About

Ramavarapu S. Sreenivas is a Professor in the Department of Industrial and Enterprise Systems Engineering at the University of Illinois Urbana-Champaign, where he also serves as Associate Head for Graduate Studies. He holds research appointments at the Coordinated Science Laboratory (CSL) and the Information Trust Institute (ITI), and is an affiliate of the Electrical and Computer Engineering Department. His academic background includes a B.Tech. in Electrical Engineering from the Indian Institute of Technology Madras, an M.S.E.E., and a Ph.D. in Electrical and Computer Engineering from Carnegie Mellon University. After completing his doctoral studies, he was a Post-Doctoral Fellow at Harvard University in Decision and Control. His research focuses on the analysis, control, and performance evaluation of Discrete-Event/Discrete-State (DEDS) systems, which are prominent in domains such as air-traffic control, automated manufacturing, computer networks, and complex organizational operations. He employs advanced methodologies including Computation, Coding, Machine Learning, and Information Theory to develop near-optimal supervisory policies for these systems. Currently, he oversees the Center for Autonomous Construction and Manufacturing at Scale (CACMS), established in 2023. Sreenivas has contributed significantly to education through teaching courses in controls, systems engineering, and financial engineering, and has been recognized multiple times for excellence in teaching, including the 2023 UIUC Campus Award for Excellence in Graduate and Professional Teaching.

Research topics

• Computer Science
• Artificial Intelligence
• Machine Learning
• Mathematics
• Theoretical computer science
• Medicine
• Algorithm
• Gerontology
• Human–computer interaction
• Discrete mathematics
• Psychology
• Operating system
• Multimedia
• Engineering
• Mathematical optimization
• Applied psychology
• Real-time computing
• Programming language

Selected publications

• Neuro-Symbolic Generation of Explanations for Robot Policies With Weighted Signal Temporal Logic

  IEEE Robotics and Automation Letters · 2026-02-09 · 1 citations

  articleOpen access

  Learning-based policies have demonstrated success in many robotic applications, but often lack explainability. We propose a neuro-symbolic explanation framework that generates a weighted signal temporal logic (wSTL) specification which describes a robot policy in a human-interpretable form. Existing methods typically produce explanations that are verbose and inconsistent, which hinders explainability, and are loose, which limits meaningful insights. We address these issues by introducing a simplification process consisting of predicate filtering, regularization, and iterative pruning. We also introduce three explainability metrics—conciseness, consistency, and strictness—to assess explanation quality beyond conventional classification accuracy. Our method—<sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TLNet</small>—is validated in three simulated robotic environments, where it outperforms baselines in generating concise, consistent, and strict wSTL explanations without sacrificing accuracy. This work bridges policy learning and explainability through formal methods, contributing to more transparent decision-making in robotics.

  PublisherDOI

• Safety Monitor for Off-Road Planning with Uncertainty Bounded Bekker Costs

  SAE technical papers on CD-ROM/SAE technical paper series · 2026-04-07

  article

  &lt;div class="section abstract"&gt;&lt;div class="htmlview paragraph"&gt;Reliable off-road autonomy requires operational constraints so that behavior stays predictable and safe when soil strength is uncertain. This paper presents a runtime assurance safety monitor that collaborates with any planner and uses a Bekker-based cost model with bounded uncertainty. The monitor builds an upper confidence traversal cost from a lightweight pressure sinkage model identified in field tests and checks each planned motion against two limits: maximum sinkage and rollover margin. If the risk of crossing either limit is too high, the monitor switches to a certified fallback that reduces vehicle speed, increases standoff from soft ground, or stops on firmer soil. This separation lets the planner focus on efficiency while the monitor keeps the vehicle within clear safety limits on board. Wheel geometry, wheel load estimate, and a soil raster serve as inputs, which tie safety directly to vehicle design and let the monitor set clear limits on speed, curvature, and stopping at run time. The method carries uncertainty analytically into the upper confidence cost and applies simple intervention rules. Tuning of the sinkage limit, rollover margin, and risk window trades efficiency for caution while keeping the monitor light enough for embedded processors. Results from a simulation environment spanning loam to sand include intervention rates, violation probability, and path efficiency relative to the nominal plan, and a benchtop static loading check provides initial empirical validation.&lt;/div&gt;&lt;/div&gt;

