Shihao Li

I am a third-year Ph.D. student in Mechanical Engineering at The University of Texas at Austin, advised by Dr. Dongmei Chen in the Advanced Power Systems and Controls Lab.

My research focuses on robot learning, autonomous systems, motion planning, and safe decision-making, with a unifying theme of data attribution—using influence functions and gradient-based valuation to understand how training data shapes controller and planner behavior. I am currently exploring applications in robot learning and embodied AI, with hands-on experience deploying vision-language-action models. My earlier work established a foundation in data-driven control and manufacturing systems.

I received my B.S. in Mechanical Engineering from Pennsylvania State University in 2023, and my M.S. in Mechanical Engineering from UT Austin in 2025.

Email / Scholar / GitHub / LinkedIn

News

[Apr 2026] Paper on LLM-assisted multi-agent control accepted and published in Journal of Manufacturing Systems!
[Jan 2026] Paper on repetitive learning MPC accepted to ACC 2026. See you in New Orleans!
[Feb 2025] Two papers accepted to NAMRC 54!
[2025] Paper on MDR-DeePC accepted to MECC 2025.
[2025] Paper on Robust Optimal Task Planning accepted to MECC 2025.
[2025] Adhesion dynamics paper published in Manufacturing Letters.

Research

I work on robot learning, autonomous navigation, and safe decision-making. A recurring theme is data attribution: leveraging influence functions and gradient-based data valuation to curate training data, certify safety, and diagnose failures in learned controllers and planners. My methods span data-driven motion planning, curriculum reinforcement learning, and LLM-based agent frameworks. Earlier projects applied predictive control to manufacturing systems. Representative projects are listed below. † denotes first-author works.

