Navigator, a web agentNavigator 网页智能体
Built the RL model for Navigator, a state-of-the-art AI agent that operates web browsers.
为 Navigator 训练了强化学习模型。Navigator 是一个领先的、能自主操作浏览器的 AI 智能体。
$ whoami
Tong Xiao肖桐
Freelance AI engineer and robot builder. 
LifeRPG player.
独立 AI 工程师，学造机器人中。LifeRPG 玩家。
I built multimodal foundation models and AI agents: Llama 3 Vision post-training at Meta and the reinforcement learning (RL) for the Navigator web agent at Yutori. Now I am a freelancer, learning/building robots. I hope everyone has a humanoid companion by the time I retire.
That intelligence correlates with is a law of this universe.
我曾做多模态基础模型与 AI 智能体：在 Meta 参与 Llama 3 Vision 后训练，在 Yutori 做强化学习（RL）训练 Navigator 网页智能体。现正在独立学习/制造机器人，希望退休前让所有人都有机器人伙伴。
智能和 正相关是本宇宙的一个定律。
News动态
All news →全部动态 →- Robot Overflow is live: a hub where we can better understand the robot industry.Robot Overflow 上线了。它能帮助我们更好地了解机器人产业。
- I made Robot Real2Sim to better understand what the Real2Sim system is and its potential value.我制作了Robot Real2Sim 来更好地了解什么是 Real2Sim 以及它的潜在价值。
- I left Yutori to spend more time with family.我离开了 Yutori，和家人多一些时间相处。
- We introduced Yutori Navigator, a state-of-the-art AI web agent. I built the RL model for it.我们推出了最先进的 AI 网页智能体 Yutori Navigator。我负责模型的强化学习。
- I joined Yutori to build AI agents for the web.我加入了 Yutori，致力于构建网页 AI 智能体。
Selected work代表工作
Llama 3 Vision post-trainingLlama 3 Vision 后训练
Co-led post-training for Llama 3 Vision — SFT, DPO, reward modeling, and rejection sampling. Worked on multimodal data, SFT, and reward modeling for Llama 4.
共同负责 Llama 3 Vision 的后训练（SFT、DPO、奖励模型、拒绝采样），并参与 Llama 4 的多模态数据、SFT 与奖励模型。
Emu Edit, image editing with textEmu Edit 文本驱动图像编辑
Led the production of Emu Edit, a diffusion model for free-form image editing from text instructions.
主导 Emu Edit（根据文字指令自由编辑图像的扩散模型）的产品化。
Face tracking on Quest ProQuest Pro 面部追踪
Shipped face tracking on Meta Quest Pro — system design, data collection and annotation, distributed training and evaluation.
在 Meta Quest Pro 上发布面部追踪。负责系统设计、数据采集与标注、分布式训练与评测。
Selected publications代表论文
All 21 publications →全部 21 篇论文与专利 →
Joint Detection and Identification Feature Learning for Person Search
CVPR 2017 · IEEE Conference on Computer Vision and Pattern RecognitionSpotlight
@inproceedings{xiaoli2016end,
title = {Joint Detection and Identification Feature Learning for Person Search},
author = {Tong Xiao and Shuang Li and Bochao Wang and Liang Lin and Xiaogang Wang},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
year = {2017}
}
Person Search with Natural Language Description
CVPR 2017 · IEEE Conference on Computer Vision and Pattern Recognition
@inproceedings{li2017person,
title = {Person Search with Natural Language Description},
author = {Shuang Li and Tong Xiao and Hongsheng Li and Bolei Zhou and Dayu Yue and Xiaogang Wang},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
year = {2017}
}
Learning from Massive Noisy Labeled Data for Image Classification
CVPR 2015 · IEEE Conference on Computer Vision and Pattern Recognition
@inproceedings{xiao2015learning,
title = {Learning from Massive Noisy Labeled Data for Image Classification},
author = {Tong Xiao and Tian Xia and Yi Yang and Chang Huang and Xiaogang Wang},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
year = {2015}
}
DeepReID: Deep Filter Pairing Neural Network for Person Re-Identification
CVPR 2014 · IEEE Conference on Computer Vision and Pattern Recognition
@inproceedings{li2014filter,
title = {DeepReID: Deep Filter Pairing Neural Network for Person Re-Identification},
author = {Wei Li and Rui Zhao and Tong Xiao and Xiaogang Wang},
booktitle = {IEEE Conference on Computer Vision and Pattern Recognition},
year = {2014}
}Lab工坊
The Lab →进入工坊 →
Robot Overflow
There are too many robot companies. We need a hub to better understand them. Check out this Robot Overflow site if you have the following questions:
- Does Tesla do everything in-house?
- Which companies use grippers?
- Do companies use robot or human data?
- How does NVIDIA support the robot industry?
机器人公司太多了，我们需要一个地方来更好地了解它们。如果你有下面这些问题，来 Robot Overflow 看看：
- Tesla 是不是什么都自己做？
- 哪些公司在用夹爪？
- 各家公司用的是机器人数据还是人类数据？
- NVIDIA 如何支持机器人产业？
Robot Real2Sim
What does Real2Sim mean for robots? Here, we break down the system and components for you and show the value for each level of capability, from L1 to L5.
Real2Sim 对机器人意味着什么？我们在这里为你拆解整个系统和各个组件，并展示从 L1 到 L5 每一级能力的价值。