Manipulation Planning From Demonstration Via Goal-Conditioned Prior Action Primitive Decomposition and Alignment

• Published in: IEEE robotics and automation letters 2022-04, Vol.7 (2), p.1387-1394
• Main Authors: Lin, Nan, Li, Yuxuan, Tang, Keke, Zhu, Yujun, Zhang, Xiayu, Wang, Ruolin, Ji, Jianmin, Chen, Xiaoping, Zhang, Xinming
• Format: Article
• Language: English
• Subjects: Decomposition; Inverse dynamics; Learning from demonstraion; manipulation planning; Planning; Reinforcement learning; Robotics; Robots; Structural hierarchy; Supervised learning; Task analysis; task planning; Training; Trajectory
• ISSN: 2377-3766; 2377-3766
• DOI: 10.1109/LRA.2021.3140127

Manipulation plays a vital role in robotics but is left unsolved. Recent work attempts to leverage the hierarchical structure of tasks via using action primitives. However, due to trajectory distribution shift, prior action primitives could hardly be adapted to new tasks. In this letter, we propose the layered action primitive planning from demonstration framework (LAPPLAND) to better utilize prior action primitives while maintaining behavior-interpretability. First, we pretrain goal-conditioned action primitives decoupled with a meta policy and an inverse dynamics model to facilitate interpretable goal state and reachable trajectory. Second, we decompose tasks with logical sub-structure into a sequence of prior action primitives and then align them for better adaption. Third, we execute the action primitives in sequence, conditioned on explicitly assigned goals to lead to the desired states. Extensive experiments in both simulated and real-world environments validate that robotic manipulation planning using LAPPLAND achieves a high success rate and is robust to the variation of the environment. We also compare LAPPLAND with three state-of-the-art methods to demonstrate its superiority.