• The lack of general and systematic treatment of uncertainties and imperfect robot control, noisy sensors, and environmental changes—is a major barrier to robust robot operation and has thwarted the widespread use of robots in natural human environments until the recent decade. The partially observable Markov decision process (POMDP) provides a principled general framework for planning under uncertainty, using a probabilistic approach. Unfortunately solving POMDPs exactly is provably intractable. Early algorithms can handle only very small POMDPs with 10–20 states, which are woefully inadequate for modeling realistic robotic tasks. The challenge for POMDP planning is to scale up and handle POMDPs with very large state spaces and complex dynamics. Our research on POMDPs spans the entire spectrum of mathematical foundations, efficient algorithms, and application to critical real-world systems: The notion of covering numbers provides theoretical justification that explains why probabilistic approximation algorithms work well empirically on the computationally intractable problem of solving POMDPs. SARSOP and DESPOT are among the fastest POMDP algorithms available today for offline and online POMDP planning, respectively. Our algorithms have found a broad range of applications in robotics and beyond, including collision avoidance protocols for unmanned aircrafts, autonomous vehicles, penetration testing for computer security, and intelligent computer tutoring system.
    Read More
  • This work focuses on designing an active learning algorithm that selects not just the most informative sampling locations to be observed but also the types of measurements (i.e., target and/or auxiliary) at each selected location for minimizing the predictive uncertainty of unobserved areas of a target phenomenon given a sampling budget ...
    Read More
  • The project is funded by ASTAR and SERI to design and develop a cloud platform that screens for both retinal vessel health assessment and diabetic retinopathy analysis. We utilize techniques such as image registrations, deep learning techniques, and image quality filters to achieve high sensitivity and specificity.
    Read More
  • Geometry and motion are ubiquitous in the physical world, at both macroscopic and microscopic levels. To operate in this world or to simulate it, we need accurate models of the environment, compact data structures to represent the models, and efficient algorithms to compute motion for physical or simulated objects ...  
    Read More
  • Led by the Medical Computing Laboratory at the School of Computing, National University of Singapore, and partners, our project aims to develop an open, adaptive trial workbench for cost effective decision support in healthcare. Our workbench architecture is based on a set of generic information management and ...
    Read More
  • We study recursively enumerable graphs, their properties, and their dependence on the equality relation...
    Read More