  PublisherDOI

• Neuro-Symbolic Generation of Explanations for Robot Policies With Weighted Signal Temporal Logic

  IEEE Robotics and Automation Letters · 2026-02-09

  articleOpen access

  Learning-based policies have demonstrated success in many robotic applications, but often lack explainability. We propose a neuro-symbolic explanation framework that generates a weighted signal temporal logic (wSTL) specification which describes a robot policy in a human-interpretable form. Existing methods typically produce explanations that are verbose and inconsistent, which hinders explainability, and are loose, which limits meaningful insights. We address these issues by introducing a simplification process consisting of predicate filtering, regularization, and iterative pruning. We also introduce three explainability metrics—conciseness, consistency, and strictness—to assess explanation quality beyond conventional classification accuracy. Our method—<sc xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">TLNet</small>—is validated in three simulated robotic environments, where it outperforms baselines in generating concise, consistent, and strict wSTL explanations without sacrificing accuracy. This work bridges policy learning and explainability through formal methods, contributing to more transparent decision-making in robotics.

  PublisherDOI

• LDLT L-Lipschitz Network Weight Parameterization Initialization

  arXiv (Cornell University) · 2026-01-13

  preprintOpen access

  We analyze initialization dynamics for LDLT-based $\mathcal{L}$-Lipschitz layers by deriving the exact marginal output variance when the underlying parameter matrix $W_0\in \mathbb{R}^{m\times n}$ is initialized with IID Gaussian entries $\mathcal{N}(0,σ^2)$. The Wishart distribution, $S=W_0W_0^\top\sim\mathcal{W}_m(n,σ^2 \boldsymbol{I}_m)$, used for computing the output marginal variance is derived in closed form using expectations of zonal polynomials via James' theorem and a Laplace-integral expansion of $(α\boldsymbol{I}_m+S)^{-1}$. We develop an Isserlis/Wick-based combinatorial expansion for $\operatorname{\mathbb{E}}\left[\operatorname{tr}(S^k)\right]$ and provide explicit truncated moments up to $k=10$, which yield accurate series approximations for small-to-moderate $σ^2$. Monte Carlo experiments confirm the theoretical estimates. Furthermore, empirical analysis was performed to quantify that, using current He or Kaiming initialization with scaling $1/\sqrt{n}$, the output variance is $0.41$, whereas the new parameterization with $10/ \sqrt{n}$ for $α=1$ results in an output variance of $0.9$. The findings clarify why deep $\mathcal{L}$-Lipschitz networks suffer rapid information loss at initialization and offer practical prescriptions for choosing initialization hyperparameters to mitigate this effect. However, using the Higgs boson classification dataset, a hyperparameter sweep over optimizers, initialization scale, and depth was conducted to validate the results on real-world data, showing that although the derivation ensures variance preservation, empirical results indicate He initialization still performs better.

  PublisherDOI

• A Shared Autonomy Architecture: Utilizing an Advanced Personalized Control System (APeCS) and Shared Autonomy Arbiter (SAB)

  SAE technical papers on CD-ROM/SAE technical paper series · 2026-04-07

  article

  &lt;div class="section abstract"&gt;&lt;div class="htmlview paragraph"&gt;The shared autonomy framework has become an option with great potential in the field of autonomous vehicles. Human and machine control decisions typically demonstrate strengths in different scenarios. As a result, the robustness of systems can be enhanced by the collaboration between humans and autonomy. A shared autonomy architecture that takes into account both human and environmental factors was proposed in this work. The authority distribution between the human operator and the autonomy algorithm was determined by the Shared Autonomy Arbiter (SAB). Designed with a two-tier structure, the SAB incorporated a policy-level decision module, as well as a numerical-level arbitration tuning module. A fuzzy inference system (FIS) was incorporated to enhance the noise tolerance of the policy selection module. Furthermore, the human factor was taken into account by applying a projection to the users’ control input. The human operator’s control decision was projected by the Adaptive Personalized Control System (APeCS) to accommodate the skill levels and habits of various users. By incorporating a broad set of factors, this framework is suitable for diverse applications that require robustness in complex environments. Two case studies were included in this work to demonstrate its effectiveness. The first presented a concept design illustrating the application of the proposed architecture on autonomous vehicles operating in varied environments. The second showed that the proposed architecture can serve as a robust testbed by taking advantage of the authority modulating mechanism. By connecting a system under assessment and an established autonomy algorithm to the SAB, the new system can be tested robustly and safely through the flexible authority distribution.&lt;/div&gt;&lt;/div&gt;

