World models for robotics机器人世界模型
Action-conditioned world models that predict how a scene will change under a robot's actions — and how to use them for planning, policy learning and evaluation.
以动作为条件的世界模型：预测场景在机器人动作下会如何变化，并将其用于规划、策略学习与评测。
Embodied AI research lab · San Francisco具身智能研究实验室 · 旧金山
World models, measured against the real world.世界模型，以真实世界为尺。
Junfan Labs studies how robots learn predictive models of the physical world — and builds the evaluations that show whether those models hold up on real hardware, from simulation to deployment and across robot bodies.
Junfan Labs 研究机器人如何学习对物理世界的预测模型，并构建相应的评测体系，检验这些模型在真实硬件上是否经得起考验——从仿真到部署，跨越不同的机器人本体。
- Predicted rollouts预测轨迹
- World-model plan世界模型规划
- Measured on hardware真机实测
- ΔEvaluation gap评测误差
Research研究
What we work on我们在做什么
Four connected problems that decide whether physical AI works outside the lab.四个彼此关联的问题，决定了物理 AI 能否真正走出实验室。
Evaluation & benchmarks评测与基准
Benchmarks and protocols that track real-world performance: matched events across synthetic, simulated and physical domains, long-horizon tasks, and systematic failure analysis.
能反映真实表现的基准与评测协议：在合成、仿真与物理三个域中对齐同一事件，覆盖长时程任务，并进行系统化的失败分析。
Sim-to-real transfer仿真到真机迁移
Methods that carry policies from GPU-accelerated simulation and synthetic data onto physical robots with fewer surprises.
让策略从 GPU 加速仿真与合成数据更稳定地迁移到真实机器人上，减少落地时的意外。
Cross-embodiment learning跨本体策略学习
Policies and datasets that transfer across robot bodies, sensors and hardware revisions, instead of starting over for each one.
让策略与数据在不同的机器人本体、传感器和硬件版本之间迁移复用，而不是每换一次硬件就从头再来。
Approach方法
How we work我们的工作方式
A research loop that ends on real hardware, not on a leaderboard.研究的闭环落在真实硬件上，而不是排行榜上。
Decide on hardware以真机为准
Simulation is where we iterate quickly; physical trials are where we decide what works.
仿真用来快速迭代，真机实验用来判断什么真正有效。
Publish what we learn公开我们的所学
We share benchmarks, evaluation protocols and technical write-ups with the research community.
我们向研究社区分享基准、评测协议与技术解读。
Built on accelerated computing建立在加速计算之上
Our roadmap builds on GPU-accelerated simulation and world foundation models, including NVIDIA Isaac Sim, Isaac Lab and Cosmos, with Jetson-class compute for on-robot inference.
我们的研究路线建立在 GPU 加速仿真与世界基础模型之上，包括 NVIDIA Isaac Sim、Isaac Lab 与 Cosmos，并以 Jetson 级算力支持机器人端推理。
Founder创始人
Junfan is a physical AI researcher working on world-action models, sim-to-real transfer and cross-embodiment policy learning. He founded Saturday Robotics, a weekly research forum in San Francisco on robot world models, and served on the program committee and as a session chair of the IROS 2026 workshop on Physical World Models for Scaling Embodied AI. He also organized the ACM SIGGRAPH 2026 Birds of a Feather session on world models for robotics, and reviews for NeurIPS 2026.
朱俊帆是一名物理 AI 研究者，研究方向包括世界-动作模型、仿真到真机迁移与跨本体策略学习。他创办了旧金山每周一期的机器人世界模型研究论坛 Saturday Robotics，并担任 IROS 2026「Physical World Models for Scaling Embodied AI」研讨会程序委员会委员和分会主席；他还组织了 ACM SIGGRAPH 2026 机器人世界模型 Birds of a Feather 专场，并担任 NeurIPS 2026 审稿人。
Before Junfan Labs, he was a machine learning engineer on a production agentic VLA autonomy stack at Tensor Auto, and a machine learning quant researcher at Société Générale. He holds an M.S. in Computer Science from Georgia Tech and an M.S. in Financial Mathematics from the University of Chicago.
