https://yipko.com/ Recent content on YK's Blog Hugo -- gohugo.io zh-cn Tue, 11 Aug 2026 19:33:01 +0800 https://yipko.com/posts/work/history-guided-video-diffusion/ Tue, 11 Aug 2026 19:33:01 +0800 https://yipko.com/posts/work/history-guided-video-diffusion/ DFoT 用逐帧独立噪声等级支持任意历史条件，History Guidance 再组合不同时间窗口和频率信息下的 score，在长视频生成中改善质量、一致性与运动幅度。 https://yipko.com/posts/work/one-step-diffusion-via-shortcut-models/ Tue, 11 Aug 2026 18:35:00 +0800 https://yipko.com/posts/work/one-step-diffusion-via-shortcut-models/ Shortcut Models 在同一个网络里联合学习局部 flow 和不同步长的有限跳跃，用二分自一致目标把多步生成能力递推到少步乃至单步推理。 https://yipko.com/posts/work/mars-policy-multimodality-only-when-it-matters/ Sat, 30 May 2026 17:36:37 +0800 https://yipko.com/posts/work/mars-policy-multimodality-only-when-it-matters/ MARS 把 flow matching 的初始源分布改成高斯噪声与历史动作先验的自适应混合：任务阶段需要多模态时保留随机性，阶段接近单模态时退回更高效的确定性动作生成。 https://yipko.com/posts/work/a2a/ Sun, 24 May 2026 05:18:15 +0800 https://yipko.com/posts/work/a2a/ A2A 把常见的 `noise-to-action` 生成路线改成了 `action-to-action`。模型会从历史动作编码出来的 latent 起点出发，去生成未来动作 latent，不再从纯随机噪声起步，因此推理更快，也更贴合机器人运动的连续性。 https://yipko.com/posts/work/rtc/ Sun, 24 May 2026 05:18:15 +0800 https://yipko.com/posts/work/rtc/ RTC 是一个纯推理阶段的方法。它不改训练、不重训 VLA；执行旧 action chunk 的同时生成新 chunk，把“已经注定会执行的那段动作”冻结下来，再把剩余部分当成 inpainting 问题补出来，从而缓解高延迟下的停顿、跳变和 mode switching。 https://yipko.com/posts/work/training-time-rtc/ Sun, 24 May 2026 05:18:15 +0800 https://yipko.com/posts/work/training-time-rtc/ Training-Time RTC 把 RTC 原本放在推理阶段做的 continuity 约束前移到了训练阶段。模型在训练时就显式见到“给定已知 action prefix，继续生成 action postfix”这个任务，因此推理时不再需要 RTC 那套额外的 inpainting guidance。 https://yipko.com/posts/work/pi0.7/ Sun, 24 May 2026 05:18:15 +0800 https://yipko.com/posts/work/pi0.7/ `pi0.7` 把 prompt 从一句任务文本扩展成“总任务 + 子任务语言 + 子目标图像 + episode metadata + control mode”的多模态上下文。这样一来，模型更容易从混合质量、混合来源的数据里学到可控、可组合、可迁移的行为。 https://yipko.com/posts/work/apt-mirror-tool-raspberry-pi/ Fri, 22 May 2026 00:00:00 +0800 https://yipko.com/posts/work/apt-mirror-tool-raspberry-pi/ 一个 APT 软件源配置工具的树莓派版本，支持 DEB822、传统格式和测速后切换。 https://yipko.com/links/ Thu, 10 Jul 2025 00:22:00 +0800 https://yipko.com/links/ <div class="friend"> <p><a target="_blank" href=https://yipko.com title=YK&#39;s&#32;Blog class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://yipko.com/img/me.jpg /> </div> <div class="frienddivright"> <div class="friendname">YK&#39;s Blog</div> <div class="friendinfo">A blog that records YK&#39;s life.</div> </div> </div> </a> <a target="_blank" href=https://pengshihan.com title=Blang&#39;s&#32;Blog class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://pengshihan.com/head0.png /> </div> <div class="frienddivright"> <div class="friendname">Blang&#39;s Blog</div> <div class="friendinfo">Blang&#39;s Blog</div> </div> </div> </a> <a target="_blank" href=https://junweiliang.me/ title=Prof.&#32;Junwei&#32;Liang class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://junweiliang.me/resources/me.jpeg /> </div> <div class="frienddivright"> <div class="friendname">Prof. Junwei Liang</div> <div class="friendinfo">The personal page of Prof. Junwei Liang</div> </div> </div> </a> <a target="_blank" href=https://precognition.team/ title=Precognition&#32;Lab class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://precognition.team/resources/precognition_lab.png /> </div> <div class="frienddivright"> <div class="friendname">Precognition Lab</div> <div class="friendinfo">Precognition Lab @ HKUST (Guangzhou)</div> </div> </div> </a> <a target="_blank" href=https://sds.cuhk.edu.cn/en/teacher/489/ title=Prof.