Time: 2PM - 3PM ET
Virtual Link: https://teams.microsoft.com/meet/2205202324560830?p=ujaza5yWTRsOGWnxsv
Marilyn Braojos
Robotics PhD Student
School of Aerospace Engineering
Georgia Institute of Technology
Committee members
Summary:
Spacecraft rely on state estimation robust to dynamic environments for increased autonomy. This proposal presents predictive modeling frameworks that increase spacecraft awareness and resilience. The first aim of this work addresses navigation challenges from extended periods of radio communication outages. Global LSTM-based architectures were benchmarked with ARIMA architectures to model GPS satellites’ clock bias for extended 6-hour horizons at < 2 ns accuracy. The second aim of this proposal transitions to mobile vehicle uncertainty awareness from proprioceptive sensor information. To mitigate the limitations of exteroceptive sensors, this aim demonstrates an uncertainty-aware terrain classification mechanism based on an evidential deep learning architecture. Extracting accurate uncertainty from out-of-distribution sensor data provides further insight into system faults and terrain hazards. The final aim of this proposed work integrates the uncertainty-aware mechanism into an adaptive extended Kalman filter to achieve real-time state estimation in the Husky A200. This capability will be demonstrated by dynamically scaling the EKF’s measurement noise covariance matrix in ROS Gazebo as the mobile robot traverses pre-defined trajectories with varying friction characteristics to represent different terrain types. Together, these aims deliver a comprehensive framework capable of safe and reliable state estimation resilient to dynamic environments and unexpected sensor faults.
]]>