Extending Behavior Trees for Representing and Planning Robot Adjoint Actions in Partially Observable Environments

Abstract

Behavior Trees (BTs) have received increasing popularity in the robotics community, serving as an efficient way of modeling robot behavior and structuring behavior interactions. Utilizing automated planning with BT task-level robot control has been proven efficiently reactive to environment dynamics, allowing iterative expansion of BT action nodes while executing the tree. However, existing BT-based planning approaches plan the new actions under the assumption of a fully observable environment model, which shows limitations when the robot has only partial observations on the environment states. In a partially observable domain, an expected robotic plan is an interleaved and adjoint sequence of robot acting and sensing actions. Each acting action is decided upon new observations received by a sensing action. It remains a challenge to efficiently obtain the partially observable states by interacting between robot sensing and acting actions. This paper first proposes a novel Adjoint Sensing and Acting (ASA) model that explicitly abstracts robot actions and specifies an adjoint interaction scheme between them to tackle this challenge. Then, we extend the classical formulation of BTs to represent and realize the ASA model, which proposes a new atomic ASA-BT structure for representing the partially observable environment states and modeling ASA action components. Finally, we cast the problem of planning adjoint acting and sensing actions as a Partially Observable Markov Decision Process (POMDP), which plans out an efficient ASA-based plan. By comparing with the classical BT approach, we experimentally demonstrate that plans from ASA-BT based planning approach can efficiently obtain the partially observable states with fewer time costs.