Enhancing Realism in Indoor Navigation
Author(s)
Chhablani, Gunjan
Advisor(s)
Editor(s)
Collections
Supplementary to:
Permanent Link
Abstract
This body of work addresses key challenges in indoor navigation for embodied AI by enhancing simulation realism and developing adaptive training strategies to relax assumptions towards realism in policies. We introduce innovative methodologies in scene reconstruction and policy learning to advance goal-conditioned navigation in complex environments. The first of the works presents a novel framework that integrates 3D Gaussian Splatting with smartphone-based captures to create high-fidelity simulation environments. Our approach enables faster scene capture while maintaining high-quality performance, overcoming limitations of traditional, costly methods. Through extensive evaluations, we demonstrate the relationship between scene reconstruction quality, measured by Peak Signal-to-Noise Ratio (PSNR), and navigation performance, achieving successful policy transfer and improved performance in real-world scenarios. The second work tackles GPS dependence in indoor social navigation tasks, proposing curriculum learning strategies that progressively reduce GPS reliance. Our approach demonstrates significant performance gains, challenging conventional assumptions about GPS-based localization. Together, these contributions improve simulation fidelity, real-to-sim-to-real gap, and localization capabilities in agents, advancing the state-of-the-art in embodied AI navigation systems.
Sponsor
Date
2024-12-08
Extent
Resource Type
Text
Resource Subtype
Thesis (Masters Degree)