Event

PASS-VLN proposes a traversability-aware local navigation benchmark and model

1 reports · 1 sources

Cong Zou et al. propose PASS-VLN, a traversability-aware benchmark and model for local-instruction vision-language navigation: the scene, instruction, starting pose, and reference path are fixed, and only the robot body width is varied; the robot is required to reach the goal when the passage is traversable and to stop early when it is blocked. The authors report that experiments show that, compared with strong VLN models fine-tuned on this benchmark, PASS-VLN improves the collision-free success rate by about 12 percentage points, and can adjust its reaching/stopping behavior according to body constraints and generalize to unseen body widths; when connected to a high-level planner, the success rate is higher and fewer local instructions are required, and deployment on a real robot as a low-level executor also demonstrated effectiveness. The benchmark and model will be open-sourced.