文章摘要
自然意志研究团队推出面向物理世界智能的大模型Natura-1，该模型为500亿参数MoE结构，基于超1亿小时第一人称人类操作数据训练，可低延迟控制20自由度灵巧手，解决了传统机器人灵巧性与通用性难兼顾的问题，能零样本完成包饺子、贴屏幕膜等未训练的复杂任务，测试表现突出，团队后续将进一步优化其泛化性等能力。

Making dumplings is a simple task for humans—using hands to hold wrappers, add filling, fold, and seal, all while adjusting to the soft, flexible material and timing movements precisely. But for robots, this kind of delicate, adaptive handwork has long been a major challenge. Now, a new system has achieved this feat: a pair of robotic hands successfully completing the entire dumpling-making process in a single, unedited 72-second video.

This system is powered by Natura-1, a large model from a research team focused on physical world intelligence. It’s a 50-billion-parameter Mixture of Experts (MoE) model, trained on over 100 million hours of first-person human operation data. The model uses language and visual inputs to control a 20-degree-of-freedom robotic hand, with a control loop latency of just 100 milliseconds.

While the dumpling task shows the model can perform specific actions, Natura-1 aims to solve a much harder problem: can a robot directly handle tasks it has never practiced before?

Dexterity and Generality: The Long-Standing Trade-Off

The balance between dexterity and flexibility has long limited robotic manipulation. Human hands have dozens of degrees of freedom, and working with soft, deformable objects requires constant real-time adjustment. Parallel grippers simplify control but cannot replicate the hand’s ability to support, twist, or coordinate two hands in complex tasks. Past robotic systems often stuck to one side of this divide: either general strategies for grippers or task-specific operating systems.

For general-purpose systems, a key barrier is data: generalizing to new tasks requires large, diverse training data, but collecting data from real robots is expensive, and teleoperating dexterous hands is slow and difficult. This creates a vicious cycle: robots can only collect data where they already work, and can only work where they have pre-existing data. To break this cycle, many teams are turning to a ready-made, scalable data source: humans.

Earlier this year, a research project used 20,854 hours of first-person human operation videos to train a model, finding a log-linear relationship between data scale and validation loss. When deployed on a 22-degree-of-freedom hand, its success rate was 54% higher than baselines without human data. Later in the year, another project pushed this scale to over 100 million hours of first-person human videos, with no robot data used during pre-training. As the data scale increased from 1,000 to 100 million hours, the model’s average score on 14 robotic tasks rose from 20% to 53%.

But both approaches still required additional adaptation: one needed small amounts of human-robot alignment data, the other used up to 10 hours of robot data per task for post-training. The next critical question is: can human experience be directly transferred to dexterous hands without per-task adaptation?

Natura-1: A Model for Two Core Capabilities

Natura-1 is the first physical world general-purpose model from its development team, founded in early 2026. Its core goal is to build a general physical intelligence system that can complete perception, prediction, reasoning, and action loops in the real world. The team’s two key judgments: general physical world models are possible, and first-person human data can be scaled. Following the methodology of large language models, they developed the model by scaling model size and data together, starting small to validate the architecture before expanding.

Unlike systems that split tasks like "understanding instructions", "predicting the world", and "generating actions" into separate modules, Natura-1 is a single MoE Transformer, jointly trained on language, images, videos, and actions. It focuses on two main capabilities:

  • Dexterous manipulation: handling multi-finger coordination, object posture adjustment, and complex contact tasks;
  • Zero-shot instruction following: completing unseen tasks based on natural language commands, including spatial constraints, multi-object relationships, and negative conditions.

What Natura-1 Can Achieve

The team tested Natura-1 in three key areas: long-horizon tasks, zero-shot testing, and emergent abilities.

