Hi there! I'm a second year PhD student here at Berkeley, working on autonomous robotics and 3D mapping. I am advised by Prof. Claire Tomlin in the Hybrid Systems Lab. I am also a member of the Berkeley AI Research lab. I did my undergrad in electrical engineering at Princeton.
Outside of research, I like to play intramural ultimate frisbee and softball, and lately I've started getting back into acoustic guitar.
I am fundamentally interested in two related areas of robotics:
- Representing 3D space efficiently, and
- Understanding how robots should make good decisions under uncertainty.
Representing 3D space efficiently
Representing 3D space efficiently is important for a couple reasons. First, no matter what sensors a robot is equipped with, and no matter what it's objective is, it somehow needs to use what it senses to guide how it accomplishes the objective. Depending on the sensor and the objective, some space representations may be more useful than others. For example, if a robot just needs to ensure it will not collide with obstacles, a coarse volumetric representation like an occupancy grid is sufficient. However, for a robot that needs to render surfaces a mesh representation may be more appropriate.
My interest in this area stems mainly from a broader interest in robot autonomy; that is, robots making decisions about how to act. In order to do this well, a robot needs to understand and represent some salient aspects of the environment.
Making good decisions under uncertainty
This is obviously a critical task for modern robotics. My interest is primarily focused on how robots gather information about their environments; that is, how can a robot choose its next actions in order to gain the most new information about its environment. I am also very interested in broader problems related to distributed optimization and robust control.
Internet flow control is a project aimed at leveraging relatively simple linear time varying system modeling and a model predictive control scheme to improve TCP/Controlled Delay performance for bulk flow scenarios on the internet. This work is in collaboration with Margaret Chapman, Nathan Hanford, and Dipak Ghosal.
Information-thoretic bounds on optimal power flow seeks to understand the fundamental limitations of a regression-based inverter control scheme that sees only local, rather than global, information on a large power grid. Thiswork is a collaboration with Roel Dobbe.
Robust stochastic reinforcement learning is still in the planning stages, but the main idea is to estimate uncertainty in value function approximation and maximize not just expected reward but expected reward with some regularization for high uncertainty. I am working on a home-brewed framework for testing these ideas -- feel free to check it out.
AtomMap is a recent project related to 3D space representation. The main idea is to represent space using non-overlapping spheres, with the goal of capturing surface orientation more accurately than existing techniques allow. See link for more details.
Mininet is my current side project. I realized that I was going to lots of deep learning-related talks, and sitting next to a lot of folks who use neural nets all the time, and yet I'd never actually played with them myself. So I built mininet, which is a lightweight C++ library for deep learning. Currently, it only supports simple, feed-forward networks, but over time I do plan to add more functionality. Take a look if you're interested!
My GitHub site: click here.
My LinkedIn page: click here.