Hybrid Model-based Adaptive Systems :

NASA’s Mars Exploration Rovers, Spirit and Opportunity, opened the door for a new era of science based on insitu science instruments, through extensive traverses and an unprecedented level of science return.  However, at times these rovers needed to tip toe at the edges of craters, while operators worried whether or not the rovers would be able to return from exploring a particularly interesting crater.  Capable autonomous vehicles must be much more agile.  Rovers should be able to estimate soil traction and reorient their center of gravity in order to safely move along steep inclines.  To reach highly inaccessible locations, future legged robots must be able to catch themselves as they stumble.  Likewise, exploration vehicles and automobiles should be able to detect and diagnose failures during their onset, when the earliest symptoms may be hidden within the noise.

To support these types of capabilities, our research has focused on model-based adaptive systems. First, a method called Gaussian Particle Filtering was developed for robustly estimating and tracking the continuous state of systems under different operational regimes and under failure. This method was applied to tracking the behavior of a ball throwing, acrobatic robot and for detecting the subtle failures of an advanced life support system at Johnson Space Center. Second, a method called conflict-directed Branch and Bound was developed for efficiently solving optimization problems that mix logical decisions with continuous decisions involving linear constraints. This method was demonstrated on cooperative air vehicle path planning, based on simulations provided by DARPA. Our future research will explore methods for learning hybrid discrete/continuous models through active exploration, mission-directed cooperative path planners, and hybrid methods for controlling and interpreting humanoid motion.

Selected Publications in this area: