Summary

NavTrust is a unified benchmark for evaluating the trustworthiness of embodied navigation agents, presented in the paper arXiv:2503.16908 by Huaide Jiang, Yash Chaudhary, and Yuping Wang. The benchmark systematically corrupts input modalities—including RGB images, depth data, and natural language instructions—under realistic scenarios and measures the resulting impact on navigation performance. NavTrust is the first benchmark to expose embodied navigation agents to a diverse range of RGB-Depth corruptions and instruction variations within a single unified evaluation framework. By providing systematic perturbations across multiple input modalities, it enables researchers to quantify how robustly navigation policies degrade when real-world sensory noise, sensor failures, or ambiguous instructions occur, addressing a key gap in reliable embodied AI deployment. Posted on zhichai.net, this paper summary highlights the benchmark's role in advancing robustness evaluation for vision-and-language navigation systems, offering a standardized protocol for comparing navigation agents under realistic corrupted-input conditions.