柏克萊讓機器人靠網路影片學靈巧操作把簡體中文原文的重點轉成繁體中文讀者更容易閱讀的脈絡。核心在於 Berkeley’s Do as I Do pipeline tries to turn ordinary web videos into executable dexterous robot trajectories.
變化在哪裡
目前可核驗的事實包括：single-view RGB input, no depth information, dual UR3e arms with Sharpa Wave dexterous hands, 50 Hz execution, 500 verified trajectories, 20 action categories, retargeting success rising from 25% to 71%, and only about 5% of web videos usable after filtering。這些細節讓新聞不只停留在發布會口號。
為什麼重要
Embodied AI is constrained by manipulation data; usable internet video could multiply the available training signal. AI 新聞已經不能只看模型分數，還要看成本、供應、監管與真實工作流程。
接下來看什麼
Scaling from 500 lab trajectories to thousands of reliable real-world skills will determine whether the approach changes robotics data collection. 下一步要看的不是更多宣傳，而是第三方測試、客戶續用與長時間穩定性。
參考來源：UC Berkeley Do as I Do research material and project results、CocoLoop。