| Samenvatting: | Legged robots promise a clear advantage in unstructured and challenging terrain, scenarios such as disaster relief, search and rescue, forestry and construction site. Dynamic locomotion on rough terrain has to guarantee stability and maximizing the cross-ability of a local set of candidate footholds. Trajectory optimization improves such performance metric while satisfying locomotion stability. Terrain conditions increase significantly the dimensionality of the optimization problem. Moreover, decoupling footstep selection and Center of Mass (CoM) motion generation may limit the success of the task. We are inspired by the observation that humans solve complex problems through intensive reasoning in the initial phases, which allows them to solve faster and naturally similar problems. In the same vein, the preview optimization allows the robot to infer the locomotion skills required on challenging terrain, and then use the data to build a locomotion policy that can be the used in real-time. A set of preview model allows us to reduce the dimensionality of the problem, which is desirable for trajectory optimization and policy reconstruction.
|
|---|