<?xml version="1.0" encoding="UTF-8"?><rss version="2.0"><channel><title>Theodore Ouyang — Portfolio</title><description>Bilingual research notes, essays, and reflections.</description><link>https://research-notes-130u.jiligualapiqiu.chatgpt.site/</link><item><title>π₀: How a Robot Turns Understanding into Continuous Action</title><link>https://research-notes-130u.jiligualapiqiu.chatgpt.site/pi0/</link><guid isPermaLink="true">https://research-notes-130u.jiligualapiqiu.chatgpt.site/pi0/</guid><description>A layered account of the VLM backbone, continuous action expert, and Conditional Flow Matching that turn vision, language, and proprioception into executable action chunks.</description><pubDate>Sun, 02 Aug 2026 04:00:00 GMT</pubDate><language>zh-CN, en</language><category>Robotics</category><category>Embodied AI</category><category>Vision-Language-Action</category><category>Flow Matching</category></item><item><title>From Vision and Instructions to Robot Actions: Two Emerging Paths</title><link>https://research-notes-130u.jiligualapiqiu.chatgpt.site/from-vision-and-instructions-to-robot-actions/</link><guid isPermaLink="true">https://research-notes-130u.jiligualapiqiu.chatgpt.site/from-vision-and-instructions-to-robot-actions/</guid><description>A practical taxonomy for VLM-to-VLA and video/world-model-to-action systems, organized by training targets, internal representations, and inference-time computation.</description><pubDate>Wed, 29 Jul 2026 04:00:00 GMT</pubDate><language>zh-CN, en</language><category>Robotics</category><category>Embodied AI</category><category>Vision-Language-Action</category><category>World-Action Models</category></item><item><title>Manifold and the Overseas World-Model Landscape</title><link>https://research-notes-130u.jiligualapiqiu.chatgpt.site/manifold-world-model-landscape/</link><guid isPermaLink="true">https://research-notes-130u.jiligualapiqiu.chatgpt.site/manifold-world-model-landscape/</guid><description>A conclusion-first desk research case study that narrows 23 leads into five risk paths while preserving 141 sources, 111 atomic claims, and the full red-team trail.</description><pubDate>Tue, 14 Jul 2026 04:00:00 GMT</pubDate><language>zh-CN, en</language><category>World Models</category><category>Embodied AI</category><category>Competitive Intelligence</category><category>Research Methodology</category><category>Agentic Research</category></item></channel></rss>