Small vision language model guides robot navigation efficiently
Structure over Reasoning: Unified Zero-Shot Navigation with a Small Vision-Language Model
Robotics
Summary
Teaching robots to find their way indoors using images and language is hard, especially when the robot has to decide where to go and when to stop by itself. The researchers tested a small computer brain that is good at recognizing if a scene looks like the goal but not good at planning routes. They built a system that asks this small brain simple questions each step and uses a map-based approach to combine the answers, making navigation decisions more reliable. Their method outperforms other similar zero-shot methods that don’t require extra training when tested on realistic simulated environments and works on a real robot too.
What this means in practice
- •For robot developers: Build navigation systems for robots that use small vision-language models to improve goal recognition and stopping decisions without complex route planning.
- •For autonomous delivery teams: Implement efficient zero-shot navigation agents on quadruped robots to navigate realistic indoor environments without extensive retraining.$Commercial implications: Enables reliable indoor delivery robots that navigate complex environments with minimal computational resources, improving deployment feasibility.
Authors
Zeqi Li, Xunyi Zhao, Zerui Li, Sihao Lin, Qi Wu, ShiJie Li
Abstract
Vision-language navigation is moving into photorealistic, physically simulated scenes, where the leading zero-shot agents put a vision-language model in the control loop to choose every move and stop. Robots favour small models, whose abilities are uneven: Qwen3-VL-4B can judge whether a single view looks like the goal, but its route reasoning is at chance, its distance estimates nearly constant, its stop votes unreliable. Our agent therefore asks it two perceptual questions per step, which direction looks like the goal and whether the view justifies stopping, plus one gated frontier query on route-bearing instructions, and treats the answers as evidence, not actions: a route-blind semantic field over a bird's-eye map accumulates the directional scores through a tempered update, a skeleton graph supplies frontiers and geodesic distances, and a two-stage terminator turns stop votes into termini through a trail-consistency gate and a budget-forced return. With every baseline re-run in one installation at one budget, the agent exceeds all four zero-shot baselines of the VLNVerse subset in mean success on both granularities, though the paired interval against the strongest coarse-grained baseline includes zero. Termination is the dominant lever in ablation, thirteen attempts to give the model more of the navigation establish no reliable gain, and the agent runs on a Unitree Go2 quadruped with off-board inference. Code, per-episode logs and a robot video will be released at https://anonymous.4open.science/r/structure-over-reasoning-PAGE.