Sai Shankar Narasimhan

I am a fifth year graduate student in the University of Texas at Austin, pursuing PhD in ECE. I am advised by Prof. Sandeep Chinchali. My research focuses on synthetic time-series data generation using diffusion-based multimodal models, and on improving forecasting by incorporating exogenous signals such as news and social media. Previously, I was a Research Assistant at Robotics Research Centre, IIIT, Hyderabad, working under Madhava Krishna.

I received my undergraduate degree in Electrical Engineering at Anna University in India, with my Bachelor's thesis advised by Ranganath Muthu. My Research Internship at IIT Madras, with P.V.Manivannan, was focussed on SLAM. I have also worked full-time for over a year at Swaayatt Robots.

Email  |  CV  |  LinkedIn  |  Github  |  Twitter  |  Google Scholar

• Nov 2025: I have successfully completed my PhD Progress Review!
• Sep 2025: I will be presenting Constrained Posterior Sampling for Time-Series Data Generation at NeurIPS 2025
• May 2025: I will be interning at Synthefy on developing multivariate time series forecasting foundation models
• May 2024: I will be presenting Time Weaver (Spotlight Talk) at ICML 2024
• Nov 2023: I will be presenting my work on Safe Networked Control as an Oral talk at ICRA 2024
• Sep 2023: Started my PhD at UT Austin under the supervision of Prof. Sandeep Chinchali
• Sep 2021: I joined UT Austin as an MS student in ECE
• June 2020: AutoLay, our benchmark for amodal layout estimation was accepted for IROS 2020
• Jan 2020: Our work, Deep Flow guided IBVS will be presented in ICRA 2020
• Dec 2019: Our work MonoLayout was accepted for WACV 2020
• June 2019: Joined Robotics Research Centre at IIIT-Hyderabad as Research Assistant

We propose Constrained Posterior Sampling (CPS), a principled framework for generating realistic synthetic time series data that provably satisfies user-specified hard constraints such as physics laws and operational limits.

We present Time Weaver, a novel diffusion-based model for conditional time series generation that can generate realistic synthetic time series data conditioned on scenario-specific descriptions and metadata.

We propose a principled framework for safe networked control of robotic systems using probabilistic verification to provide safety guarantees under network impairments such as delays and packet losses.

We introduce AutoLay, a new dataset for amodal layout estimation in bird’s eye view. Further, we propose VideoLayout, a real-time neural net architecture that leverages temporal information from monocular video, to produce more accurate and consistent layouts.

We present MonoLayout, a deep neural network for real-time amodal scene layout estimation from a single image. We leverage adversarial feature learning to hallucinate plausible completions for occlusions.

We propose a novel system consisting of deep neural networks that systematically integrates depth cues with flow features.

We propose a novel Kalman Filter (KF) algorithm, that leverages the statistical correlation between the measured variables.

We propose a solution to reduce the range and bearing error in SONARs significantly. Using the results from the Gaussian Correlation Inequality, we derive probabilistic transformations that can improve the measurements of the SONAR, thus reducing the sensor error.

We developed a collaborative Aerial Vehicle (UAV), Ground Vehicle (UGV) platform that can be used to aid / automate search and rescue missions in disaster zones.