American Control Conference(ACC), 2026, [paper], [code]
This project models perception uncertainty for safe navigation by representing uncertainty as multiple Gaussians, the approach captures fine-grained structures that baseline method overlook. This enables more effective motion planning while maintaining probabilistic safety guarantees.
Project available on GitHub, 2025, [code]
This is a team project of Deep Learning for Robotic Manipulation course at Columbia. It learns an RL policy for active 3D shape exploration using tactile sensing, enabling robots to perceive objects that are visually occluded. The system is built and evaluated in IsaacSim/IsaacLab with diverse geometries, demonstrating effective online tactile-based reconstruction of unseen objects.
Project available on GitHub, 2024, [code]
This is a team project from the Deep Learning for Computer Vision course at Columbia. We improved a SOTA text-guided shape editing framework ChangeIt3D by leveraging a region-attended editor that leverages shape correspondence to enable more precise, localized geometry modifications.