Policy-Driven Behavior Trees

Abstract

The research discussed in this paper aims to simplify the creation and modi?cation of adaptive behavior for a robot. The following research question was formulated. How can a framework for robots be created that simpli?es the online creation and modi?cation of adaptive behavior? To answer this question, two sub-questions answer how this can be technically realized and how this framework can be operationally embedded. The technical realization has been done by writing policies in Drools, the creation of behavior trees in a virtual environment created in Unity and their combination in a policy engine. TNO developed the Policy Engine TNO (PET) that allowed the policies and the behavior trees to communicate with each other, creating a policy-driven behavior tree. Both policies and behavior trees are intuitive to use and encourage non-experts to experiment with the creation and alteration of adaptive behavior strategies. The framework entails this combination and was speci?cally developed for a house search mission, simulated in a 3D virtual test environment. An ontology was written in the policy engine so that the robot is able to classify objects in the virtual environment in relation to other objects. The robot may recognize a weapon in the virtual environment and know that this is classi?ed as a dangerous object. With this ontology, several policies were created to function as a conflict-solving mechanism and allow the robot to follow these policies and overrule any previously issued policies. The scenario that followed from this house search mission was carefully created in collaboration with domain experts that are closely working together with TNO. With the creation of this scenario and a simulation run to prove that the integration of policies and a behavior tree works, the sub-question regarding operationally embedding was ful?lled.