SleepWalk: A Three-Tier Benchmark for Stress-Testing Instruction-Guided Vision-Language Navigation
Researchers have introduced SleepWalk, a new benchmark designed to rigorously test instruction-guided vision-language navigation capabilities of AI models. This benchmark features a three-tier difficulty system, focusing on localized, interaction-centric embodied reasoning within 3D environments. Initial evaluations on frontier vision-language models revealed significant challenges, particularly with complex instructions, spatial reasoning under occlusion, and interaction constraints, indicating a need for further advancements in grounded multimodal reasoning and embodied agents. AI
IMPACT Provides a new evaluation framework to drive progress in embodied AI and grounded multimodal reasoning.