Researchers from the Chinese Academy of Sciences have developed AeroVerse-SkyPlan, a vision-language model for UAV embodied task planning that integrates an aerospace embodied chain-of-causality to improve reasoning in complex urban environments. The model simulates human cognitive decision-making through four stages: environment perception, spatial reasoning, path description, and result summarization, enabling deeper logical inference and better interpretability compared to conventional input-output mapping approaches.
Key takeaways
- Introduces a chain-of-causality instruction fine-tuning paradigm that structures reasoning into four stages, enhancing the model's ability to handle complex, long-range UAV missions.
- Proposes an automated data generation pipeline that uses GPT-4o with 3D point cloud mapping to create high-quality causal chain data (SkyAgent-AeroChain3k) at low cost, balancing quality and efficiency.
- On the SkyAgent-Plan3k dataset, AeroVerse-SkyPlan outperforms GPT-4o by 35%–47% on the BLUE metric, demonstrating significant gains in task planning performance across models of varying parameter sizes (1B to 8B).
For industrial automation, this work highlights the potential of causal reasoning in embodied AI for aerial applications such as delivery, surveillance, and inspection, where reliable long-horizon planning is critical. The automated data generation method also offers a scalable approach to training specialized models without excessive manual annotation.
Source: 《机器人》期刊 (robot.sia.cn) · Published 2026-08-24 · “AeroVerse-SkyPlan：空天具身因果链增强的无人机具身任务规划模型”