Steppability is the ability of a legged robot to place a stable foothold at a given location. This approach predicts steppability directly in the camera image space with a semantic segmentation model trained on synthetic primitive-shape scenes (regions labeled steppable, passable, or non-passable), and uses the learned model as a visual search heuristic inside an interleaved graph-search and trajectory-optimization footstep planner, enabling online reactive contact planning for quadrupeds in unknown terrain without building a Cartesian height map.
Steppability is the ability of a legged robot to place a stable foothold at a particular location in its surrounding terrain. This paper predicts steppability directly in the camera image space using…