Natura-1: A 50B Model Teaching Dexterous Robot Hands From 1M Hours of Human Video

Most humanoid and manipulation demos are built task by task: collect teleoperated demonstrations, train a policy, polish a video. Nature Will is arguing for the opposite. On October 13, 2026 it introduced Natura-1, what it calls "our first general foundation model for the physical world," a 50-billion-parameter mixture-of-experts (MoE) model pretrained on more than one million hours of egocentric human video that follows plain-language instructions zero-shot and carries them out with 20-DoF dexterous hands.

Everything below comes from Nature Will's own announcement and research page. The company's demo videos and benchmark have not been independently replicated, and we could not view the videos from our research environment, so we describe only what the text states.

TL;DR: what Nature Will says Natura-1 is

Watch: Nature Will's launch video

Why dexterous hands are the hard part

Nature Will's framing is that the physical world was built around the human hand, so "no machine can be truly general without being dexterous." The difficulty is real: contact dynamics leave little room for error, and the action space grows exponentially with the number of hand joints. The company says prior work has mostly split into general-purpose policies for simple grippers or task-specific systems for dexterous manipulation.

That is the gap Natura-1 targets. It is a bet that one large model can span both: general in what it can be asked to do, dexterous in how it does it. For context on where hand hardware is heading, see our coverage of 1X NEO's 25-DoF hands and physical API.

Where does the data come from?

The central claim is about data. Robot data collection, Nature Will argues, is bounded by robot capability: a robot can only gather data where it already works, and it only works where data exists. Humans already operate everywhere, so the company trains on human experience instead.

The announcement gives a detailed account of the pipeline:

• Scale. The data infrastructure manages tens of petabytes in a unified lakehouse, with throughput of hundreds of thousands of hours of egocentric video per week.
• Cleaning. It deduplicates at video and clip level across sources, screens for visual and sensor anomalies, and enforces per-task quotas to keep the distribution close to real human activity.
• Labels. It recovers dense 3D hand poses from monocular and stereo video and annotates each activity hierarchically, from overall goal to subtasks to individual steps.
• Mix. The dataset is "primarily built through controlled data acquisition, complemented by large-scale internet ego videos." It includes monocular, stereo and multi-camera recordings plus wrist-mounted views.

Nature Will is candid about the mapping problem: "no robot hand maps exactly onto the human hand," so it relies on anthropomorphic dexterous hands as the closest correspondence available. How much is lost in that translation is the central open question for any human-video approach, and the same question hangs over other efforts such as Xiaomi Robotics-1's 100,000-hour UMI-based VLA and the human-camera data collection described in India.

The training and inference stack

The company says it built the stack in-house. Training uses a distributed engine co-designed with the 50B MoE, with batches streaming from object storage rather than a shared file system and custom attention kernels that it says considerably outperform baselines such as FlexAttention on its workloads. It reports sustaining over 50% Model FLOPs Utilization (MFU) across long jobs on hundreds of GPUs.

At inference, a Rust orchestrator overlaps observation streaming, computation and action transfer, and fused CUDA kernels trim critical-path latency to hold a strict 100-millisecond observation-to-action loop on physical hardware. For a 50B-parameter model, that latency target is the engineering claim most worth scrutinizing: a mixture-of-experts design activates only part of the network per step, but Nature Will does not state how many parameters are active or what hardware serves the loop.

The benchmark: what it does and does not show

Nature Will evaluated on physical dual-arm robots (ROKAE AR5 with Wuji hands) using a benchmark that measures dexterous manipulation along twelve capability dimensions. It chose two baselines: pi0.5, an established vision-language-action model, and GPT-6-Astra, a frontier multimodal model used directly as a coding-based policy.

The results need careful reading:

• On the dexterous-hand platform, the company says neither baseline produces a scored result. pi0.5 fails even after fine-tuning on its dexterous-hand data, and GPT-6-Astra fails to score within the 60-minute per-task limit.
• On a separate gripper platform (ARX AC One), pi0.5 (fine-tuned on matching gripper demonstrations) and GPT-6-Astra score 13.5 and 16.7, against Natura-1's 53.4 (Base) and 59.1 (Agentic) with dexterous hands.

The company itself notes that no publicly recognized dexterous-hand foundation model exists, so a like-for-like baseline is unavailable. That makes the comparison suggestive, not definitive: different hardware, a baseline used as a code-writing policy rather than a purpose-built controller, and a benchmark designed by the model's authors. Treat the headline numbers as vendor-reported until someone else runs the same tasks. If you want a refresher on how to judge such claims, our piece on benchmark claims versus verified results lays out the checklist.

Nature Will says none of the four showcased benchmark tasks appear in post-training data, nor do their scenes, objects or action combinations, and that all evaluations use a single fixed training set with no task-specific tuning. It shows each execution next to the closest retrieved training demonstrations so readers can judge the distance. One example: a task asking the robot to place an apple in a bowl, seat a bottle on a tray and carry the loaded tray to a placemat, where the training data had separate picking and carrying examples.

