Machine Learning for Motion Planning (MLMP) @ ICRA2021

Deep Neural Network-based Fast Motion Planning Framework for Quadrupedal Robot*

Deep Neural Network-based Fast Motion Planning Framework for Quadrupedal Robot*

by Jinhyeok Jang, Heechan Shin, Minsung Yoon, Seungwoo Hong, Hae-Won Park, and Sung-Eui Yoon

Korea Advanced Institute of Science and Technology (KAIST)






Abstract

We present a motion planning framework that generates the motion of a quadrupedal robot in a short time using a deep neural network. Our planner gets the initial robot state, target goal pose, and terrain heightmap as input and generates a trajectory of a quadrupedal robot. The planner contains deep neural networks that extract features from input. These features guide the planner to generate a precise trajectory. We achieved the planning time within 230ms for 2 seconds long trajectory over various terrain types.



Contents

Paper: PDF (5.37MB)