NavTrust Benchmark Evaluates Embodied Navigation Under Input Corruption
August 7, 2026
NavTrust introduces a unified benchmark for Vision-Language Navigation (VLN) and Object-Goal Navigation (OGN) that systematically corrupts RGB, depth, and instruction modalities. Testing seven SOTA models reveals significant performance degradation when agents encounter realistic sensory and linguistic variations.
HOW THIS AFFECTS YOU
●
builderThis helps you identify failure modes in embodied agents before deploying them in unconstrained environments.
●
researcherYou can use this to evaluate model robustness against real-world sensor noise.