🌶️🤖𝐖𝐞 𝐜𝐨𝐨𝐤𝐞𝐝 𝐭𝐡𝐞 𝐫𝐨𝐛𝐨𝐭𝐢𝐜𝐬 𝐰𝐨𝐫𝐥𝐝’𝐬 𝐟𝐢𝐫𝐬𝐭 𝐩𝐥𝐚𝐭𝐞 𝐨𝐟 𝐌𝐚𝐩𝐨 𝐓𝐨𝐟𝐮.🤖🌶️ This was not just a cooking demo. It was a 30-minute, long-horizon robotics challenge packed with delicate, continuous, high-dexterity actions. With only a small number of demonstrations, our model learned to perform complex manipulation over an extended sequence. At the same time, we pushed hard on motion control and system-level optimization, making the robot move smoothly while keeping the high success rate. For us, this plate of Mapo Tofu is more than a dish. It is a small but meaningful step toward home robots that can handle real everyday tasks with the dexterity, consistency, and reliability people expect.

We should stop optimizing robot policies against a single overall reward. Trajectories differ along many axes, such as speed, precision, and subtask completion, and one can be better on some while worse on others. If we collapse all of that into a single overall axis we lose this structure making the reward ambiguous and harder to optimize. Blog: freeform-pl.github.io/fpl.we… Paper: arxiv.org/abs/2606.32027

Interaction with the real world is the major bottleneck in robot learning. So what would robot RL look like if we didn’t need to limit compute per interaction? Our latest work, Off-Policy Generative Policy Optimization (OGPO, accepted to ICML26) embarks on answering this question (spoiler alert: when done correctly, it helps massively!). 🧵(1/N)

Impressive long-horizon robustness!

This is incredible! This isn't a 24-hr stream doing essentially the same tasks over and over. This is a 30-min continuous sequence of various subtasks; any of them failing along the way we won't be getting no Mapo Tofu! Great stuff from @fancy_yzc & the PokeBot team.

When will I be invited to see such a long-horizon demo? It’s awesome.👏

Doesn’t seem that crazy anymore to have a robot chef. Impressive long horizon work!

more long horizon task and more robust manipulation systems!!'

Long horizon specialist models on the edge is what a lot of robots applications need

Impressive work—stable real-world long-horizon execution. Excited to see home robots moving forward! Congrats @fancy_yzc

Long horizon multi step task: cooking mapo tofu. It definitely feels like affordable, capable home robots are on the way

This is such a nice surprise. If robots can help me get mapo tofu, I’m officially sold 🤖🌶️

At the same time, we made many optimizations across motion control and the full robotic system, enabling smooth motion while maintaining a high success rate.

Video world model imaginations🌎💭can miss critical but plausible outcomes of robot actions. Introducing 𝙎𝙩𝙧𝙚𝙨𝙨𝘿𝙧𝙚𝙖𝙢: inference-time steering for video WMs, imagining plausible✅, high-impact⚠️ futures for 𝙧𝙤𝙗𝙪𝙨𝙩🛡️ policy evaluation and improvement. (1/15)

If the end goal of robot hands is to perform human motion, then we should optimize the hardware design with human motion - and on a large scale! We can generate both a high-dof generalist hand, and also low-dof specialized hands from human demonstration.

this took so long for me to understand: the bottleneck to more innovation is not more high intelligence people, but more people having an interest in hard problems it's impossible to create new useful things if you don't get immense happiness from making that thing

Children learn from play. Can robots do the same? We propose 𝐏𝐥𝐚𝐲𝐟𝐮𝐥 𝐀𝐠𝐞𝐧𝐭𝐢𝐜 𝐑𝐨𝐛𝐨𝐭 𝐋𝐞𝐚𝐫𝐧𝐢𝐧𝐠, a paradigm that gives embodied coding agents a play stage before downstream tasks arrive, and instantiate it with 𝐑𝐀𝐓𝐬 (Robotics Agent Teams), where robots discover reusable skills through curious play. Co-led with @jiaxin_ge_

Keeping human at the center is what allows us to cut through the noise. Unlike many previous works, UME is designed for humans — not the robot. Over the past few days, I’ve received a lot of great feedback on our hardware design. Truly appreciate all of it! I wanted to clarify a core philosophy behind our choices: Our hardware design is human-centered, the only goal is to maximize human comfort and recover the full range of motion of the human arm. Why? Because data quality and efficiency matters. When a user struggles to hold an awkward end-effector or faces joint limits, his time is wasted for nothing, and the data goes into trash. This is exactly the reason we committed to our hardware software combination of Coaxial Configuration and Universal Retargeting. Inspired by @StanfordHAI @Stanford

