Megan Minju Lee

RIO: Flexible Real-time Robot I/O for Cross-Embodiment Robot Learning

Preprint, Under Review

POLARIS: Hybridized classical planning and reinforcement learning framework

Hybridized classical planning and reinforcement learning framework for long-horizon robot manipulation. An LLM decomposes tasks into subgoals via PDDL predicates, then a policy assignment module dynamically selects between classical motion planning (RRT*) and learned RL policies (PPO) per subgoal, combining the reliability of classical methods with the adaptability of learned behaviors, with automatic fallback upon failure.

ASPEN: Spectral-Temporal Fusion for Cross-Subject Brain Decoding

Preprint, Under Review

ASPEN architecture improves cross-subject generalization in brain-computer interfaces by dynamically weighting more consistent spectral features against temporal signals, significantly reducing the need for individual user calibration.

VLM-Guided Subgoal Planning for Indoor Navigation in Habitat-Lab

Improve robot navigation efficiency and success rates in the Habitat simulator by integrating a "Think" action that allows PPO-based reinforcement learning policies to leverage Vision-Language Models for generating semantic subgoals.

CAPSTONE: Multi Agent Geometric Inspection and Classification

Bimanual manipulation system designed to perform active inspection through coordinated motion and 3D reconstruction.

Bumpkin: A Fist-Bumping Robot Companion

Robot arm that can identify when a human wishes to fist bump, track the closest fist, continuously tracks the target motion with the use of dynamic trajectory execution, and proceeds to bump when the fist is within a threshold distance from the tracked fist.

Nonlinear Trajectory Optimization for Multiple Robot Arms

Centralized multi-robot offline planner set up as a trajectory optimization problem. Given a set of kinematic collision-free paths for each robot, the controller tries to optimize a safe, smooth, and dynamically feasible local trajectory that can be directly executed on all robots online using DIRCOL and ALTRO.