&#32;SUN,&#32;Qilin class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://sds.cuhk.edu.cn/sites/sds.prod.dpsite03.cuhk.edu.cn/files/styles/w1200/public/2022-01/%E5%AD%99%E5%90%AF%E9%9C%96%E7%B2%BE%E4%BF%AE%20%282%29.jpg?itok&#61;nbbmM8qW /> </div> <div class="frienddivright"> <div class="friendname">Prof. SUN, Qilin</div> <div class="friendinfo">The personal page of Prof. SUN, Qilin</div> </div> </div> </a> <a target="_blank" href=https://jiaming-zhou.github.io/ title=Jiaming&#32;Zhou class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://jiaming-zhou.github.io/images/jiaming_hk.jpg /> </div> <div class="frienddivright"> <div class="friendname">Jiaming Zhou</div> <div class="friendinfo">The personal page of Jiaming Zhou</div> </div> </div> </a> <a target="_blank" href=https://axi404.top title=Axi&#39;s&#32;Blog class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://axi404.top/avatar/avatar.png /> </div> <div class="frienddivright"> <div class="friendname">Axi&#39;s Blog</div> <div class="friendinfo">一只可爱小猫</div> </div> </div> </a> <a target="_blank" href=https://hazintihoa.github.io/ title=Zhihao&#32;Tian class="friendurl"> <div class="frienddiv"> <div class="frienddivleft"> <img class="myfriend" src=https://github.com/hazintihoa.png /> </div> <div class="frienddivright"> <div class="friendname">Zhihao Tian</div> <div class="friendinfo">The personal page of Zhihao Tian</div> </div> </div> </a> </p> </div> <br/> <br/> <br/> <br/> <br/> <div style="font-size: 20px;" class="youlian">👉 The Format of Links Submitted</div> <div style="font-size: 16px;"> <table> <thead> <tr> <th></th> <th></th> </tr> </thead> <tbody> <tr> <td>Name：</td> <td>YK&rsquo;s Blog</td> </tr> <tr> <td>Site：</td> <td><a href="https://yipko.com">https://yipko.com</a></td> </tr> <tr> <td>Icon：</td> <td><a href="https://yipko.com/img/me.jpg">https://yipko.com/img/me.jpg</a></td> </tr> <tr> <td>Description：</td> <td>A blog that records YK&rsquo;s life.</td> </tr> </tbody> </table> </div> <br/> <div style="font-size: 20px;">👉Friendly Links Exchange Application Requirements</div> <blockquote> <p>Adhering to the principle of exchanging friend links, regularly updating articles, not having too many advertisements, and controlling the word count of personal descriptions to within 15 words.</p> https://yipko.com/posts/work/llama-cpp-cuda-server-compile/ Sat, 03 May 2025 18:04:20 +0800 https://yipko.com/posts/work/llama-cpp-cuda-server-compile/ 本指南记录了如何在 Ubuntu 22.04 上编译和部署 Llama.cpp]，启用 CUDA 加速和优化选项，并通过 `llama-server` 启动一个本地推理 API 服务（兼容 HuggingFace）。 https://yipko.com/posts/work/cuda-switcher-with-conda/ Tue, 15 Apr 2025 18:28:30 +0800 https://yipko.com/posts/work/cuda-switcher-with-conda/ Linux 系统上快速切换不同 CUDA 版本的小工具，支持 Conda 环境自动切换，兼容系统默认 CUDA 设置。文末附有 AI Agent 自动化配置教程。 