Towards a behavior tree-based robotic software architecture with adjoint observation schemes for robotic software development

Nowadays, autonomous robots are increasingly accomplishing tasks in the dynamic world where environment states may change unexpectedly and be partially observable. The robot tasks in dynamic environments generally expect the robot to continuously deliberate upon the task goal while effectively obtaining environmental information with sensor and actuator actions. Implementing the underlying robotic software for such tasks can be rather difficult and tedious. The software developers need to synthetically implement the decision-making issues of controlling and planning, as well as the interactions between robotic sensing and actuating components, which is much more challenging than general-purpose software development. The existing software engineering practices focus on the general-purpose software development issues of modularity and communication, without specialized architectural solutions for the implementation of robotic controlling and decision-making processes, which still limits the implementation efficiency of robotic software in dynamic environments. This paper proposes a general-purpose scheme of adjoint observation between robotic sensing and actuating components, which specifies the integral control loop of controlling, planning, and data flows. The adjoint observation scheme solves the problem of effectively exploring the environment for effective observations by the integral control loop. Then we utilize the Behavior Tree component software architecture for concrete implementation of adjoint observation schemes. More specifically, we propose the Parallel and Fallback tree structure for concrete implementation of adjoint control flows. We also extend the BT architecture with an online planning component and mutual data store mechanism, enabling continuous planning and efficient data communication between robotic sensing and actuating processes. In the experiment, we select the Classical BT approach and Pure ROS-based approach as baseline approaches, to validate the task effectiveness of the adjoint observation scheme and development efficiency of the supporting software architecture.