I Have No Eyes and I Must Chirp: A Bio-Inspired, Neural Network Based, Echolocating Robot
Authors
Mikkelsen, Mads Wenneberg ; Andersen, Asbjørn Møller
Term
4. semester
Education
Publication year
2026
Submitted on
2026-06-04
Abstract
This thesis describes the development of a bio-inspired robot that uses echolocation and neural networks to operate in environments with poor visibility where GPS is not available. The system uses an ultrasonic speaker and two ultrasonic microphones to send out sound waves and record their echoes. Based on these echoes, a neural network was designed and trained using a dataset collected specifically for this project. The neural network estimates distances to objects in the surroundings, and these distance predictions are compared with measurements from a lidar sensor, which serves as the ground truth reference. During testing, the system showed that this approach is viable and can be used for distance estimation in practice, but several issues also became clear. The results suggest that many of these limitations are caused by the small and not very diverse training dataset for the neural network. This indicates that using a larger and more varied dataset would likely improve the accuracy and robustness of the system.
Denne opgave beskriver udviklingen af en bio-inspireret robot, der bruger ekkolokation og neurale netværk til at navigere i omgivelser med dårlig sigtbarhed, hvor GPS ikke kan bruges. Systemet består af en ultralydshøjtaler og to ultralydsmikrofoner, som udsender lydbølger og opfanger deres ekko. Ud fra disse ekkoer har vi designet og trænet et neuralt netværk på et datasæt, der er specielt indsamlet til projektet. Det neurale netværk beregner afstande til objekter i omgivelserne, og disse afstande sammenlignes med målinger fra en lidar-sensor, som fungerer som sand referenceværdi. I testfasen viste systemet sig at være lovende og kunne i praksis bruges til afstandsbestemmelse, men flere problemer kom også frem. Meget tyder på, at mange af udfordringerne skyldes det begrænsede og lidt varierede datasæt til træning af det neurale netværk. Resultaterne peger derfor på, at et større og mere mangfoldigt datasæt vil kunne forbedre systemets nøjagtighed og robusthed.
[This abstract has been rewritten with the help of AI based on the project's original abstract]
Keywords