Abstract:
Motion planning is a fundamental problem in robotics; how can we move robots efficiently and safely? Motion planning can be solved using several paradigms with their own strengths and weaknesses. This talk dives into Heuristic Graph Search and its application to motion planning by converting it to a problem of finding a start-goal path in a graph. Heuristic Graph Search employs cost-to-goal estimates (heuristics) to find paths efficiently and can find solutions with solution quality bounds (e.g. finding optimal solutions). The first part of this talk motivates and describes Heuristic Graph Search in a beginner friendly manner. Non-motion planning people are encouraged to attend! The second part of the talk discusses a novel method to use machine learning to learn local heuristics which can significantly speed-up heuristic graph search. Unlike prior works that attempt to estimate the entire cost-to-go, which is hard to estimate and generalize to new maps, we define local heuristics which estimate the additional cost-to-go associated with escaping local regions. We show this is easier to learn and generalize to new maps and can provide large speed-ups in kinodynamic planning for navigation.

Committee Members:
Maxim Likhachev (Advisor)
Zeynep Temel
Sebastian Scherer
Shivam Vats (PhD Student)