Embedded Files
Started doing R&D through MIT's UROP as a way to earn money on the side while studying. UROP exposed me to writing out research projects, planning deliverables, and exposing myself to multiple fields. Once I considered plans after B. Sc. graduation, going back to grad school became a top choice.
Below are significant projects, including Master Thesis and Bachelor Thesis as well as summer and semesterly projects.
Links for further detail on my projects are a WIP!
A system for Autonomous Long Range Navigation in unmapped field environments combining geometric and visual features for efficient frontier exploration and expansion.
Using the DINOv3 foundation model, a Graph Convolutional Network, and a learned Q-function, we assign scores to frontiers in a constructed navigation graph to minimize time and cost to a goal far away, in unknown space.
Keywords: Robot Learning, Reinforcement Learning, Visual Foundation Models, Autonomous Navigation, Motion Planning, Field Robotics, Sensor Fusion, Legged Robots
Outcome:
ICRA 2026 Workshop on Field Robotics (Poster)
ICRA 2026 Workshop on Long-Term Deployments in the Wild (Poster)
M.Sc. Thesis conducted at NASA Jet Propulsion Laboratory, California Institute of Technology (2026)
This exploratory work uses an Uncertainty-Aware forward dynamics world model to guide a quadruped locomotion policy to learn on terrains with a high epistemic uncertainty, thus strengthening the predictive value of the model.
We show that when presented with out-of-distribution observations, the world model carries less autoregressive error for a 5 second horizon, and the epistemic uncertainty output also serves as an indicator of Out Of Distribution areas.
Keywords: Uncertainty Quantification, Epistemic Uncertainty, Legged Locomotion, Reinforcement Learning, Out of Distribution Deployments, World Models, Autoregressive World Models, Policy Optimization, Network Interpretability, Legged Robots
Semester Project conducted at the Robotic Systems Lab, ETH Zurich(2024)
This work adapts the Laplace Code (a local temporal difference code for distributional reinforcement learning) into a deep reinforcement learning setting, to recover temporal evolutions in the reward distribution. We scale the Laplace Code beyond tabular settings.
The work is tested on several Atari Games and dynamics environments, to simulate an evolving reward structure as the episode progresses. We use a Z-Transform and a multiplexer acting on 3 neural networks, each acting on a different reward discount factor scale. Results show similar performance to other distributional architectures. (Quantile and Categorical).
Keywords: Distributional Reinforcement Learning, Laplace Encoding, Reward Scales, Deep RL, Partial Observability, Reinforcement Learning Algorithms, Stochastic Methods
Semester Group Project conducted at ETH Zurich, for the Foundations of Reinforcement Learning Course (Optimization & Decision Intelligence Group, ETHZ)(2024)
We use an Event Camera with an IMU to perform proprioceptive state estimation at moments of impact (i.e. falls, jumps, crashes). Using an aggregate network that takes in Events as a primary visual source we show accurate state estimation at an overlooked time scale through learned methods.
This work has potential applications towards providing data and detail at impact times and high speed motion, while still providing useful state estimates at longer time scales.
Keywords: Sensor Fusion, Event Cameras, Neuromorphic Vision, Visual Learning, Robot Learning, Event-Based Vision, Legged Robots
B. Sc. Thesis / MIT SuperUROP project advised by Biomimetic Robotics Laboratory, MIT (2021-2022)
Various research projects as team lead of State Estimation & Tracking at MIT Driverless. Competed in the Indy Autonomous Challenge and Roborace. Deployed on a Dallara AV-21 chassis (autonomy-retrofitted Indy Lights car).
Four Projects:
Multi-Agent Prediction for Autonomous Motorsports
Decoupled Multi-Modal Perception for Autonomous Racing
Robust Modeling & Controls for Racing on the Edge
Fast & Modular Autonomy Software for Autonomous Racing Vehicles
Outcome:
Fast and Modular Autonomy Software for Autonomous Racing Vehicles - Field Robotics, 2024
ICRA 2022 Workshop on Opportunities and Challenges with Autonomous Racing
Indy Autonomous Challenge Competitions at Monza, Las Vegas (CES), Indianapolis Motor Speedway, Texas Motor Speedway, Goodwood Festival of Speed, WeatherTech Raceway Laguna Seca. (2022-2024)
Internal Project Presentations
Keywords: Field Robotics, Autonomous Racing, State Estimation, Adversarial Robotics, Multi-Agent Robots, Decentralized Robots, Game-Theoretic Motion Planning, Sensor Fusion, Visual Estimation, Opponent Pose Estimation
(2020-2024)
Creation of a human-like running gait on a bipedal robot using a direct transcription optimization problem. In contrast to other approaches, we do not use imitation learning or joint waypoint following.
We optimize over a stance phase and a flight phase for a 5DOF biped legged robot. By minimizing cost of transport through the sum of squared torques across the joints for a target velocity. We deploy an MPC controller to track the trajectory. We provide analysis on convex hull reachability and Lyapunov stability.
Keywords: Trajectory Optimization, Legged Robots, Legged Locomotion, Bipedal Robots, Humanoids, Model-Based Control, Optimal Control, MPC, Legged Robots
Semester Group Project for Underactuated Robotics Course (Robot Locomotion Group, MIT) (2022)
Throughout my time at ETH Zurich and MIT, I have taken the opportunity to work on other exploratory research projects! They span from learning a locomotion policy to reason about moving platforms, measuring impulse output from electrospray thrusters, a calibration technique for magnetometers, and many more! From these I mention technical learnings.
Keywords: Bio-Inspired Robotics, Legged Locomotion, Reinforcement Learning, Quadruped Robots, Machine Learning, Computer Vision, FPGAs, CubeSats, Electrospray Thrusters, Legged Robots
(2019-2024)
Page updated
Google Sites
Report abuse