Ziang Xie Rose Hills
Learning Invariant Features f or Robotic Perception
A key challenge in bringing robots out of the manufacturing setting and into homes and offices is that of perception: though many robots are equipped with numerous sensors, there is currently no reliable computer algorithm which can take as input images of unpredictable, cluttered environments and identify each object in the image and estimate its 3D pose.Recently, there have been many advances in the field of unsupervised feature learning via neural networks, specifically in the learning of sparse features. Given the vast amount of data online, as well as the ability to render photorealistic images in simulation, there arises the possibility of enforcing feature invariances using structured data. This summer I plan to work on a project using such data to construct features which possess invariance to different lighting conditions and shifts in viewpoint, as well as encapsulate depth information through RGB-D sensors to improve robotic perception.