Reinforcement Learning on Synthetic Navigation Data allows Safe Navigation in Blind Digital Twins

Ilias Sarbout, Mehdi OUNISSI, Dan Milea, Daniel Racoceanu

Keywords: Electronic Travel Aids, Virtual Environment, Semantic segmentation, Reinforcement Learning

TL;DR: This study describes an innovative method for extraction of low-dimensional cues for navigation in blinds.

Abstract: Limited access to dedicated navigation data in visually impaired individuals is a significant bottleneck for developing AI-driven assistive devices. For this purpose, we have developped a virtual environment designed to extract various human-like navigation data from procedurally generated labyrinths. Using reinforcement learning and semantic segmentation, we trained a convolutional neural network to perform obstacle avoidance from synthetic data. Our model outperformed state-of-the-art backbones including DINOv2-B in safe pathway identification in real world. In conclusion, despite being trained only on synthetic data, our model successfully extracted features compatible with safe navigation in real-world settings, opening new avenues for visually impaired.

Primary Area: applications to neuroscience & cognitive science

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 10204