Path Smoothing Through Learning to Optimize

In this paper, we propose a novel method based on the learning-to-optimize (L2O) framework for path smoothing. We adopt a hierarchical path-planning framework with a global collision-free guiding path generated beforehand. This path is then broken into segments to be fed into the L2O framework. We present a new cost function in L2O to indicate the smoothness of the expected path, which explicitly constrains the changes between successive path segments. The optimizer in the L20 framework takes as the input the gradient of the designed cost function and the hidden state of the last time step. It incrementally updates the path vertex coordinates and iteratively optimizes the path. The developed framework is verified through simulation experiments. It is observed that with the dedicated designed cost function, the Long Short-Term Memory (LSTM) unit in the developed L2O framework relies more on the input guiding path and releases the dependence on analyzing the environment, which further prompts the generation of the optimal path.