Binary Tracking for Spatial QA and Navigation with Open Vision-Language Models
Researchers have developed BinTrack, an open-source agent for spatial question answering and navigation in robots, designed to operate without reliance on unstable or costly closed-source models like GPT-4o. BinTrack utilizes a binary search approach over trajectory segments, achieving up to a 22.8% accuracy improvement over existing open-source methods and matching closed-source performance on the SpaceLocQA benchmark. The system also offers a 1.5x inference speedup. Additionally, the team released GangnamLoop, a new outdoor benchmark dataset collected with a real quadruped robot. AI
IMPACT This open-source approach could enable more robots to perform spatial reasoning and navigation tasks reliably without depending on cloud-based models.