Researchers at Shandong University have developed a zero-shot object navigation method that addresses the real-time and accuracy limitations of existing approaches. The method integrates a lightweight open-vocabulary detector (YOLO-World) with a traditional CNN-based detector, using spatial constraints and confidence fusion to reduce false positives while maintaining low latency. This optimization is critical for open-world navigation where targets may be unseen during training.
Key takeaways
- Combines open-vocabulary and traditional detection: Spatial IoU filtering and confidence thresholds suppress erroneous detections from the open-vocabulary model, improving reliability for navigation decisions.
- Feedback-driven semantic extraction: A cross-attention network fuses target semantics with environmental features and incorporates historical state feedback, enabling the agent to adjust its exploration and avoid inefficient loops.
- Performance gains: In experiments, the method outperformed TDANet, improving unknown-object search accuracy by 19.5% and 22.0% under 18/4 and 14/8 category splits, respectively, and known-object accuracy by 6.6% and 6.3%.
- Real-world validation: Tests in physical environments confirmed adaptability across settings and generalization to unseen targets, highlighting practical applicability for service robots.
For manufacturers deploying autonomous mobile robots in dynamic settings, this work demonstrates a path to more reliable zero-shot navigation without heavy computational overhead, potentially enabling robots to handle novel objects and layouts with greater autonomy.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-08-24 · “基于开放词汇目标检测的零样本目标导航方法”