Robot Motion Planning with Optimization-based Algorithm and Neural Network: Unsupervised Path Regression

Abstract

Planning collision-free motions for robots is challenging in environments with complex obstacle geometries. In this study, we combine the optimization-based algorithm and neural network to obtain a short and collision-free path through unsupervised path regression. The network can help to predict an educated initial path for an optimization-based planner, which will converge to a better solution. We compare different path representations, world representations, and also their combinations to improve the result. The motion planning problem is extended from a sphere robot in a single 2D world to a more complex static arm robot in multiple worlds. Through our experiments, we can find that in multiple worlds, the relative path with the signed distance field is the best combination for both robots. This is not only proved by metrics of feasible rate and loss but also by the visualization in the map.