NavTrust: Benchmarking Robustness of Embodied Navigation Agents
Researchers have introduced NavTrust, a novel unified benchmark designed to assess the reliability of embodied navigation agents amid real-world corruptions. This benchmark methodically disrupts input modalities—RGB, depth, and instructions—within realistic contexts to evaluate their effects on navigation efficacy. It stands out as the first benchmark to challenge embodied navigation agents with a variety of RGB-Depth corruptions and instruction alterations in a cohesive framework. An analysis of seven leading approaches demonstrates significant performance decline when faced with realistic corruptions, underscoring the fragility of current models. This research is accessible on arXiv (2603.19229v2) and fills a crucial gap, as prior evaluations have largely concentrated on ideal conditions.
Key facts
- NavTrust is a unified benchmark for embodied navigation.
- It corrupts RGB, depth, and instruction modalities.
- It is the first benchmark to combine RGB-Depth corruptions and instruction variations.
- Seven state-of-the-art approaches were evaluated.
- Substantial performance degradation was observed under realistic corruptions.
- The benchmark addresses the gap of real-world robustness in navigation.
- The paper is available on arXiv with ID 2603.19229.
- The announcement type is replace-cross.
Entities
Institutions
- arXiv