Following policy guidelines provided by the University, we have cancelled Research Expo this year due to the potential transmission of the Coronavirus.
159. Sparse Learning-based Occupancy Mapping and Safe Navigation in Unknown Environments
Department: Electrical & Computer Engineering Research Institute: Contextual Robotics Institute Faculty Advisor(s): Nikolay Atanasov Primary Student: Thai Duong
Abstract Autonomous mapping and navigation in unknown environments is challenging, especially for large and complex environments. State of the art approaches in online mapping face two challenges: (1) maintaining and updating a memory-efficient representation for large environments and (2) quantifying uncertainty in the map representation. Our work proposes Gaussian Process (GP) regression to capture occupancy correlation among map elements in a fully Bayesian model. We exploit a set of latent points and properties of the kernel function to develop an efficient storage and update method for probabilistic mapping from streaming sensory data. Second, we employ a machine learning method based on the Kernel Perceptron to learn a sparse representation of the obstacle surfaces in the form of support vectors. This approach not only provides efficient map storage and maintenance but enables fast collision checking with theoretical guarantees. Given our sparse probabilistic occupancy map, a second major challenge is to achieve safe and stable navigation for a robot system. To address this control problem, we propose an adaptive control strategy based on a virtual first-order system, called a reference governor. The governor embeds directional preference to accommodate the geometry of local obstacles, allowing the robot to slow down when going through unsafe regions such as narrow passages and speed up otherwise. We show that our controller is able to navigate in complex environments much faster than common control designs while providing joint stability and collision avoidance guarantees.
Industry Application Area(s) Control Systems | Other: Intelligence Systems, Robotics