- Co-developed the Obstacles Out-of-Place Scoring (OOPS) framework to systematically evaluate the spatial reasoning, safety compliance, and contextual understanding of Vision-Language Models (VLMs) in navigation tasks.
- Curated a structured dataset of real-world scene anomalies to stress-test models including ChatGPT, Gemini, and InternVL.
- Performed rigorous error analysis to isolate systemic model failure modes, spatial inconsistencies, and overconfidence trends, formulating concrete optimization strategies for human-aligned, risk-aware AI evaluation.
Skills: Vision-Language Models (VLM), Spatial Reasoning, Error Analysis, Dataset Curation
- Trained robust locomotion policies for a Unitree Go2 quadruped within NVIDIA Isaac Lab, deploying scalable training pipelines across the NYU Greene High-Performance Computing (HPC) cluster.
- Designed multi-objective reward structures targeting precise foot placement, torque smoothness, slip reduction, and balance.
- Applied domain randomization over irregular terrains, velocity commands, and external forces to ensure broad policy generalization.
Skills: Reinforcement Learning, Isaac Lab, PyTorch, High-Performance Computing (HPC), Python
- Implemented and cross-evaluated Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC) reinforcement learning algorithms to balance a multi-link dynamic inverted pendulum system on a cart in NVIDIA Isaac Lab.
- Designed targeted reward formulations to maximize upright stability, eliminate transient oscillations, and penalize excessive control effort.
Skills: Reinforcement Learning, PPO, SAC, Isaac Lab, Reward Design
- Built a low-latency, vision-based hand gesture recognition pipeline processing live camera feeds to map human intent directly to physical robotic control coordinates.
- Utilized MediaPipe and OpenCV to isolate hand landmarks and classify spatial patterns reliably under varying illumination.
Skills: Computer Vision, MediaPipe, OpenCV, Python