ReferTrack: Referring Then Tracking for Embodied Visual Tracking
南方科大& 腾讯开源ReferTrack:单目VLA 刷新具身视觉跟踪 ...
This paper proposes ReferTrack, a new paradigm for embodied visual tracking that combines referring and tracking. The model uses a single forward-facing camera, selects targets via index detection boxes, and employs TVBI temporal visual boundary information to achieve state-of-the-art results on EVT-Bench. It is deployable on quadruped and humanoid robots, showing strong simulation-to-reality transfer capabilities.
Source: Chinese robotics — Robot learning research · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
- Environment
- SimulationOpen source S10
- code: Available Artifact link Open source S4
- paper: Available Artifact link Open source S4
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
EVT‑Bench基准上单目设置拿下SOTA,在目标极易混淆的任务上,效果甚至优于部分多相机基线;同时完成四足、人形机器人部署,证明强大的仿真到现实迁移能力。
Open source S10
消融实验(在干扰跟踪DT任务上开展) Oracle实验组:去掉Refer‑CoT,直接给TVBI喂真值目标框,SR达到81.5%,和专家策略差距很小。
Open source S22
保留Refer‑CoT,移除TVBI模块:SR下降到70.4%。
Open source S23
Refer‑CoT+TVBI全部移除:SR暴跌至55.7%。
Open source S24
损失函数 总损失由三部分加权组合: 航点损失:预测航点和专家轨迹做 MSE 均方误差,权重系数 指代损失:交叉熵损失,监督 Refer‑CoT 输出正确的目标索引。
Open source S18
阶段2 :导航仿真样本(130万)与Refer‑QA指代问答样本(130万)按1:1混合训练;视觉编码器冻结,其余全部参数参与更新。
Open source S20
任务成功判定标准:机器人和目标保持 1‑3米跟随距离 ,并且目标持续出现在相机视野内。
Open source S13
作者 :Hanjing Ye , Tianle Zeng , Jiazhao Zhang , Shaoan Wang , Zibo Zhang , Weisi Situ , Yuchen Zhou , Yonggen Ling , Hong Zhang 单位 : 南方科技大学RCV实验室, 腾讯Robotics X, 北京大学, 福田实验室 论文标题 :ReferTrack: Referring Then Tracking for Embodied Visual Tracking 论文链接 :https://arxiv.org/abs/2607.20061 代码链接 :https://github.com/MedlarTea/referTrack 研究背景 具身视觉跟踪EVT ,简单说就是:机器人只靠自身机载摄像头,听懂人类自然语言指令,持续跟随指令描述的特定行人。
Open source S4
Source:Chinese robotics — Robot learning research · 163.com