My research aims to build a unified algorithmic framework where a robot efficiently infers the optimal value function from a bounded set of interactions with both humans and the environment. It ties together insights from motion planning and imitation learning and applies them to robots deployed in the wild.

**2017 - 2020**
| **Bayesian Motion Planning**
Motion planning algorithms, in the absence of any prior knowledge, search *everywhere* for a solution, considering solutions that are almost surely invalid. Can we reduce search time if we are in a Bayesian setting where we have priors on edge validities (e.g. from a database of plausible worlds)? In this work, we explore near Bayes-optimal search strategies for efficiently solving various motion planning problems. Papers: ICRA'20, IJCAI'18, NeurIPS'17, ISRR'17 |

**2017 - 2018**
| **Learning to Search**
How should search algorithms leverage past experience to speed up search? Search algorithms rely on heuristics to balance *exploration*, i.e., discovering promising new states, and *exploitation*, i.e., expanding the current best state. We show how such heuristics can be learned by offline imitation of optimal planners. Papers: RSS'19, CoRL'17, Videos: Oral |

**2019**
| **Learning to Sample**
How should planners efficiently search over a *continuous, high-dimensional* state space? Planners typically sample the space to create a roadmap. A desirable roadmap is one that is sparse, allowing for speedy search, with nodes spread out at various bottleneck regions along potential paths. We explore techniques for learning distributions that can be sampled to generate such roadmaps. Papers: IROS'19 |

**2015 - 2016**
| **Learning Planner Ensembles**
Can we train a learner to select the best planner for a given scenario? We argue that this learning problem is hard - small changes in obstacle configurations dramatically alter planner performance. Instead, we learn to hedge our bets and predict an *ensemble* of diverse planners that can be run in parallel such that at least one has good performance. Papers: ICRA'16, ICRA'15 |