I develop planning and learning methods that help robots act effectively in long-horizon, persistent environments.
I am a Ph.D. candidate in Computer Science at George Mason University and a member of the Robotic Anticipatory Intelligence & Learning (RAIL) Lab, advised by Dr. Greg Stein. My research sits at the intersection of robotics, machine learning, and task and motion planning, with a focus on anticipatory decision-making for household and service robots.
More broadly, I am interested in enabling robots to reason not only about how to complete the task at hand, but also about how their current actions shape the cost and feasibility of future tasks in the same environment.
My work focuses on long-horizon robot autonomy in environments where actions have persistent side effects. In these settings, choices that appear locally efficient can make future tasks significantly harder. I study how robots can anticipate those downstream consequences and make better decisions in the present.
In particular, I work on anticipatory task and motion planning, learning-augmented planning, and lifelong robot autonomy. My recent research integrates learned anticipatory cost models into classical planning systems to guide search toward world states that improve performance across sequences of future tasks.
Selected recent updates.
Recent work in anticipatory planning, task and motion planning, and graph learning.
We present a learning-guided task and motion planning framework for persistent environments, enabling robots to account for the future consequences of their current actions and achieve more organized, preparation-aware behavior.
We introduce anticipatory task planning, a formulation that helps robots solve current tasks while reasoning about the expected cost of unseen future tasks.
We model floor plans as graphs and use graph neural networks to improve room classification performance by leveraging relational structure in building layouts.
Selected technical projects beyond published papers.
I built a system that injects learned anticipatory cost models into the Fast Downward planning stack, allowing neural cost estimation to be composed with standard symbolic heuristics such as FF, max, and goal-count.
In a 2D navigation setting, I formulated the domain in PDDL and used PDDLStream to plan obstacle relocations that open a feasible path to the goal.
I compared Proximal Policy Optimization (PPO) and Deep Q-Networks (DQN) on the Lunar Lander benchmark to study differences in learning stability, sample efficiency, and control performance.