of service, data corruption, and even unsafe actuation behavior. Key contributions: Type confusion in ROS 2 deserialization We show how malformed or adversarial message layouts can trigger type confusion during ROS 2 message reconstruction. ROS-aware fuzzing pipeline We built a fuzzing pipeline that directly targets the ROS 2 middleware stack, uncovering vulnerabilities that traditional fuzzers miss. Validated on real robotic platforms We confirmed the impact on physical robotic systems, demonstrating that these issues affect real deployed environments. 🔍 New vulnerability class identified We uncovered a previously undocumented class of deserialization vulnerabilities and followed responsible disclosure practices with the affected maintainers. Grateful for the collaboration of Jose Toribio and Jeremy Blackstone. 📄 Paper: https://lnkd.in/detsJ72H #RoboticsSecurity #ROS2 #CyberSecurity #SystemsSecurity #Fuzzing #SecurityResearch

41

6 Comments

The Spring issue of IEEE Solid-State Circuits Magazine is out! https://lnkd.in/gxbjNtKw This issue is dedicated to Integrated Circuits for AI and ML applications, and was guest edited by Marian Verhelst and Charlotte Frenkel. It features four excellent articles that highlight core challenges and methodologies in AI hardware acceleration (see https://lnkd.in/gayteYmV for more detail). The feature articles are: 1) "Design and Optimization of Efficient Digital Machine Learning Accelerators" by Wei Tang, Sung-Gun Cho, Jie-Fang Zhang, and Zhengya Zhang examines AI acceleration through optimized computational kernels, such as matrix–matrix multiplication, leveraging architectures like SIMD and systolic arrays (10.1109/MSSC.2025.3549361). 2) "SRAM- and eDRAM-Based Compute-in-Memory Designs, Accelerators, and Evaluation Frameworks," by Yifan He, Xiaofeng Hu, Hongyang Jia, and Jae-sun Seo addresses in-memory computing (IMC) as a means to reduce data movement overhead, covering both analog and digital IMC approaches (https://lnkd.in/gHiukWKx). 3) "Open Source Heterogeneous SoCs for Artificial Intelligence," by Francesco Conti, Angelo Garofalo, Davide Rossi, Giuseppe Tagliavini, and Luca Benini explores the integration of heterogeneous systems on chip, demonstrating how open source platforms such as PULP support scalable AI hardware development (https://lnkd.in/gAQgK-8d). 4) " Sparsity-Aware Hardware," by Man Shi, Adrian Kneip, Nicolas Chauvaux, Jiacong Sun, Charlotte Frenkel, and Marian Verhelst discusses sparsity-aware hardware design, illustrating how structured and unstructured sparsity can be leveraged for significant performance improvements using an adapted roofline model (https://lnkd.in/gv8eTyvd). The issue also features regular columns: 1. Behzad Razavi’s “The Analog Mind” column (https://lnkd.in/gxKyqZr7) continues the series on CMOS inverter applications 2. Chris Mangelsdorf’s “Shop Talk: What You Didn’t Learn in School” column picks up where the previous issue left off, continuing the discussion on Miller compensation in "Miller’s Wrong Half-Plane Zero" (https://lnkd.in/gngTrywV). This time, the focus is on the right half-plane zero and its potential to degrade stability if not properly addressed. 3. Farhana S.’s “Circuits From a Systems Perspective” traces the historical evolution of wireless charging in everyday devices (https://lnkd.in/gnVhqUyF). Lot's of interesting articles to read! Enjoy! #ieeesscc #analog #ieee #ai #ml #solidstatecircuits #hardwareacceleration #semiconductors #inmemorycomputing

260

3 Comments

Happy to see our Perspective now out in npj Robotics on a performance-driven approach for soft robotics. https://lnkd.in/e2466K2Q From our side, some of this thinking has been shaped by work on the Hopcopter—where compliance isn’t just for safety, but a way to store and redirect energy for impulsive motion. That experience really convinced me that soft elements can increase performance when thoughtfully integrated. Of course, this only works in synergy: compliant mechanisms paired with rigid structures (beautifully explored in Kaushik Jayaram’s work) and emerging high-performance soft actuators (Kevin Chen and others) are what make these hybrid systems compelling. If this resonates, we’re organizing a RoboSoft 2026 workshop on Fast and Impulsive Robots Enabled by Soft Actuators and Mechanisms (April 7, Kanazawa). Would love to continue the conversation there. https://lnkd.in/eD2QWJuE

68

Major technical education CS degree teaches you OS, ML, algorithms — but not how to actually use your tools. MIT's "Missing Semester" fills exactly that gap. It covers the stuff no one formally teaches you: → Shell & scripting → Vim & text editors → Git (beyond just git push) → SSH & remote machines → Debugging & profiling → Security & cryptography You'll spend thousands of hours with these tools in your career. Might as well learn them properly. Free. Online. MIT quality. missing.csail.mit.edu #ComputerScience #SoftwareEngineering #MIT #LearningAndDevelopment #CodingTools

1

Shocked by what AI coding agents can do now. I built an incompressible CFD solver with coupled level-set/VOF interface tracking and cut-cell embedded boundaries on top of the AMReX framework. In about two weeks. ~26K lines of C++, 12 modules, 467 unit tests, 7 validated benchmarks. And it works. I didn't read a single line of the code myself. I set up 10 agent sessions, kicked off the orchestration loop to auto-resubmit tasks, went to sleep, and checked the results a few hours later. Anyone who's touched Adaptive Mesh Refinement knows: this stuff normally takes years. Ghost-cell communication, flux refluxing, subcycling, PLIC reconstruction, level-set reinitialization across AMR boundaries... each one is months of work on its own for building and debugging. AMReX is a well-established framework, but even building on top of it is no small thing. The documentation is massive, the API runs deep (FillPatch, EBFactory, MLLinOp...), and you'd normally spend weeks just reading source code before writing a single line of your own. I built custom orchestration scripts and MCP interfaces that encode the physics, the numerical pitfalls, and the solver architecture. Then I set up a multi-agent team, a lead agent coordinating parallel sub-agents, each owning a module. Project management at machine speed. The part that surprised me most: the agents found actual bugs in AMReX itself, and built missing functionality from scratch to get thing move forward. All benchmarks validate against published results. Mass conservation within 2%. Correct vortex shedding. Physically accurate two-phase topological transitions. What an exciting time to be a builder! #CFD #AIAgents

187

33 Comments