Research Experience:

1. Fine-tuning Large Language Models (LLMs) with LoRA to generate 6-DoF waypoints from structured prompts (Task/Rules/Env + expert demo), using chain-of-thought guidance and post-generation verification (collision, bounds, step-size, etc.) for safe object navigation in cluttered door layouts.

2. Training loop that logs the first failure point when the LLM’s path collides or violates limits, then feeds it back as a negative example in the next epoch to improve long-horizon plans and orientation choices.

3. Designed a disturbance-aware MPC + pose-optimization controller for human–robot co-transportation with a two-step iterative process to improve safety, tracking, and robustness vs. non-disturbance-aware baselines. Implemented the developed algorithm on a Fetch mobile manipulator.

4. Built a mutual-adaptation framework that models a distribution over human choices (risk + distance), adds an individual's time-varying stubbornness, and enables mode transitions (robot leads <=> follows) to maximize team performance, validated with real human feedback data and simulations.

5. Introduced pose selection under human-led to mitigate preference uncertainty, which empirically reduced system cost across diverse environments compared to no pose optimization.

6. Engineered whole-body control for a redundant 7-DOF arm + mobile base with singularity-aware pose candidates and derived Jacobians and linearized dynamics to form MPC input matrices per candidate pose.

7. Developed a Conditional Variational Autoencoder to generate feasible joint-pose candidates and integrated fully parallelized candidate evaluation to keep planning in real-time.

8. Added sequence modeling for human motion with LSTM-based prediction to initialize MPC when future human poses are unknown, and executed receding-horizon control with first-input application for stability.

9. Conducted extensive simulations and hardware trials demonstrating lower instantaneous and accumulated cost, tighter tracking, and milder compute scaling vs. non-pose-optimized or non-disturbance-aware baselines. The computation scales mildly with horizon due to parallelization.

10. Deployed the full stack in ROS/Gazebo and on Fetch hardware, and implemented real-time pipelines publishing base twists and 7-joint velocities from MPC controllers.