Staircase Negotiation Learning for Articulated Tracked Robots with Varying Degrees of Freedom

Tracked robots capable of negotiating 3D terrains require delicate control, most often tailored to a specific platform or setting. For staircase traversal in particular, autonomous robot behaviours are difficult to obtain due to the increased risk of accident and stochasticity. Based on a previously developed reinforcement learning based framework that allows learning staircase ascent for an articulated tracked robot, in this work we extend our work to allow also staircase descent and further investigate the role of a manipulating arm in the stability and smoothness of the traversal. By relying on a precise simulation of geometry and kinematics of a real robot, we demonstrate prototype policies for staircase ascent and descent, optionally under the influence of an integrated active arm and different penalty criteria. The obtained results are qualitatively and quantitatively compared and show that the robot can learn plausible behaviors effectively, when guided by appropriate reward and penalty criteria.