PINN∗: A Combined Approach for Path Planning around Irregular Obstacles

Sammanfattning

As interest in complex robotics systems increases, the shortcomings oftraditional path planning methods stand out. Traditional methods areoften physics-agnostic and generate suboptimal paths for real deployment.Additionally, they may not support arbitrary obstacles, hindering theirusefulness in real environments. Physics-Informed Neural Networks (PINNs)are a new paradigm in which neural networks can learn physical constraintsdirectly. In this paper, PINNs are applied to the path planning of a robotgoverned by a laterally constrained unicycle model traveling through a courseof irregularly shaped obstacles. The proposed model utilizes a signeddistance function for obstacle detection and some input from the traditional 𝐴∗algorithm for additional guidance. Two types of boundary conditions, hard andsoft constraints, are examined and have varying results for different obstacles.Additional constraints are placed on the control inputs and the smoothnessof the path is optimized for. Experimental results show that the model, aftertraining, is successful in traversing 3 distinct types of obstacles in a way that isconsistent with the unicycle model and the additional constraints. The modelis shown to be sensitive to its initialization and is very dependent on the guiding𝐴∗ algorithm, which could limit its real world deployment. PINNs show greatpotential for solving physics-bound path planning problems however, furtherresearch is necessary to develop more general and robust models ready for real world usage.