A research team in Southwest Jiaotong Universit has published their latest study on 15 January 2026 in Frontiers of Computer Science co-published by Higher Education Press and Springer Nature, proposing a novel Bidirectional Chain-of-Thought (BiCoT) framework for zero-shot object navigation.
Zero-shot object navigation requires an AI agent to find a specific target in unseen environments without prior training. Traditional learning-based approaches struggle to generalize to new environments, while existing zero-shot methods often rely only on the target's perspective. To address this limitation, the BiCoT framework enables an embodied AI agent to reason about navigation paths from both the target's side and the agent's current observations.
BiCoT constructs two Chain-of-Thought (CoT) graphs—one predicting objects near the target and another detecting objects in the agent's view. By assessing the relevance between these graphs using a large language model, the agent efficiently explores areas with the highest correlation to the target.
Experimental results on the MP3D and HM3D benchmarks demonstrate that BiCoT significantly improves success rate (SR) and navigation efficiency (SPL) over previous zero-shot methods—by more than 3.0% on MP3D and 5.5% on HM3D.
Future research directions include enhancing the object-reasoning process and integrating multi-modal learning to further improve navigation capabilities in real-world applications.
