研报速递在密集且动态的人群中实现机器人视觉导航，需要多样化且可靠的训练数据集；然而，在现实世界中采集此类数据成本高昂，且难以充分覆盖各种交互行为及罕见事件。
为了实现导航数据的规模化生成，本研究开发了两种利用高保真仿真技术和网络级第一人称视角视频的方法。首先，我们开发了NavIsaacLab，这是一个集成了对抗性运动控制功能的GPU并行仿真平台。该平台能够生成动作自然且交互多样化的虚拟人群，从而加速导航策略的训练与标准化评估。
其次，我们建立了一条自动化流水线，通过基于视觉语言模型（VLM）的推理标注，将网络视频转化为城市导航样本。导航基础模型利用历史相机观测数据和点导航目标，学习结构化的场景理解与运动决策能力。
为了覆盖非结构化城市环境中的“长尾”场景，我们从语义稀缺性和轨迹预测误差两个维度对网络视频数据集进行了挖掘。随后，基于反思机制（reflection-based）的微调进一步提升了模型从罕见及复杂案例中学习的能力，从而增强了其在多样化城市导航场景中的泛化性能与鲁棒性。
Jiankun Wang (Senior Member, IEEE) received the B.E. degree in Automation from Shandong University, Jinan, China, in 2015, and the Ph.D. degree from the Department of Electronic Engineering, The Chinese University of Hong Kong, Hong Kong, in 2019. He is currently an Assistant Professor with the Department of Electronic and Electrical Engineering, Southern University of Science and Technology, Shenzhen, China. During his Ph.D. study, he spent six months at Stanford University, CA, USA, as a Visiting Student Scholar, supervised by Prof. Oussama Khatib. His current research interests include robot motion and path planning, human-robot interaction, and artificial intelligence. He serves as the Associate Editor of IEEE Transactions on Automation Science and Engineering, IEEE Robotics and Automation Letters, and Biomimetic Intelligence and Robotics.
