Deep Learning on Point Clouds: From Paper Folding for Soft Robots to Unsupervised Robotic Mapping
Deep learning on point clouds, in addition to images, offers many new possibilities for robotics. In this talk, I will explain three recent projects in my group. The first one starts with a question: can deep networks learn paper-folding/zhezhi/origami, the art of turning a piece of paper into a 3D shape. I will explain how we design a 3D point cloud auto-encoder that essentially resembles the paper-folding operations in its decoder, leading to better reconstructions of 3D shapes and better linear separable latent features, yet being more parameter-efficient than its competitors. Then, I will show how this new decoder can be used for addressing a challenging robotics task: soft robot proprioception, achieving real-time (>400Hz) full body 3D shape reconstruction with high accuracy (<=1% relative error) from only self-observing images inside a soft robot's body. Lastly, I will discuss our CVPR'19 oral paper, DeepMapping, where we proposed a new unsupervised deep learning method to address the point cloud mapping (and SLAM) problem, which conventionally involves many complex and hand-engineered processes.
Dr. Chen Feng
Dr. Chen Feng earned his Bachelor's degree in geospatial engineering from Wuhan University in China. Then he went to the University of Michigan at Ann Arbor and earned a master's degree in electrical engineering and a Ph.D. in civil engineering in 2015, where he studied robotic vision and learning and attempted to apply them in civil engineering. After graduation, he became a research scientist in the computer vision group at the Mitsubishi Electric Research Labs (MERL), focusing on visual SLAM and deep learning. In 2018 August, he became an assistant professor jointly in the Department of Mechanical and Aerospace Engineering and the Department of Civil and Urban Engineering at New York University Tandon School of Engineering, where his lab AI4CE (A-I-force) aims to advance robotic vision and machine learning through multidisciplinary use-inspired research that originates from civil/mechanical engineering domains. More information can be found at https://ai4ce.github.io/.
Host: Professor Ziming Zhang