We address local trajectory planning for a mobile robot in the presence of static and dynamic obstacles. The trajectory is computed as a numerical solution to a Model Predictive Control (MPC) problem, with collision avoidance incorporated by adding obstacle repulsive potential to the MPC cost function. Our approach estimates this repulsive potential using a neural model. We explore three strategies for handling dynamic obstacles: treating them as a sequence of static environments, predicting a full sequence of repulsive potentials at once, and predicting future potentials step by step in an autoregressive mode.
0 commit comments