Dyna Robotics 发了 Dyna-2，只靠看人类视频就能学会机器人操作，没碰过机器人数据也能越看越强，跟以前的做法很不一样。
Dyna Robotics 发布世界动作模型 Dyna-2，仅用 100 万小时人类视频预训练，从未接触过任何机器人轨迹数据。实验发现，人类视频数据量从 1000 小时增至 100 万小时，模型在未见过的机器人任务上表现持续提升，呈现跨四数量级的缩放规律。这表明具身差距更像适应问题而非知识问题，可能改写机器人预训练范式。
What surprised me most: Dyna-2 never trained on a single robot trajectory, yet the more human video ...
What surprised me most: Dyna-2 never trained on a single robot trajectory, yet the more human video it saw, the better it got at robot tasks. The embodiment gap looks less like a "knowledge problem" than an "adaptation problem." That distinction could rewrite robot pretraining. Dyna Robotics @DynaRobotics Today we are introducing Dyna-2, a world-action model pre-trained on one million hours of human video. At this scale, for the first time, we discovered several new scaling laws: • world-action models exhibit scaling law on human data across four orders of magnitude, from 1000 to 1,000,000 hours, • this human data scaling law implied a scaling law on never seen robot data, • both data and objective matter; world modeling and scaling on video data are essential for cross-embodiment scaling transfer to emerge 🧵 Your browser does not support the video tag. 🔗 View on Twitter 🔗 View Quoted Tweet 💬 0 🔄 0 ❤️ 0 👀 586 ⚡