We propose a novel method for learning a symbolic representation that is appropriate for the high-level planning of loco-manipulation tasks. This method is necessary for constructing a goal-directed sequence of motor skills of various levels of flexibility. Previous work has addressed this issue in the context of blackbox and parameterized motor skills. We extend this work to include a more flexible form of motor skill (namely dynamic option), of which the behavior is dictated by a low-level planner. The execution of the dynamic option aims to facilitate the execution of actions for multiple steps in the future. Particularly in loco-manipulation tasks, the dynamic option is suitable for modeling the locomotion action, which is executed in order to pick and place several objects in reachable range of the arm. The introduction of dynamic option also invokes a new systematic approach for learning symbols, and a new framework that uses a low-level motion planner to inform high-level planning. Unlike previous work, our proposed approach is able to learn symbolic representation, from the data collected from human user playing a video game that simulates loco-manipulation tasks. The features acquired via task abstraction reveal additional useful information about the underlying loco-manipulation coordination strategy, compared to the human performance features (e.g., task completion time and accuracy) and the task plan pattern.

Demo

Publication

• Heramb Nemlekar, Max Merlin, Zhaoyuan Ma and Zhi Li, “Learning Symbolic Representations for Loco-Manipulation Planning”, submitted to 2019 Robotics: Science and System (RSS).