Rohit Sonker

Rohit Sonker

ML/Robotics Carnegie Mellon University

About me

I’m Rohit Sonker, a Research Assistant at Auton Lab at Carnegie Mellon University advised by Prof. Jeff Schneider, pursuing MS Robotics at CMU. I am broadly interesed in developing AI algorithms for applications in real world domains. Currently, I work on Offline Reinforcement Learning, Optimal Experiment design with Bayesian Optimization, and multi-modal foundational models combining text and time-series data, with application to the field of Nuclear Fusion.

Prior to this I have ~5 years of experience in industry. I led the Data science and Machine learning team at CS DISCO India Office working on Machine learning and Large Language Model applications to help attorneys in E-discovery, document review and building cases. Before that, at PwC, I led several projects on the development of AI based intelligent clinical trial design leveraging predictive models, simulations and LLMs.

I graduated with a combined bachelors and masters from IIT Kanpur, where I worked on offroad driving in varying terrain conditions with Prof. Ashish Dutta

Research Projects

AI for Nuclear Fusion

Bayesian Optimization for Experiment Design -

Worked on developing a multi-scale bayesian optimization algorithm to achieve different goals pertaining to plasma control in nuclear fusion. Our algorithm has been tested at DIII-D tokamak facility in San Diego to mitigate tearing mode instabilities and also at K-STAR tokamak in South Korea for betaN (normalized plasma pressure) maximization. I will be presenting this at ICML 2025. If you are visiting and would like to chat please reach out !

Multi-Modal Foundational Models for Nuclear Fusion -

I am actively working on multi-modal state transition model which combines text and trajectory/timeseries data. Our goal is to allow text conditioned timeseries generation and timeseries conditioned text generation.

Reinforcement Learning -

Designed pipeline for distributed training and hyperparameter tuning of Reinforcement Learning agents for closed loop control of fusion tokamak actuators. I work on developing controllers for plasma profile tracking (eg. Density, Rotation, Differential Rotation). We recently ran this algorithm on DIII-D facility, results and full paper currently under submission!

Adaptive Models for Dynamics Learning in Changing Scenarios

Worked on latent parameter recurrent state space models for learning dynamics (state-transistion probability functions) for various robotics tasks. These models used Kalman updates to account uncertainty in data, and latent variables to dynamically adapt to different regions of the environment. Some examples include - robot moving over varying terrain and manipulator picking different loads. Models achieved state of the art results in robotics tasks.

Terrain Height to Dynamics Learning

Worked on adding terrain height to learning dynamics models of a robot moving over various terrain conditions. Using this model, we also built a Model Predictive Controller (MPC) for trajectory tracking.