Behavior Trees

My research aims to address the question: ’How can the robot deliberation in dynamic and unpredictable environments benefit from the task representation language?’. The question arises from the fact that manual design and automated planning techniques represent the most used approaches in robots autonomy and they both rely on the chosen task representation language. My past research has helped to answer this question by investigating alternative task representation languages. In this setting, I identified Behavior Trees , widely adopted in the video game industry, as an innovatory task representation language for robot deliberation and I analyzed their main advantages. The main advantages of BTs lie in their modular structure and reactive execution that enable the designer to shape complex robotics behavior by aggregating simpler ones.

Check HERE a non-exausitive list of robotics companies using Behavior Trees:

Below, there are some selected works I did on the topic.

First book on Behavior Trees.

The book, titled Behavior Trees in Robotics and AI: An Introduction, is published by CRC Press Taylor & Francis. You can buy the hardcover copy and the ebook either on the CRC Press Store or Amazon.

Please make sure the pages are printed in color (see details here).

The free pre-print is available here.

Behavior Trees in Formal Verification

We introdices Behavior Trees in the RobMoSys framework and we are working towards the development of formal verification tools to assess the correctness of defined behaviors in the form of Behavior Trees.

More info here.

Blended Acting and Planning using Behavior Trees

We proposed a framework that automatically synthesizes a Behavior Tree that satisfies a given goal. It is an attempt to address the challenges regarding blending planning and acting: hierarchically organized deliberation and continual planning and deliberation.

Reinforcement Learning Using Behavior Trees

Taking advantage of modularity and reactiveness of Behavior Trees, we propose a model-free Automated Planned framework using Genetic Programming to derive an action plan for an autonomous agent to achieve a given goal in unknown environments. The advantages of the proposed approach are based on the advantages of BT over a common FSM. 

Stochastic Behavior Trees

Stochastic Behavior Trees are an extension of Behavior Trees presented at ICRA 2014 where each leaf node is described by its success/failure probabilities and execution times. The recursive structure of the tree then enables us to step by step propagate such probabilities and time from the leaves to the root.

Structural Properties of Behavior Trees

Using standard tools of robot control theory, it is possible to evaluate structural properties such as robustness and safety of an action plan described by a Behavior Tree. In IROS 2014 we presented how to build action plans that are safe and robust.