Teacher-Guided Terrain-Aware Learning for Recovery of Quadruped Robots

Boyuan Deng*, Xu Yang, Yilin Mo, Nikolaos Tsagarakis
Istituto Italiano di Tecnologia, Tsinghua University

Abstract

We present an optimal posture formulation and posture evaluation machanism for robotic recovery on uneven terrain and a novel teacher-student reinforcement learning framework that enables a student policy to infer external terrain conditions using only proprioceptive sensing.

Framework and Fall Simulation


Diverse Terrains Depolyment

We show that our policy performs on the non-flat environments, also we provide the pt that anyone could be free to download it and deploy it on the real robot.

Flat
Foam_1
Foam_2

Roots environments

Stairs

Mass


Cross-Platform Validation

We show that our policy performs on the non-flat environments with Kyon robot which is Wheeled-Quadruped.

Box Grid
Pyramid Slope
Random Rough
Pyramid Stairs
Inverted Pyramid Stairs
Inverted Pyramid Stairs

Bibtex




  

Acknowledgements:

Template for this Website