CovNav preprint improves robot navigation with selective slow reasoning

Researchers report that CovNav, a training-free robot navigation framework, improved success rates over the strongest zero-shot baseline by 0.8 and 1.7 percentage points on HM3D v1 and v2, while using slow reasoning for only 12.4% of queries. The framework asks a local language model whether a target object is likely near an observed object. It uses calibrated uncertainty to send only ambiguous queries to slower reasoning, retaining most of the navigation gain from always using that mode with less latency. The preprint also reports TurtleBot 4 experiments, but its quantified success-rate improvements come from navigation benchmarks.