Humans learn new physical skills quickly: we reuse prior experience, infer structure from a small number of examples, try things, recover from mistakes, and improve through interaction. I am interested in how to build robot agents that learn in a similarly resourceful way.

I am a first-year Master's student in Robotics at Carnegie Mellon University, advised by Prof. Max Simchowitz and Prof. Aviral Kumar. My research is at the intersection of reinforcement learning, generative modeling, and representation learning. I am especially interested in how large pretrained robot policies and world models can become adaptive systems: not just models that imitate or predict, but agents that can plan, assign credit, correct their behavior, and improve after deployment.

The questions I currently care about are: What structure in pretrained robot policies makes adaptation possible? How much task information is really needed before autonomous learning can take over? How should value functions be designed for long-horizon, compositional behavior? And how can agents adapt when deployment exposes states, failures, and task variations that were not covered in their training data? I am drawn to problems where representation learning, generative modeling, reinforcement learning, and control cannot be separated cleanly, because real agents need all of them to work together.

Before CMU, I studied at IIT Kharagpur, where I built a broad foundation across mathematics, software, and robotics. At the Autonomous Ground Vehicle Research Group, I worked on SLAM and RL pipelines for F1TENTH autonomous cars, and later led the 25-member team as Executive Head. I also worked with Prof. Aritra Hazra on offline and meta-RL, and with Prof. Balaraman Ravindran at IIT Madras on sim-to-real transfer. Those years made me a full-stack builder: I learned to care not only about the algorithm, but also about the code, the sensor pipeline, the robot, the experiment, and whether the system actually works under pressure.

I have also been fortunate to work on research that connects modern learning systems to real embodied problems. At CMU's Robotics Institute, I worked with Yufei Wang, Prof. David Held, and Prof. Zackory Erickson on learning rewards from preferences for real-world assistive manipulation. I collaborated with Prof. Sherry Yang and Prof. Bo Dai on self-improving video generation for planning and control.

I am drawn to research where careful scientific experimentation changes what we believe about learning systems, and where those insights lead to better algorithms for agents that must act in the real world. Outside research, I spend time with films, tennis, debating, games, books, and Carnatic/Tamil music.