Task and Motion Planning with Behavior Trees for Locomotion and Manipulation

Robust and flexible behavior generation for robot task execution remains a difficult problem due to the wide variety of constraints, disturbances, and uncertainties present in the environment. Behavior Trees (BTs) have emerged as a powerful Control Architecture for reliable and flexible autonomous action execution that is capable of authoring complicated logic. More specifically, BTs can execute high level actions based on a tree of composed primitive policies to accomplish a task. This thesis examines the design of two unified robot frameworks that integrate BTs as a robust middleware between high-level decision making as well as low-level motion planning and control. Firstly, a framework for legged locomotion with disturbance rejection is proposed, which incorporates model based trajectory optimization for motion primitive generation and a hybrid dynamic tracking controller for bipedal stability. Each node of the BT is integrated as a single walking step that can be sequentially combined for multi-step locomotion plans with robustness to disturbances. A high level decision maker, Linear Temporal Logic with reactive synthesis, is then used to automatically generate scalable actions with the understanding of potential disturbances. Secondly, a manipulation framework was designed that incorporates a variety of motion primitives with a supporting perception framework for solving generic manipulation problems. At the task planning level, the ROS2 PlanSys2 architecture was used to generate a plan based on BTs and solve multi-step plans. Assembly and disassembly tasks are specifically demonstrated using pick, place, alignment, locking, unlocking, insertion, removal, and many other action primitives for the maintenance of industrial and aerospace environments.

Date

2022-05-03

Resource Type

Text

Resource Subtype

Thesis

Full item page