Researchers have introduced NavVerse, a new physics-enabled benchmark designed to evaluate embodied navigation for robots that must transition between indoor and outdoor environments within a single continuous episode. Existing benchmarks often assess these environments separately and abstract away crucial execution details like boundary traversal and adaptation. NavVerse includes 100 indoor, 50 urban outdoor, and 50 indoor-to-outdoor scenes, totaling 10,000 episodes across Object Navigation, Vision-and-Language Navigation, and Place Navigation tasks. Initial experiments with reinforcement learning and vision-language agent baselines indicate that current systems struggle significantly with this cross-context navigation, with modular methods showing a better safety profile despite lower success rates. AI
IMPACT This benchmark could accelerate the development of more capable robots for real-world applications by addressing the critical challenge of seamless indoor-outdoor transitions.
RANK_REASON The cluster describes a new benchmark for robot navigation published on arXiv.
Read on Hugging Face Daily Papers →
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
