A teacher agent generates training tasks by monitoring the robot's live performance, while grounding the curriculum in real-world task distributions.
- +6.8% success — wheeled navigation in constrained environments
- +6.1% success — quadruped locomotion in confined 3D spaces
A teacher re-weights auxiliary rewards as the robot's proficiency grows — training wheels that fade away on their own.
- +122.62% off-road mobility, 3× faster training
- 5/5 physical off-road trials vs 2/5 for expert-designed rewards
Plan with full robot dynamics near the robot, and progressively simpler dynamics along the horizon — full fidelity exactly where it matters.
- Improved performance across three augmented planners
- Won the simulation phase of the 2025 BARN Challenge
Surface normal consistency resolves the double-sided mapping problem for thin structures like walls and doors.
Who am I
Ph.D. student in Computer Science at George Mason University's RobotiXX Lab, advised by Dr. Xuesu Xiao. I study how robots can learn complex behaviors efficiently — by training on the right task, at the right difficulty, at the right time. Before this: M.S. at Carnegie Mellon, B.S. at Cincinnati, and a summer building statistical testing infrastructure at AWS.