Robotics, Neural Networks, and Deep Learning

GRAND Seminar Friday, April 12, 12noon, Room: ENGR 2901

Professor SeYoung Oh

  • Affiliate Professor, Dept. of ECE, Univ. of Maryland
  • Professor Emeritus, Dept of EE, Pohang University of Science and Technology, S. Korea


Robot is a Moving computer then Computer becomes a Robot Brain. But moving safely around the world is a GRAND Challenge such that the Robot’s level of intelligence (IQ) will determine its chance of survival in the world. I will first start with the various robot navigation tasks and discuss the required intelligence. Navigation requires Environment Perception and Path Planning. The traditional approach to realize Robot (Artificial) Intelligence requires a well-defined model of the Environment as well as that of the Robot. The Analytical Approach using Probabilistic techniques like Kalman and Particle Filters still dominates with no true learning capability assuming all the models are given. Then we point out how humans solve the same problem by learning with neural networks (NN) in the brain. We then show what learning and generalization can bring us for Robotics. NN used in the early days for modeling of robotics and control did not require many hidden layers due to the low dimensionality of data – state variables and low D images. So, where does deep learning (DL) come in?

Most of the successful DL applications occurred in Computer vision and Speech but not so much in Robotics. I believe all these DL success stories can equally apply to the field of Human Robot Interaction (HRI) and Environment perception component of Robot Navigation. However, DL should honor the robotic constraints such as real time control and dynamic environments. In object detection, Robot vision presents more challenge than image classification in the sense that it needs to deal with images with different viewpoints, occlusion and scene clutter thus making inadequate most of the ImageNet trained from internet images. Robot cars will serve as the main target application to which DL is being applied.

Short Bio:

Prof. SeYoung Oh is a professor at Postech (Pohang Univ of Science and Technology) for 30 years and is now visiting Univ of Maryland, College Park. He has spent 40 years in robotics and neural networks and AI. His areas of past publication includes:

Key Applications:

  • Intelligent Vehicles -Autonomous Driving, and Advanced Driver Assistance Systems (ADAS)
  • Mobile and Service Robots also including Cleaning Robots (Low grade sensor based SLAM)
  • Human Robot Interaction - Face and Gesture Recognition
  • Robotic Unmanned Aerial Vehicles (UAVs) and Drones

Key Enabling Technology:

Machine Intelligence and Machine Learning

  • Neural Networks including Deep Learning Neural Network (DNN)
  • Nature-Inspired Optimization :
    • Swarm Intelligence – Particle Swarm Optimization (PSO)
    • Evolutionary Optimization – Evolutionary Programming (EP)

Mobile Robot Navigation

  • Localization, Mapping, and Simultaneous Localization and Mapping (SLAM).
  • Path Planning