PhD Candidate, Mechanical Engineering · Virginia Tech
Slides from recent talks. Click a slide to step through it.
Unifying Physical Demonstrations and Language Corrections for Neuro-Symbolic Robots
Robots learn skills from both physical demonstrations and language corrections. Providing demonstrations before corrections lifts zero-shot success from 45% to 72%.
From Local Corrections to Generalized Skills: Improving Neuro-Symbolic Policies with MEMO
Aggregates local language corrections into a skillbook of generalized skills, reaching 88% success on a real robot with about 1.5 corrections per trial.
Safe Interaction via Monte Carlo Linear-Quadratic Games
A stochastic local search for real-time, adjustable safety. In a user study, MCLQ cut collisions without slowing the robot's task.
Fine-Tuning Robot Policies While Maintaining User Privacy
A model-agnostic framework that uses keys to transform a policy's intermediate features, so a single network personalizes to multiple users without leaking information between them.