  PublisherDOI

• Backward Fuzzy Driving Control for 6×4 Off-Road Vehicles

  SAE technical papers on CD-ROM/SAE technical paper series · 2026-04-07

  article

  &lt;div class="section abstract"&gt;&lt;div class="htmlview paragraph"&gt;Off-road autonomous vehicle systems must be able to operate across unstructured and variable terrain while avoiding obstacles. This presents significant challenges in vehicle and control system design, especially for less conventional platforms such as 6×4 vehicles. While forward driving autonomy has developed and matured in recent years, effective reverse navigation remains an under-explored area of vehicle co-design. Reversing 6×4 vehicles have limited rear steering authority, an extended wheelbase, and asymmetric traction, which introduce complex dynamics into any control system that is used. To address this need, a robust and experimentally validated fuzzy logic control architecture for 6×4 reverse navigation was developed during the course of this project. This architecture incorporates both near-field and long-range path data with adaptive outputs controlling steering and velocity based on a rule base that covers the whole vehicle state space. This method has low computational cost and is robust to terrain changes, wheel slip, and actuator lag. To accomplish this, the controller coevolves with the vehicle design parameters, making this an effective co-design strategy. The vehicle design constraints are embedded into the controller through constraint-aware membership functions and rule tuning, reducing the need for terrain-specific calibration. The architecture is modular and scalable across numerous similar platforms, supporting rapid reconfiguration and vehicle design exploration for future autonomous off-road vehicles such as those used in expeditionary environments.&lt;/div&gt;&lt;/div&gt;

  PublisherDOI

• LDLT L-Lipschitz Network Weight Parameterization Initialization

  ArXiv.org · 2026-01-13

  articleOpen access

  We analyze initialization dynamics for LDLT-based $\mathcal{L}$-Lipschitz layers by deriving the exact marginal output variance when the underlying parameter matrix $W_0\in \mathbb{R}^{m\times n}$ is initialized with IID Gaussian entries $\mathcal{N}(0,σ^2)$. The Wishart distribution, $S=W_0W_0^\top\sim\mathcal{W}_m(n,σ^2 \boldsymbol{I}_m)$, used for computing the output marginal variance is derived in closed form using expectations of zonal polynomials via James' theorem and a Laplace-integral expansion of $(α\boldsymbol{I}_m+S)^{-1}$. We develop an Isserlis/Wick-based combinatorial expansion for $\operatorname{\mathbb{E}}\left[\operatorname{tr}(S^k)\right]$ and provide explicit truncated moments up to $k=10$, which yield accurate series approximations for small-to-moderate $σ^2$. Monte Carlo experiments confirm the theoretical estimates. Furthermore, empirical analysis was performed to quantify that, using current He or Kaiming initialization with scaling $1/\sqrt{n}$, the output variance is $0.41$, whereas the new parameterization with $10/ \sqrt{n}$ for $α=1$ results in an output variance of $0.9$. The findings clarify why deep $\mathcal{L}$-Lipschitz networks suffer rapid information loss at initialization and offer practical prescriptions for choosing initialization hyperparameters to mitigate this effect. However, using the Higgs boson classification dataset, a hyperparameter sweep over optimizers, initialization scale, and depth was conducted to validate the results on real-world data, showing that although the derivation ensures variance preservation, empirical results indicate He initialization still performs better.

  PublisherOA PDF

• Intelligent Farm Management Using Artificial Intelligence

  2025-01-22 · 1 citations

  book-chapter

  Agriculture, coupled with farming, has been an important source of food in our country. Crops have been identified as an important source for agriculture. Rice and wheat have been the most important crops for agriculture. There are various factors affecting the quality of crops, and some are also destroyed. To handle the various factors that influence crops in a negative way, we discuss farm management coupled with artificial intelligence (AI), which would be of help to farmers in various ways.

