Event

StopWatch Improves Object Goal Navigation with Explicit Goal Grounding

1 reports · 1 sources

Xinhao Liu and other researchers propose the StopWatch method, which addresses the main failure mode in object goal navigation through explicit goal grounding, and achieves state-of-the-art results on six benchmark datasets across three scene collections. The paper diagnoses that object goal navigation policies based on vision-language models (VLMs) perform well in exploration, yet often fail to notice that the goal is already in view or to estimate its distance; goal grounding is the primary cause of failure. The researchers also corrected annotation flaws in these benchmarks, enabling more accurate evaluation of future methods. The researchers state that, when using ground-truth grounding, their policy nearly reaches the level of the expert that generated the demonstrations, indicating that the remaining room for improvement lies in perception rather than policy learning.