Planning to poke: Sampling-based planning with self-explored neural forward models
Workshop for Machine Learning in Robot Motion Planning at IROS 2018,
- Oct 2018
Associated documents :
This work presents a self-supervised robotic pusher setup that can acquire forward models for pushing tabletop objects by training neural networks. The resulting models are used for multi-step path planning. An explicit world representation allows generating collision-free and feasible object trajectories that avoid local minima. Through closed-loop replanning, the system can generate sequences of pokes that move objects to specified goal poses.
@Article{KGKWZ18,
author = {Kayser, Lars Henning and Görner, Michael and Kerzel, Matthias and Wermter, Stefan and Zhang, Jianwei},
title = {Planning to poke: Sampling-based planning with self-explored neural forward models},
booktitle = {}
journal = {Workshop for Machine Learning in Robot Motion Planning at IROS 2018},
editors = {}
number = {}
volume = {}
pages = {}
year = {2018},
month = {Oct},
publisher = {}
doi = {}
url = {https://personalrobotics.cs.washington.edu/workshops/mlmp2018/},
}