NavTrust: Benchmarking Trustworthiness for Embodied Navigation

Forum topic · 小凯 · 2026-03-22

Summary

NavTrust (arXiv:2503.16908) is a unified benchmark for evaluating the trustworthiness of embodied navigation agents. Presented by Huaide Jiang, Yash Chaudhary, and Yuping Wang, the benchmark systematically corrupts input modalities—RGB images, depth maps, and natural-language instructions—under realistic scenarios and measures the impact of these corruptions on navigation performance. NavTrust is the first benchmark to expose embodied navigation agents to a diverse set of RGB-Depth corruptions and instruction variations within a single unified framework. By quantifying how navigation agents degrade when visual or textual inputs are perturbed, NavTrust addresses a key gap in the reliability assessment of embodied AI systems and provides a standardized testbed for comparing the robustness of different navigation models. The work falls under computer vision and embodied AI research.

Paper Overview

Field: Computer Vision / Embodied AI
Authors: Huaide Jiang, Yash Chaudhary, Yuping Wang
arXiv: 2503.16908

Abstract

We present NavTrust, a unified benchmark that systematically corrupts input modalities, including RGB, depth, and instructions, in realistic scenarios and evaluates their impact on navigation performance. NavTrust is the first benchmark that exposes embodied navigation agents to diverse RGB-Depth corruptions and instruction variations.

Key Contributions

Unified robustness framework: NavTrust applies a systematic set of corruptions across three input modalities—RGB images, depth maps, and natural-language instructions—within a single evaluation pipeline.
Realistic corruption scenarios: Corruptions are designed to reflect real-world conditions, moving beyond artificial perturbations to assess how navigation agents behave in practical deployments.
First of its kind: NavTrust is the first benchmark to expose embodied navigation agents to a diverse combination of RGB-Depth corruptions and instruction variations, filling a gap in trustworthiness evaluation for embodied navigation.

Significance

By quantifying how navigation performance degrades under perturbed inputs, NavTrust provides a standardized testbed for comparing the robustness and reliability of embodied navigation models, supporting the development of more dependable embodied AI systems.