Building Reactive Task Plans for Real-World Robot Applications - Chris Paxton (BT seminar #5)

ABSTRACT:
Robots are increasingly an important part of our world, from working in factories and hospitals to driving on city streets. As robots move into more unstructured environments such as homes, however, we must be able to create complex, reactive task plans that can deal with stochastic actions, unreliable sensors, and that above all are intuitive and easy to build. To this end, we created the Behavior Tree-based CoSTAR system -- which allows novice end users to create task plans for industrial robot task plans, shown in a 35-person user study to be highly user friendly and offer a useful set of tools for creating task plans. Some of this technology was later spun out into a startup -- Ready Robotics. We describe a variant on Behavior Trees which supports symbolic task planning and gives guarantees on performance. Finally, we describe a case study on an example of a reactive household manipulation task. Reactive task plans based on behavior trees allow us to build robust, responsive, and interpretable systems that can perform challenging multi-step tasks alongside humans.
BIOGRAPHY:
Chris Paxton is a research scientist at Meta AI, in FAIR Labs. He previously worked at NVIDIA research in their robotics lab. He got his PhD in Computer Science in 2019 from the Johns Hopkins University in Baltimore, Maryland, focusing on using learning to create powerful task and motion planning capabilities for robots operating in human environments. During his PhD, Chris led the team that won the 2016 KUKA Innovation Award with CoSTAR, a project which enabled for end-user instruction of collaborative robots. His goal is to tie together language, perception, and action, in order to make robots into robust, versatile assistants that can work alongside humans for a variety of applications. Recently, his work won the ICRA 2021 best human-robot interaction paper award, and was nominated for best systems paper at CoRL 2021.