David B
- Research Program Mentor
PhD candidate at Harvard University
Expertise
robotic manipulation, control theory, reinforcement learning, online machine learning, humanoid robots
Bio
I am a passionate researcher that uses robotic hardware platforms (drones, arms) to test novel algorithms at the cutting edge of control theory and reinforcement learning. The robotics field is undergoing rapid progress --- it's an exciting time to be in this field. Control theory is a sub-field of applied mathematics that deals with the control dynamical systems. Modern control theory is rich with a wide variety of interesting control strategies, whose performance can be guaranteed mathematically. By combining the rigorous performance guarantees from control theory with the exceptional practicality of machine learning algorithms, autonomous hardware systems (a.k.a robots) can become safer, more trustworthy, and ultimately more beneficial to society. In terms of my personal interests and hobbies, I love being active. I've recently started learning how to row and I've also done a lot of distance running --- a great hobby for the mountains of Nevada where I'm from.Project ideas
Robotic Grasping Simulation using Python
In this project, the student will use Python to implement a simulation of an industrial robotic arm and end effector picking an object from a bin and placing it in another bin. Although this task may sound trivial, it will require skills in mathematical optimization, control theory, and rigid-body kinematics. The student will utilize Drake, a Python library for model-based design and verification of robotics, developed at MIT and the Toyota Research Institute. Depending on progress, this project can go as deep as desired --- for example, by adding algorithms for motion planning, pose estimation, and impedance control.
Fundamentals of Control Engineering
Control engineering is a discipline concerned with stabilizing and regulating dynamical systems. Modern control theory, a field of applied mathematics, provides the theoretical foundations for controlling "real" systems, like a rocket, car, motor, or robot, to name only a few. Mechanical, electrical, chemical, and computational systems require control engineering to properly operate. In this project, the student will use control theory to study a system of their choice --- for example a motor or DC circuit. Programming tools such as MATLAB and Python may be used. The final product will be a research paper or presentation.
How accurately can large language models predict time series data, such as financial markets and weather patterns?
Large language models (LLM) have recently been shown to perform well as predictors of time-series data. In machine learning (ML), time-series data is notoriously challenging to predict, due to ML's inherent ability to interpolate, but not extrapolate. Most people that use LLMs use them via applications such as ChatGPT. However, there are open source LLMs that can be downloaded and run programmatically. These "raw models" are well-suited for scientific study. For example, these raw models can be examined in Python, and the LLM's "next token prediction" can be exploited to predict numerical sequences, instead of sequences of words. In this project, the goal is for the student to find a "niche" or "angle" that is unique, such that the results could possibly be the first ever in its respective research topic. Overall, this project will enable the student to learn about time series models (in financial markets and weather patterns, for example), machine learning, LLMs, and Python programming.
How well can large language models predict the behavior of electrical and mechanical systems?
Large language models (LLMs) and generative pre-trained transformers (GPT) have demonstrated incredible capabilities, not only in their abilities with regards to natural language, but also in their capabilities to act as "zero-shot time series forecasters". Recent research papers have shown that LLMs, when given a sequence of discrete points from some signal (for example, a sine wave), can predict the future behavior of the signal with high accuracy. Given this surprisingly accurate results, I am curious on how well LLMs can perform when predicting the behavior of electrical and mechanical systems. LLMs are incredibly popular, yet so new that so many aspects of them are still unexplored. This project allows the student to discover new knowledge about LLMs. Fortunately, the project is niche, but also has a low barrier to entry, which makes it perfect for a high school student.