https://yipko.com/posts/work/coppeliasim-rlbench-troubleshooting-guide/ Sun, 23 Feb 2025 18:59:00 +0800 https://yipko.com/posts/work/coppeliasim-rlbench-troubleshooting-guide/ A comprehensive guide to troubleshooting common issues encountered with CoppeliaSim and RLBench. https://yipko.com/posts/work/nvidia-gpu-ecc-disable-failure-solution/ Fri, 21 Feb 2025 20:09:00 +0800 https://yipko.com/posts/work/nvidia-gpu-ecc-disable-failure-solution/ 本文介绍了在 Nvidia Tesla L40 上无法关闭 ECC 功能或者驱动设置无法生效的解决方案，包括使用 nvidia-smi 命令、手动卸载 NVIDIA 内核模块以及重载 GPU 和驱动等步骤，方法适用于所有具备ECC功能的GPU。文末附有AI Agent自动化执行教程，可直接用于自动化部署。 https://yipko.com/posts/work/gpu-remote-desktop-research-setup/ Wed, 25 Sep 2024 18:42:00 +0800 https://yipko.com/posts/work/gpu-remote-desktop-research-setup/ 远程科研一晃一天，环境搭建履步维艰。笔者折腾了一下实验室服务器/工作站（headless机器）之间的远程桌面，由于要使用Isaac Sim等软件，需要让远程桌面跑在GPU上，故写一篇文章来记录一下实现流程，分享详细的手动配置步骤。本文底部还包含AI Agent自动化安装教程，可直接用于自动化部署。 https://yipko.com/about/ Thu, 21 Mar 2024 15:04:28 +0800 https://yipko.com/about/ <p>I am <strong>YE, Ke (叶柯)</strong> <span><button type="button" aria-label="Hear YE, Ke pronounced in Mandarin" title="叶柯 · Yè Kē" onclick="const audio = this.nextElementSibling; audio.currentTime = 0; audio.play().catch(() => { audio.hidden = false; audio.controls = true; });"><span aria-hidden="true">🔈</span></button><audio src="https://yipko.com/audio/ye-ke.mp3" preload="none" hidden aria-label="Mandarin pronunciation of 叶柯"></audio><noscript><a href="https://yipko.com/audio/ye-ke.mp3">Listen to the pronunciation</a></noscript></span>, a PhD student at <a href="https://www.hkust-gz.edu.cn/">HKUST(GZ)</a>, advised by <a href="https://junweiliang.me/">Prof. Junwei Liang</a> at the <a href="https://precognition.team/">Precognition Lab</a>.</p> <p><a class="pub-link" href="mailto:wiriest.listen-0a@icloud.com">📬 Email</a> <a class="pub-link" href="https://cv.yipko.com">📕 CV</a> <a class="pub-link" href="https://github.com/YipKo">👨🏻‍💻 GitHub</a></p> <h3 id="background">Background</h3> <table> <thead> <tr> <th style="text-align: left"></th> <th style="text-align: left"></th> </tr> </thead> <tbody> <tr> <td style="text-align: left">🎓 Education</td> <td style="text-align: left">M.Phil. in Artificial Intelligence, HKUST(GZ)<br>B.Eng. in Automation, Northeastern University</td> </tr> <tr> <td style="text-align: left">🔬 Previously</td> <td style="text-align: left">Research Assistant, CUHK-Shenzhen<br>Advisor: <a href="https://sds.cuhk.edu.cn/en/teacher/489">Prof. Qilin Sun</a></td> </tr> </tbody> </table> <h3 id="research">Research</h3> <p><strong>Embodied AI · Robot Learning · Multimodal Intelligence</strong></p> <p>I study how robots learn manipulation from human demonstrations and generalize to new tasks and noisy observations.</p> https://yipko.com/posts/work/the-malfunction-of-xrdp/ Sat, 24 Feb 2024 19:39:50 +0800 https://yipko.com/posts/work/the-malfunction-of-xrdp/ xrdp 连接黑屏、剪贴板不工作、文件传输异常、硬盘无法挂载、没有声音的解决方案。文末附有 AI Agent 自动化修复教程，可直接用于自动化执行。 https://yipko.com/posts/work/inference-bug-on-nvidia-gtx-1660/ Mon, 02 Oct 2023 00:04:45 +0800 https://yipko.com/posts/work/inference-bug-on-nvidia-gtx-1660/ NVIDIA GTX1600 系列显卡（包括1650 1660Ti 1660Super在内）在Pytorch推理时，输出会变为NaN，即全空张量 https://yipko.com/posts/work/jetson-nano-setting-up/ Sun, 05 Jun 2022 00:17:58 +0800 https://yipko.com/posts/work/jetson-nano-setting-up/ Jetson 环境搭建以及模型部署 https://yipko.com/posts/work/simulated-annealing/ Fri, 21 Jan 2022 00:27:25 +0800 https://yipko.com/posts/work/simulated-annealing/ 最优化问题中很常用的算法