转自小红书：用 AI 读论文两年半，我认为最好用的 prompt xhslink.com/o/6XFS2HHAQVZ Github 中英文 Skill & Instruction 版: github.com/FeijiangHan/Paper… TLDR：总结论文不能被动吸收，要主动思考作者从 0 获得 idea 的逻辑，以及思考如何“攻击”这篇论文。 1. 论文提出并解决的研究问题是什么（调研和补充背景）？为什么这个问题是重要的？解决这个问题能带来哪些价值？ 2. 这个问题之前被解决了吗？之前的研究为什么存在不足？ 3. 【重要】在正式讲方法之前，先重建作者可能的思考路径。不要使用论文自己的贡献作为前提，只使用之前已有的背景、失败模式、经验观察和相关工作。思考和模拟作者本人的思路和受到的 inspiration 以及 intuition，引导我理解为什么作者可以基于已有知识想到这篇论文的 idea 4. 这篇论文提出方法的 Intuition 是什么？易懂清晰 concise 的告诉我这篇论文核心 idea 的本质 5. 这篇论文的具体方法是什么？结合一个真实的例子讲解：输入、处理、输出完整的 pipeline 6. 这篇论文的核心数学推导过程是什么（让 0 基础的我从理论视角理解方法）？如果有，请给我补充理论背景（我的数学比较差），告诉我理论的基础和 intuition；如果没有，可以说明并跳过这一点 7. 这篇论文是如何设计实验来验证提出的方法和 claim 的？按照下面格式总结：提出了什么问题->设计了什么实验验证这个问题->问题的答案是什么。不需要很多数据细节，只需要核心思路 8. 总结这篇论文的 take-aways 9. 【重要】这篇论文最脆弱的假设是什么？ 10. 如果我有 1 周时间，能做一个最小复现实验验证它的哪一点？ 11. 【重要】如果我反对它，我会怎么设计反例？ 12. 调研、思考、基于你的信息提出一个 follow up 的 idea，要 novel，不是增量研究，是从方法缺陷 Limitation和需求出发思考可能的新的有价值的研究 我认为 3 是最重要的，但又被很多人忽略。 我用 AI 总结论文收获最大的往往不是论文方法本身（毕竟大多数方法几个月后可能就过时了），而是作者是如何被启发的。 读论文要反向推理作者的思考过程，才能锻炼基于 related work 和 inspiration 提出 idea 的能力。 所以要按照上面的顺序总结，先了解背景，然后尝试自己思考，如果你是作者的话能不能想到解决方法，再结合AI逆向蒸馏作者的思考过程，然后修正你的思考。（这类似一种 meta learning，学习如何做 research，而非被论文细节干扰） 9-12 也很重要。我们要看这篇论文最大的限制，如何最小复现核心 claims，如何反例设计攻击它，然后尝试提出新的 ideas 补充要求： * 风格参考 Andrej Karpathy 和 Kaiming He，要求有真人的语感 * 使用详细的、准确的 claim，每句话都要有信息量，避免大空话和泛泛而谈 * 使用流畅的文本，避免滥用破折号、引号，保持输出内容清洁流畅，易读性高 * 使用真人逻辑，避免使用[不是...而是]这种 AI 的低信息量结构 * 请严格区分四类信息：论文原文明确声称的内容、相关文献中的已有结论、基于证据的合理推断、仍然不确定的猜测。不要把推断写成事实。 个人实践经验：LLMs 有上下文限制，一次完整总结 12 个问题可能会存在一些小细节缺失（但对我来说够用了，反正可以继续提问）；如果追求精度，可以每个问题单独提问，或者调用 subagent 单独处理，最后整合。 这个版本还可以继续扩充和优化，大家取长补短即可。

Another crazy CVPR 2026 world model result. “Envisioning the Future” forecasts where points in a scene will move, step by step, from a single image. No dense video needed. - 3,000x faster than video models, 10x fewer parameters, and 5x more accurate under a fixed compute budget. - An autoregressive diffusion model rolls sparse point trajectories forward through short, predictable steps, modeling uncertainty as it grows. - Why it matters: Rollouts get cheap enough to simulate thousands of futures and plan over them, hitting 78% billiard planning accuracy vs 16% for the best dense video baseline.