创办 Junfan Labs 之前，他在 Tensor Auto 担任机器学习工程师，参与量产级智能体 VLA 自动驾驶系统的研发；更早在法国兴业银行（Société Générale）担任机器学习量化研究员。他拥有佐治亚理工学院计算机科学硕士和芝加哥大学金融数学硕士学位。
Selected publications代表性论文
Co-authored by Junfan Zhu.朱俊帆参与合著。
-
Agents' Last Exam (ALE)
NeurIPS 2026 · E&D Track
A long-horizon agent benchmark of professional workflows across 55 fields.
覆盖 55 个领域专业工作流的长时程智能体基准。
-
An Interactive Enhanced Driving Dataset for Autonomous Driving (IEDD)
Scientific Data · 2026
Interaction-dense driving scenarios with a hierarchical benchmark for vision-language models.
高交互密度的驾驶场景数据集，附面向视觉语言模型的分层评测基准。
-
SSP: An Event-Matched Syn2Sim2Phy Cross-Domain Evaluation Framework for Autonomous Driving VLA Models
arXiv · 2026
Evaluates driving VLA models on matched safety-critical events across synthetic, simulated and physical domains.
在合成、仿真与物理三个域中对齐同一安全关键事件，评测驾驶 VLA 模型。
-
GestureLSM: Latent Shortcut based Co-Speech Gesture Generation with Spatial-Temporal Modeling
ICCV 2025
Compresses multi-step diffusion sampling into a few steps for coordinated full-body gesture generation.
将多步扩散采样压缩为少数几步，生成全身协调的伴随语音手势。
-
QuantEval: A Benchmark for Financial Quantitative Tasks in Large Language Models
arXiv · 2026
Knowledge QA, quantitative reasoning and strategy coding, with a backtesting framework.
涵盖知识问答、量化推理与策略编程，并配备回测框架。
Community社区
Saturday Robotics
A weekly, in-person research forum in San Francisco on robot world models and embodied intelligence, founded by Junfan Zhu. Sessions are built around recent research and the people doing it, and draw PhD students, industry researchers and founders from Google DeepMind, NVIDIA, Stanford, UC Berkeley, CMU, MIT and leading robotics startups.
由朱俊帆在旧金山创办的每周线下研究论坛，聚焦机器人世界模型与具身智能。每期围绕前沿研究和做研究的人展开，参与者包括来自 Google DeepMind、NVIDIA、斯坦福、UC Berkeley、CMU、MIT 以及众多领先机器人创业公司的博士生、业界研究员和创业者。
Partners include Samsung Research America, Dyna Robotics, Robotics Center of Silicon Valley, Stanford Robotics Club and MIT Robotics Club.
合作伙伴包括 Samsung Research America、Dyna Robotics、硅谷机器人中心（Robotics Center of Silicon Valley）、斯坦福机器人俱乐部和 MIT 机器人俱乐部。
Saturday Robotics is a separate organization from Junfan Labs.Saturday Robotics 是独立于 Junfan Labs 的组织。
30+weekly sessions since 2026期每周论坛（始于 2026 年）
3,600+subscribers on LumaLuma 订阅者
200–300sign-ups per weekly session每期报名人数
380+registrations for the IROS 2026 Research NightIROS 2026 研究之夜报名人数
Conference events顶会活动
- CVPR 2026 Research Night研究之夜Denver · Jun 2026丹佛 · 2026 年 6 月
- ACM SIGGRAPH 2026 Birds of a FeatherBoF 专场Los Angeles · Jul 2026洛杉矶 · 2026 年 7 月
- IROS 2026 Robotics Research Night机器人研究之夜Pittsburgh · Sep 2026匹兹堡 · 2026 年 9 月
- CoRL 2026 Research Night, with GenRobot研究之夜（与 GenRobot 联合）Austin · Nov 10, 2026奥斯汀 · 2026 年 11 月 10 日
Contact联系
Work with us与我们合作
Research collaborations, evaluation partnerships and speaking invitations are welcome.欢迎研究合作、评测合作与演讲邀请。