Konrad Cop
supervisor: Tomasz Trzciński
In order to navigate through an environment a robot must be able to estimate a path which avoids all obstacles on its way. If a new object appears in the environment, the information about the shape, which a robot perceives, can be incomplete as the observation is done from a single viewpoint only. To avoid driving around the obstacle for full observation we propose a system that predicts the obstacle footprint based on the partial view form the onboard sensors. Our system is based on autoencoder to predict a 2D footprint and a PointNet -based network to estimate potential shift of a partial cloud origin. To provide close to reality data we propose a data generation pipeline that emulates sensor views using simulation environment and meshes of real object models. The output of the system is estimated 2D footprint which can be directly used in navigation occupancy grid.
Our solution is validated on a synthetically generated dataset and can heavily reduce additional motion which a robot would need to perform to obtain the full view of the obstacle.