nature will 发布 500 亿参数 MoE 具身基础模型 Natura-1
The hardest thing to give a robot is a capable pair of hands, and Natura-1, from @naturewill is a strong attempt at it.
nature will 发布首个面向物理世界的通用基础模型 Natura-1，这是一个 500 亿参数的 MoE 模型，在 100 万小时第一人称人类经验数据上预训练，再用有限的机器人数据做后训练以适配灵巧操作。该模型可零样本跟随自然语言指令，推理空间约束、多物体关系和否定条件，并能自主执行长时程连续流程。Rohan Paul 指出，它同时处理了机器人常难以兼顾的三件事：抓取后的双手精细控制、无需额外训练的新任务，以及需要记住早期步骤并从错误中恢复的长序列。
查看原帖Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct…
展开完整原文
Introducing Natura-1, our first general foundation model for the physical world. It follows natural-language instructions zero-shot and turns them into action with dexterous hands. Natura-1 is a 50-billion-parameter mixture-of-experts (MoE) model pretrained on one million hours of egocentric human experience. On top of this foundation, we conduct efficient post-training on a limited amount of robot data to adapt pretrained world knowledge to dexterous manipulation, yielding a model that generalizes to a wide range of unseen scenarios. Natura-1 follows complex natural-language instructions zero-shot, reasoning through spatial constraints, multi-object relationships, and negative conditions. It can also autonomously carry out long and continuous workflows to perform long-horizon executions. Natura-1 provides early evidence that generalization—and even emergent behavior—can coexist with the complexity of dexterous hands. Further progress depends on the full stack: infrastructure for faster training and real-time inference, pretraining data at greater scale and coverage, and models that transfer to robots with less adaptation. We expect it to be a long-term research and engineering effort.
正文
当前显示原文，中文全文译文尚未就绪。
The hardest thing to give a robot is a capable pair of hands, and Natura-1, from @naturewill is a strong attempt at it.
Nature Will trained their new Natura-1 on 1M hours of egocentric human data
first-person video of people using their hands, then used that to control real robot hands
Robot demonstrations are slow to collect, so human footage is the faster route to scale, provided the skills survive the move from human hands to robot hands.
One model handles 3 things robots usually fail at together: fine 2-handed control after the grasp, new tasks without extra training, and long sequences where it has to remember earlier steps and recover from mistakes. Many robot systems show 1 of these in a narrow setting.