Dyna Robotics 推出 Dyna-2，一个用 100 万小时人类视频预训练的世界动作模型，证明机器人预测能力随人类视频数据量单调提升。该模型在未见过的客户现场实现 87% 生产通过率，远超 Dyna-1 的 46%。团队首次发现人类数据规模定律，并验证其可迁移至未见机器人数据。
Maybe robots don’t only need more robot data.
Humans already provide a gigantic record of interacting with the physical world. Why not learn from that ?
Dyna Robotics just proved robot prediction improves monotonically as human video pre-training reaches 1Mn hours.
They just launched Dyna-2, a world-action model pre-trained using 1 million hours of human video. The key breakthrough is a human-to-robot transfer scaling law.
More human video went in, better manipulation came out. And somehow, that improvement carried over to robots the model had never trained on.
At unseen customer sites, Dyna reports an 87% production pass rate for Dyna-2 versus 46% for Dyna-1 at identical post-training budgets.
So there is now an obvious next experiment: keep scaling.
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 🧵在 X 查看被引用的帖子
来源：Rohan Paul · x.com