	Gradient-Based Data Valuation Improves Curriculum Learning for Game-Theoretic Motion Planning ^† Shihao Li, Jiachen Li, Dongmei Chen arXiv preprint, 2026 page / arXiv / code Gradient-based data valuation (TracIn) outperforms metadata-based approaches for training game-theoretic motion planners. TracIn-weighted curriculum on GameFormer achieves mean planning ADE of 1.704 m vs. 1.822 m for interaction-difficulty curriculum (p=0.021) on nuPlan.
	Behavioral Score Diffusion: Model-Free Trajectory Planning via Kernel-Based Score Estimation from Data ^† Shihao Li, Jiachen Li, Jiamin Xu, Dongmei Chen arXiv preprint, 2026 page / arXiv / code A model-free trajectory planner that computes the diffusion score function directly from data via kernel-weighted estimation. Using triple-kernel weighting with a coarse-to-fine strategy, BSD achieves ~98.5% performance parity with model-based baselines using only 1,000 trajectories.
	From Natural Language to Certified H-infinity Controllers: Integrating LLM Agents with LMI-Based Synthesis ^† Shihao Li, Jiachen Li, Jiamin Xu, Dongmei Chen In Preparation, 2026 arXiv A multi-agent LLM framework that synthesizes certified H-infinity controllers via convex optimization, achieving 100% synthesis success on 14 COMPleib benchmarks.
	Data-Attributed Adaptive Control Barrier Functions: Safety-Certified Training Data Curation via Influence Analysis Jiachen Li, Shihao Li, Dongmei Chen arXiv preprint, 2026 arXiv Combines data attribution with adaptive control barrier function learning for autonomous navigation. Reduced prediction error by 35.6%, expanded the certified safe operating region by 39%, and achieved collision-free navigation through 16 obstacles.
	VLM-in-the-Loop: A Plug-In Quality Assurance Module for ECG Digitization Pipelines Jiachen Li, Shihao Li, Soovadeep Bakshi, Wei Li, Dongmei Chen arXiv preprint, 2026 arXiv A closed-loop quality assurance module that wraps ECG digitization systems with VLM feedback via tool grounding. Improved verdict consistency from 71% to 89%, recovered 41.2% of failed limb leads, and achieved 98.0% "Excellent quality" on 428 clinical HCM images.
	IF-CPS: Influence Functions for Cyber-Physical Systems — A Unified Framework for Diagnosis, Curation, and Safety Attribution Jiachen Li, Shihao Li, Soovadeep Bakshi, Jiamin Xu, Dongmei Chen arXiv preprint, 2026 arXiv Three CPS-adapted influence function variants—safety, trajectory, and propagated influence—that trace neural controller failures back to training data. Achieves AUROC 1.00 in Pendulum poisoning detection and 0.92 in HVAC.
	Stochastic Trajectory Influence Functions for LQR: Joint Sensitivity Through Dynamics and Noise Covariance Jiachen Li, Shihao Li, Soovadeep Bakshi, Jiamin Xu, Dongmei Chen arXiv preprint, 2026 arXiv A three-level influence hierarchy for stochastic LQR that jointly attributes closed-loop cost to training trajectories through both learned dynamics and noise covariance estimation via the Riccati equation.
	Auction-Based Task Allocation with Energy-Conscientious Trajectory Optimization for AMR Fleets Jiachen Li, Soovadeep Bakshi, Jian Chu, Shihao Li, Dongmei Chen arXiv preprint, 2026 arXiv A two-stage hierarchical framework for multi-robot task allocation and trajectory optimization. Sequential auctions with closed-form bids plus energy-aware optimal control achieve 11.8% average energy savings over nearest-task allocation across 505 scenarios.
	An LLM-Assisted Multi-Agent Control Framework for Roll-to-Roll Manufacturing Systems Jiachen Li, Shihao Li, Christopher Martin, Wei Li, Dongmei Chen Journal of Manufacturing Systems, 2026 A lifecycle multi-agent framework automating control engineering for R2R manufacturing—from system identification to sim-to-real adaptation with safety-guaranteed deployment. Achieved robust tension regulation under 50% model uncertainty.
	Curriculum-Based Soft Actor-Critic for Multi-Section R2R Tension Control ^† Shihao Li, Jiachen Li, Christopher Martin, Zijun Chen, Dongmei Chen, Wei Li Manufacturing Letters, 2026 · Accepted arXiv A curriculum-based Soft Actor-Critic for R2R tension control, trained in three phases with progressively wider reference ranges. A single policy handles 20–40 N step changes with no scenario-specific retuning on a three-section benchmark.
	A Repetitive Learning Model Predictive Control Method for Nonlinear Systems with Application to Roll-to-Roll Manufacturing Christopher Martin, Shihao Li, Jiachen Li, Wei Li, Dongmei Chen American Control Conference (ACC), 2026 · Accepted A control framework that gives machines "muscle memory." By updating internal linear models with data from previous cycles, RLMPC reduced tracking error by 64% in roll-to-roll manufacturing while maintaining real-time 3ms computation.
	Adaptive Trajectory Bundle Method for Roll-to-Roll Manufacturing Systems Jiachen Li, Shihao Li, Christopher Martin, Wei Li, Dongmei Chen IEEE Transactions on Mechatronics, Under Review arXiv An adaptive trajectory bundle method using derivative-free sequential convex programming for R2R manufacturing. Reduced tension error by 4.3% over gradient-based control and improved velocity response by 11.1%, with real-world validation on a six-zone system.
	Modeling of Adhesion Dynamics in Roll-to-Roll Lamination Processes ^† Shihao Li, Christopher Martin, Enrique Velasquez Morquecho, Zijun Chen, Dongmei Chen, Wei Li Manufacturing Letters, vol. 44, pp. 552–558, 2025 page A first-principles process model integrating Hertzian contact theory and transient heat transfer to predict adhesion quality. Identified a dual-heating strategy that reduces process time by 62.6%.
	DM-MPPI: Datamodel for Efficient and Safe Model Path Integral Control Jiachen Li, Shihao Li, Xu Duan, Dongmei Chen arXiv preprint, 2025 arXiv Extending datamodels from machine learning to Model Predictive Path Integral control. Achieved 5x sample reduction (2000 to 400 samples) while improving obstacle avoidance safety through learned influence prediction.
	AURORA: Autonomous Updating of ROM and Controller via Recursive Adaptation Jiachen Li, Shihao Li, Dongmei Chen In Preparation, 2025 An autonomous framework that maintains control of complex high-dimensional systems. AURORA continuously monitors performance and automatically rebuilds Reduced Order Models and retunes controllers when system dynamics drift, achieving 100% success rate on dynamic benchmarks.

Template adapted from Jon Barron. Last updated April 2026.