Long-horizon workflows, with an asterisk

The long-horizon section covers six single-take workflows: making wontons, packing a shipping box, serving tea, Chinese calligraphy, applying a screen protector and making a yogurt bowl. Several last multiple minutes, the longest listed at about three minutes.

The footnote is important. These six demos, marked with a dagger, "use additional tuning on top of Natura-1," while "all other results share the same weights." So the most impressive-looking sustained tasks are not the zero-shot ones. The company frames them as showing "the depth of a trained skill," as opposed to the benchmark and open-ended trials, which test physical generalization.

Its description of the wonton task gives a sense of the claimed behavior: the hands must support a thin sheet and filling while leaving the edges free, separate one wrapper from a stack, add filling with a long-handled tool, then fold and seal, adjusting pressure so the wrapper does not tear. We cannot verify that from text alone.

Zero-shot instruction following

For open-ended instruction following, Nature Will says the trials were recorded during researchers' day-to-day interactions, with unscripted and often deliberately extreme prompts and no reference trajectories. The demos include ruling out bottle colors, tracking a moving lemon, passing a tissue and stacking cups with an orange and a tissue.

The clearest example of language control is a three-way test with a toothbrush case, a water bottle and a bin. Asked to move both items at once, to use the left hand first, or to use the right hand first, the robot reportedly followed the sequencing constraint each time. Nature Will also publishes a nearly 30-minute single-take recording of everyday interactions.

The caveat: "recorded during day-to-day interactions" does not tell us the success rate, the number of failures not shown, or how many attempts each clip took. A selected video is a floor on what the system can do, not a measure of how often it does it.

Emergent behaviors: the most interesting claim

The most striking section describes behaviors the company says it did not plan to test. When a grasp fails or a path is blocked, Natura-1 reportedly finds alternatives: sliding a card over the edge of a box to create a graspable overhang, using a rod to pull an out-of-reach basket closer, or rotating the wrist mid-air to steady a toppling stack. In one case, it grasps the near end of a rod threaded through a basket's handles, pulls the basket into reach, and uses the other hand to retrieve the yarn.

Nature Will reads this as "an early sign of more general physical intelligence," suggesting that diverse human experience gives the model an implicit grasp of physical mechanics. That is an interpretation. The same behavior could result from strong generalization over tool-use patterns in the human video, and a handful of curated clips cannot distinguish the two. It is still a result worth testing with held-out, adversarial setups.

How it compares to other physical-AI efforts

Natura-1 lands in a crowded field with different philosophies:

Yann LeCun has argued that scaling language models alone will not solve physical agency, a view we covered in his Moravec paradox comments. Natura-1 is one answer: build the physical foundation model from human experience rather than from text.

What to check before believing the headline

1. Independent replication. Has anyone outside Nature Will run the benchmark or the zero-shot tasks?
2. Success rates. What fraction of attempts succeed, including failures and resets?
3. Hardware dependence. Does the model transfer to other hands, or only to the Wuji hands it was tuned for?
4. Baseline fairness. Using GPT-6-Astra as a coding policy is one way to test a frontier model; a purpose-trained policy might do better.
5. Active parameters and serving cost. What does it take to hold a 100 ms loop?
6. Data provenance. How much of the 1M hours is controlled collection versus internet video, and under what consent and licensing terms? The announcement does not say.

What it means for builders

If you build robotics software, the practical takeaways are modest but real. First, the data argument is the one to watch: if human egocentric video really can pretrain a dexterous policy, data collection costs for robot teams fall sharply. Second, the gap between the zero-shot results and the tuned long-horizon demos suggests that task-specific tuning still matters for sustained, multi-minute work. Third, no weights or API were announced, so there is nothing to integrate yet.

What we don't know

• Whether weights, an API or a license will be released
• The active parameter count and the hardware used for inference
• Success rates and failure modes across the benchmark
• How the dataset was sourced, consented and licensed
• Whether the result holds on hands other than those tested

Nature Will says 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. It expects the work to be "a long-term research and engineering effort." That is a fair description. Natura-1 is best read as a serious data-and-systems bet with promising but self-reported results, not yet a product or an independently verified milestone.

Details come from Nature Will's October 13, 2026 announcement and research page. The page text we reviewed was cut off partway through the data section, so later sections on infrastructure and model design may contain details we have not covered.

Related reading

• Reka Rho-1: a 19B model that outputs robot actions
• 1X NEO's 25-DoF hands and physical API
• Xiaomi Robotics-1 and 100,000 hours of UMI data
• Gemini Robotics 2 and whole-body intelligence
• NVIDIA Cosmos 3: open physical-AI world model guide
• Workers wearing cameras to train humanoid robots
• Yann LeCun on LLMs, physical agents and Moravec's paradox
• Figure Helix 02 and collaborative humanoids