| |
This research is supported by Amazon.com.
This program aims to develop new fast approaches to extract 3D object pose estimation and depth estimation from 2D images. The resulting 3D object understanding should be sufficient to support motion planning decisions for robotic manipulation of items in the scene or mobile robots moving through the environment. We aim to develop a system that expects as input raw videos from multiple camera viewpoints. Our system aims to extract bounding boxes, estimate 3D pose, and estimate depth. We will characterize the robustness of our system across different types of objects, materials, environmental lighting conditions.. Our eventual goal is to enable robots in 3D environments with human co-habitants, who are also concurrently interacting with objects in the environment.