Influence of Data Collection Parameters on Performance of Neural Network-based Obstacle Avoidance

Neural networks are becoming wide-spread, including applications in mobile robotics and related fields. Most state-of-the-art approaches to training neural networks use video cameras for generating training datasets. However, these data are hard and time-consuming to collect resulting in a bottleneck of neural network training procedure. Thus, the paper briefly presents simulation-based LiDAR data collection for the training of neural networks for obstacle avoidance. The influence of two data collection parameters in simulation (distance to obstacles and number of LiDAR points) on the performance of the realworld mobile robot is analysed in more depth. Experimental testing was performed in a narrow corridor (augmented with additional obstacles) in order to fully test the neural networks and detect possible limitations. For a better understanding of proposed algorithms and analysis of their performance in real-life scenarios, a simple test-bed was devised with Turtbebot 2 as a test vehicle although it can be applied on similar mobile robot platforms. Based on obtained results, and with safety in mind, conclusions are drawn and possible future improvements proposed.