For long tasks, in addition to dumpling-making, the system demonstrated five other tasks: making tea, assembling a yogurt bowl, packing packages, applying screen protectors, and writing calligraphy. Each task lasted over 3 minutes, with no human intervention, requiring handling liquids, particles, soft tools, and flexible parts in sequence. For example, applying a screen protector involved wiping the screen, aligning the frame, pressing the phone, removing the release film, and squeezing out air bubbles. Packing packages required adjusting grip for different items, closing the box, applying tape, and scanning barcodes. These tasks require real-time adjustments as objects change shape and contact conditions, not just fixed trajectories.

But these are task-specific demonstrations; the real test is zero-shot performance. Zero-shot here means the model is not trained on test tasks, and completes them directly after pre-training and post-training. The team built a benchmark with 12 capability dimensions, where test tasks and scenes were not in post-training data. For example, a task required putting an apple in a bowl, a bottle on a tray, and carrying the tray—combining separate actions from training data. On this benchmark, other models scored poorly on dexterous hands, while Natura-1 scored significantly higher, with faster execution times. The team also tested open-ended commands, like placing items in a basket with different hand order requirements, which Natura-1 completed correctly. The "unseen" here refers to post-training data; human pre-training data may have similar actions, like pressing a keyboard key. The key is whether human experience can be directly used on real robots without corresponding robot data.

The team also observed emergent abilities: unexpected behaviors not in test plans. For example, reaching for a yarn basket on a high shelf, the model used a rod to pull the basket close, then took the yarn. When it couldn’t pick a playing card from the edge, it pushed it to the box edge and picked it up with two hands. These steps had no corresponding robot demonstration data. As data scale increased, the model adjusted and re-tried when tasks failed, showing recovery abilities, likely learned from large-scale human data that includes adjustment and retry experiences.

Where Natura-1’s Capabilities Come From

The team says these abilities don’t happen by chance; they require intentional design of data, model, and infrastructure.

For data: Natura-1 uses 100 million hours of filtered first-person human experience, with 61.7 million segments, 318,000 tasks, and 14,000 scenes, mostly collected in controlled settings plus internet videos. The team prioritized scarce task categories, processing raw videos into video-language-action trajectories with a custom pipeline that has lower joint position error than open-source methods. Choosing human data and dexterous hands is intentional: tools and tasks are designed for humans, so robots can adapt to existing setups instead of changing them, and dexterous hands preserve more hand movement information than grippers, reducing data loss.

For infrastructure: the team’s philosophy is scalability. Data side: manages large volumes of data, with a pipeline that produces hundreds of thousands of hours weekly, scaling with compute. Training side: custom framework uses hundreds of GPUs with high utilization, and custom attention operators are far faster than optimized versions. Inference side: a Rust-based service framework reduces end-to-end latency to 100ms, enabling real-time control.

For the model: Natura-1 is an auto-regressive model that sequences language, images, videos, and actions, with a MoE Transformer backbone. It treats videos and actions as two projections of the same physical state, using video as memory to predict future states, training with flow matching to connect video and action. Language is used to guide the model to follow instructions, preventing it from ignoring commands. Scaling curves show that as training data increased, action prediction error on human and robot data dropped significantly, proving human experience translates to robots.

Conclusion

Natura-1 shows that dexterity and generality can coexist. The team’s next goal is to improve generalization across tasks and scenes, making the model more reliable for real-world complex tasks. They plan to add tactile data, train on larger datasets, reduce adaptation costs, and run on cheaper hardware, a long-term effort.

The approach of using human first-person data is gaining industry consensus, with other teams expanding to even larger data scales. The focus now shifts to how well human video is converted to actions and language, how much adaptation is needed, and which hardware is used. Natura-1 takes the challenging path of using high-degree-of-freedom hands to handle human experience, with one model for unseen tasks. Its success depends on hand stability and cost, and the scalability of human experience data. In a world designed for humans, learning from them is a promising direction, and Natura-1 provides an early test of this path.

以上内容不代表本平台立场，仅供读者参考