CVPR Findings2026
CoTFly: Making UAVs Think Where to Fly Next Through Visual Chain-of-Thought Reasoning
Meiqi Wang, Longnyu Xu, Jun Liu, Hewu Li, Han Qiu
Abstract
Recent vision-language navigation (VLN) has adopted vision-language models (VLMs) to predict actions for unmanned aerial vehicles (UAVs) to reach target locations by following natural language instructions and visual observations. However, current methods, mainly relying on imitation learning, may hesitate midway in long-horizon navigation tasks, especially when given target-oriented instructions (i.e., specifying only the final destination). Our insight is that introducing explicit reasoning can help UAVs to determine where to fly next at each intermediate location before acting. Thus, we propose CoTFly, a UAVspecific visual chain-of-thought (CoT) reasoning method for VLN tasks, especially the long-horizon navigation. CoTFly generates intermediate reasoning that can provide informative sub-goal landmarks and explicit visual guidance for further action making. To achieve a high-quality supervision for visual CoT learning, we first deploy an automated data generation pipeline with an efficient trajectoryaware sampling strategy to collect visually grounded subgoals with bounding boxes. We then fine-tune pre-trained large VLMs through two-round training to both enable visual CoT reasoning and enhance action supervision. Experiments show that CoTFly, built upon LLaVA-NeXT, achieves +11% in navigation success rate and reduces more than 20-meter navigation errors, superior to the state-ofthe-art VLN methods 1 .
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
- Visual Instruction Tuning cites Learning Transferable Visual Models From Natural Language Supervision
- Visual Instruction Tuning cites Flamingo: a Visual Language Model for Few-Shot Learning
- Learning Transferable Visual Models From Natural Language Supervision cites An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale
- Flamingo: a Visual Language Model for Few-Shot Learning cites Learning Transferable Visual Models From Natural Language Supervision
Builds on16
- Learning Transferable Visual Models From Natural Language SupervisionAlec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh et al.ICML 2021 · 47,906 citations
- An Image is Worth 16x16 Words: Transformers for Image Recognition at ScaleAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn et al.ICLR 2021 · 21,477 citations
- LoRA: Low-Rank Adaptation of Large Language ModelsEdward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu et al.ICLR 2022 · 18,833 citations
- Visual Instruction TuningHaotian Liu, Chunyuan Li, Qingyang Wu, Yong Jae LeeNeurIPS 2023 · 11,349 citations
- Flamingo: a Visual Language Model for Few-Shot LearningJean-Baptiste Alayrac, Jeff Donahue, Pauline Luc, Antoine Miech et al.NeurIPS 2022 · 6,707 citations
Related papers
- CoT-VLA: Visual Chain-of-Thought Reasoning for Vision-Language-Action ModelsQingqing Zhao, Yao Lu, Moo Jin Kim, Zipeng Fu et al.CVPR 2025
- FantasyVLN: Unified Multimodal Chain-of-Thought Reasoning for Vision-and-Language NavigationJing Zuo, Lingzhou Mu, Fan Jiang, Chengcheng Ma et al.CVPR 2026 · 4 citations
- Aux-Think: Exploring Reasoning Strategies for Data-Efficient Vision-Language NavigationShuo Wang, Yongcai Wang, Wanting Li, Xudong Cai et al.NeurIPS 2025 · 28 citations
- FlightGPT: Towards Generalizable and Interpretable UAV Vision-and-Language Navigation with Vision-Language ModelsHengxing Cai, Jinhan Dong, Jingjun Tan, Jingcheng Deng et al.EMNLP 2025 · 4 citations
- ACoT-VLA: Action Chain-of-Thought for Vision-Language-Action ModelsLinqing Zhong, Yi Liu, Yifei Wei, Ziyu Xiong et al.CVPR 2026 · 43 citations
