Obstacle negotiation learning for a compliant wheel-on-leg robot

Abstract

Building versatile control for wheel-on-leg robots on various uneven terrains still constitutes a challenge. In this paper, we propose to combine a continuous state space Q-learning algorithm with a compliant wheeled locomotion structure capable of measuring forces applied by the environment. These forces are used to choose a behavior from the modulation of actuation distribution to simple configuration adjustments. Then, the robot does not need any prior knowledge on the ground geometry and is able to react to non-anticipated obstacles. The learned policy proves to be generic and allows the robot to negotiate complex obstacles that were not seen during learning.

MARCEL: Mobile active rover chassis for enhanced locomotion

“…Moving the relative wheel positions allows the rover to bring itself into a situation it can deal with using the assistance of the aforementioned bogie torque. The reconfiguration then offers the possibility to sequence the crossing of complex obstacles by isolating the difficulty on one wheel at a time, as suggested previously in Bouton et al (2017). Moreover, as seen in Section 2.2, relative wheel displacements can also play a role in reducing the friction required to overcome the obstacles in combination with the bogie torque.…”

“…To reduce the amount of actuation while keeping a control over the stance and load distribution on wheels, the Complios (Bouton et al, 2020) combines structural compliance with a wheel-on-limb kinematics. This way, it requires only one motor per limb and can autonomously cross complex obstacles (Bouton et al, 2017). However, its mechanism, involving a lot of cables and springs, is still way too complex to be realistically considered for space applications.…”