Ego-Pi: Egocentric Human Data for Humanoid Manipulation
Ego-Pi explores fine-tuning large Vision-Language-Action (VLA) models for dexterous humanoid manipulation using aligned egocentric human demonstrations and robot data. The project focuses on cross-embodiment generalization, bimanual coordination, and scalable data curation for real-world humanoid learning.
GRaD-Nav-Onboard: Vision-Based Drone Navigation
This project studies onboard, vision-only navigation for drones in dynamic environments. We integrate Gaussian Splatting, velocity estimation networks, and reinforcement learning to enable safe, real-time navigation with dynamic obstacles on embedded platforms.
Safety-Critical Control for Quadruped Robots
This line of work focuses on safety-critical control for quadruped robots using Model Predictive Control (MPC) combined with Control Barrier Functions (CBFs). The approach ensures safe navigation and motion generation under physical and environmental constraints.