Learning blind stair-climbing with explicit stair parameters estimation for quadrupeds

Abstract

In the pursuit of enhancing legged robot mobility, stair navigation emerges as a significant challenge within urban contexts. While prior research has demonstrated the potential of Reinforcement Learning(RL) in designing robust blind controllers for rough terrains, the specific focus on staircase environments has remained limited. This study introduces a RL framework tailored to concurrently train a blind quadrupedal controller to traverse and explicitly estimate the stair geometry. By harnessing the inherent structural features of staircases, the developed controller can predict the next step and reduce collisions with consecutive step edges, optimizing locomotion efficiency. Experimental results demonstrate significant advancements in stair ascent performance, efficiently traversing stairs with slopes of up to $32\circ $ and step heights of up to $18 cm$, achieving an estimation accuracy of 88% and a collision probability of 6% after the second step. The controller’s capabilities extend beyond stair ascent, as evidenced by successful performance in stair descent and random command tracking on stairs. Furthermore, its adaptability is showcased in diverse challenging terrains, including various rough terrains, although additional experimental tests in these environments were not conducted. This work contributes in enhancing the existing blind controller research by offering insights into stair-specific scenarios.