Ground Traversability Estimation for Quadruped Robot with Modified DWA Algorithm

For path planning of robots, it is very much essential that the traversability map of regions on which robot is intended to navigate must be available, the robot must have an idea about surface roughness along with obstacles of surface which it traverses. There are various methods for estimation of traversability on rough terrain, but those which are based on deep learning are found to be most efficient. In this work we have followed deep learning approach in which terrain heightmaps are converted into images and the quadruped robot is simulated on various terrains, the part of terrain where robot gets stuck the corresponding image patch is labeled as non-traversable and the part of terrain where robot travels easily the corresponding heightmap image patch is labeled as traversable. Further the data obtain from simulations is feed to CNN classifier which is then used for estimating traversability of unknown terrains for quadruped. In this paper the algorithm is proposed, which uses this trained CNN classifier for making robots traverse through smoother and safer paths.