The challenge of unstructured robot grasping from limited perceptual inputs is key to advancing generalized robot capabilities. I hope to integrate recent deep learning methodologies in 3D reconstruction and imitation learning to create an end-to-end policy for robotic manipulation, allowing robots to infer optimal grasps from a single image and a natural language query. I plan to explore point cloud methods that leverage off-the-shelf segmentation and single-shot 3D reconstruction methods to generate useful scene representations, where grasp sampling algorithms are used to deduce ideal scenarios. Our final intent is to produce a universal model that facilitates grasping in a wide range of environments, bypassing the challenges of specialized data collection and intricate camera setups.

I am participating in SuperUROP because I want to dive deeply into research and further my knowledge within artificial intelligence for robotic perception and control. I hope to take my in-class learning a step further through a meaningful and impactful project that can help advance the field as I better understand the world of AI research.