  PublisherDOI

• On the Enumeration of all Unique Paths of Recombining Trinomial Trees

  ArXiv.org · 2025-10-03

  preprintOpen access

  Recombining trinomial trees are a workhorse for modeling discrete-event systems in option pricing, logistics, and feedback control. Because each node stores a state-dependent quantity, a depth-$D$ tree naively yields $\mathcal{O}(3^{D})$ trajectories, making exhaustive enumeration infeasible. Under time-homogeneous dynamics, however, the graph exhibits two exploitable symmetries: (i) translational invariance of nodes and (ii) a canonical bijection between admissible paths and ordered tuples encoding weak compositions. Leveraging these, we introduce a mass-shifting enumeration algorithm that slides integer "masses" through a cardinality tuple to generate exactly one representative per path-equivalence class while implicitly counting the associated weak compositions. This trims the search space by an exponential factor, enabling markedly deeper trees -- and therefore tighter numerical approximations of the underlying evolution -- to be processed in practice. We further derive an upper bound on the combinatorial counting expression that induces a theoretical lower bound on the algorithmic cost of approximately $\mathcal{O}\bigl(D^{1/2}1.612^{D}\bigr)$. This correspondence permits direct benchmarking while empirical tests, whose pseudo-code we provide, corroborate the bound, showing only a small constant overhead and substantial speedups over classical breadth-first traversal. Finally, we highlight structural links between our algorithmic/combinatorial framework and Motzkin paths with Narayana-type refinements, suggesting refined enumerative formulas and new potential analytic tools for path-dependent functionals.

  PublisherOA PDFDOI

• LDLT $\mathcal{L}$-Lipschitz Network: Generalized Deep End-To-End Lipschitz Network Construction

  ArXiv.org · 2025-12-05

  preprintOpen access

  Deep residual networks (ResNets) have demonstrated outstanding success in computer vision tasks, attributed to their ability to maintain gradient flow through deep architectures. Simultaneously, controlling the Lipschitz constant in neural networks has emerged as an essential area of research to enhance adversarial robustness and network certifiability. This paper presents a rigorous approach to the general design of $\mathcal{L}$-Lipschitz deep residual networks using a Linear Matrix Inequality (LMI) framework. Initially, the ResNet architecture was reformulated as a cyclic tridiagonal LMI, and closed-form constraints on network parameters were derived to ensure $\mathcal{L}$-Lipschitz continuity; however, using a new $LDL^\top$ decomposition approach for certifying LMI feasibility, we extend the construction of $\mathcal{L}$-Lipchitz networks to any other nonlinear architecture. Our contributions include a provable parameterization methodology for constructing Lipschitz-constrained residual networks and other hierarchical architectures. Cholesky decomposition is also used for efficient parameterization. These findings enable robust network designs applicable to adversarial robustness, certified training, and control systems. The $LDL^\top$ formulation is shown to be a tight relaxation of the SDP-based network, maintaining full expressiveness and achieving 3\%-13\% accuracy gains over SLL Layers on 121 UCI data sets.

  PublisherOA PDFDOI

Frequent coauthors

• William R. Norris

  57 shared

• Ahmet Soylemezoglu

  United States Army Corps of Engineers

  50 shared

• Dustin Nottage

  50 shared

• Busi Campos

  Laboratoire d'Analyse et d'Architecture des Systèmes

  36 shared

• Luca Bernardinello

  36 shared

• Giorgio De Michelis

  36 shared

• Jan Van Leeuwen

  36 shared

• P Estrailler

  Cornell University

  36 shared

Education

• Ph.D., Industrial Engineering

  University of Illinois at Urbana-Champaign

  1990

• M.S., Industrial Engineering

  University of Illinois at Urbana-Champaign

  1986

• B.S., Mechanical Engineering

  Indian Institute of Technology Madras

  1982

Awards & honors

• 2023 UIUC Campus Award for Excellence in Graduate and Profes…