Computer Science · Computer Vision and Pattern Recognition
Neural Network Based Collision-Free Trajectory Generator for Quadcopters
Praveen Jawaharlal Ayyanathan, Ehsan Taheri, Tyler Flegel, Nan Li
See on the mapAbstract
We present a neural-network based framework for collision-free trajectory generation for quadcopters navigating in environments with static obstacles. Using trajectories generated through a successive convexification method, three neural network (NN) models: 1) NN-Trajectory, 2) NN-Parameter, and 3) NN-Segment were developed to approximate and reconstruct optimal trajectories for one- and two-obstacle scenarios. Hyperparameter tuning and architecture optimization for the NNs were performed using Optuna ensuring efficient performance. A cascaded structure using NN-Parameter and NN-Segment is proposed to enhance the interpretability of the framework. Additionally, a safety margin correction step was introduced to dynamically adjust position-level information of parameter points (points closest to the obstacles), ensuring collision-free trajectories with negligible loss of accuracy or optimality. The framework was validated across multiple scenarios, demonstrating its capability.