Learning Behavior Trees from Demonstration
2018.6 - 2018.9
Paper
Kevin French, Shiyu Wu, Tianyang Pan, Zheming Zhou, Odest Chadwicke Jenkins, “Learning Behavior Trees from Demonstration” IEEE International Conference on Robotics and Automation (ICRA) 2019
Goal
To allow the end user to teach new tasks to robots without expert knowledge, we propose a new learning from demonstration pipeline that incorporates Behavior Trees, a control architecture, as a new form of policy produced by learning from demonstration.
Contribution
Developed BT-Espresso algorithm, which converts the learned decision tree from demonstration into a Behavior Tree. Incorporated several boolean simplification algorithms into the system, including Espresso heuristic logic minimizer, tabular method, which efficiently prune the behavior tree without performance loss. Further improved the system using multi-threading and multi-process methods. Developed project GUI based on PyQt module.