Natura-1 Brings a 50B Model to Dexterous Robot Hands
Nature Will claims one model can handle unfamiliar instructions and objects, but tuned demonstrations and hardware-dependent benchmarks leave important questions unresolved.
Listen
AI narration
11:27
0:00 / 11:27
AI SummaryGenerated from this article
Nature Will unveiled Natura-1, a 50-billion-parameter model designed to control dexterous robot hands by following language instructions and visual observations without requiring separate training for each task. The company claims the system can handle unfamiliar manipulation tasks across different objects and situations within a 100-millisecond observation-to-action loop.
Nature Will's evidence consists of company demonstrations and benchmarks on its own hardware, with six long-horizon tasks receiving additional tuning beyond the base model. The company reports scores of 53.4 for Natura-1 Base and 59.1 for an Agentic variant, but has not released model weights, API access, or code for independent verification. Outside evaluation remains unavailable to test the claimed generalization capabilities.
Nature Will says Natura-1 can carry out unfamiliar manipulation tasks with dexterous robot hands using shared model weights, without a separately trained policy for every task. The language-vision-action system is designed to turn instructions and camera observations into physical behavior across different objects and situations.
In its October 13 announcement, the company describes a 50-billion-parameter mixture-of-experts model pretrained on more than one million hours of first-person human experience. Natura-1 controls 20-degree-of-freedom dexterous hands on dual-arm robots within a reported 100-millisecond observation-to-action loop.
So far, the evidence consists of Nature Will’s demonstrations and benchmark results, with no independently reproduced evaluation. Six long-horizon demonstrations received additional task-specific tuning. The announcement also does not disclose public weights, API access, code or commercial availability.
Natura-1 presents an ambitious approach to general dexterous manipulation. Outside builders cannot yet inspect and test it as a robotics foundation model.
Shared Weights Are the Central Claim
A task-specific robot policy can learn a particular movement sequence under particular conditions. Nature Will proposes a reusable manipulation policy: a shared model that interprets language, observes the scene and generates actions appropriate to the current task. The company positions Natura-1 as jointly handling perception, instruction following and dexterous control.
The output is physical action, including how the hands manipulate objects while the scene changes. Asking an LLM to describe how a robot should perform a task is a different problem.
Nature Will’s Natura-1 demonstration page includes examples designed to illustrate that distinction. In one, the robot is instructed to place an apple in a bowl, position a bottle on a tray and carry the loaded tray to a specified placemat. The company contrasts that execution with training examples showing component skills, such as picking up fruit and carrying an already loaded tray.
Another example combines extracting a tissue, wiping an area beside a cup and disposing of the tissue. Both examples raise the same question: can a model assemble familiar capabilities into a new sequence while respecting the objects and spatial relationships named in the instruction?
Nature Will says the illustrated benchmark tasks, scenes, objects and action combinations were absent from Natura-1’s robot post-training data. That is a meaningful holdout claim, though it does not establish that the model has never encountered related activities anywhere in its training history.
Work with Zeniteq
Let’s work together
We’re open to thoughtful collaborations with teams building in AI. Explore the ways we can work together.
The million-hour human-video corpus complicates any absolute claim of novelty. These examples are best read as company-reported generalization beyond the robot adaptation set, not proof of entirely unseen physical concepts.
Human Video Provides Scale, but Robots Still Need Adaptation
Collecting robot experience is constrained by what robots can already do. Nature Will’s approach draws on human activity as a much broader source of manipulation examples.
According to the company, Natura-1’s pretraining pipeline processes egocentric video, meaning footage recorded from a person’s viewpoint. It recovers dense three-dimensional hand poses from monocular and stereo recordings, then annotates activities at several levels, from overall goals to subtasks and individual steps. The resulting training material is more structured than a collection of unlabeled videos, with the intention of connecting what a person sees, what their hands do and what the activity is meant to accomplish.
Nature Will describes a 50B mixture-of-experts architecture developed to combine visual world modeling and action generation. In general, mixture-of-experts models route computation through selected model components rather than necessarily activating every parameter for every input. The headline parameter count alone therefore does not establish inference cost.
Human hands and robot hands differ in geometry, movement limits and interaction with objects. Watching one is not equivalent to controlling the other. Nature Will says it follows human-data pretraining with a limited amount of robot manipulation data to adapt the model to physical hardware.
Natura-1 is thus not presented as a system that learns exclusively from human video and then controls a robot without robot training. The claim is that large-scale human experience provides a useful foundation, reducing how much robot-specific data is needed afterward.
The supplied evidence does not quantify that reduction against a matched alternative. Demonstrating data efficiency would require comparisons that hold the robot, tasks and evaluation conditions constant while varying the training approach.
A 100-Millisecond Loop Is Not a Reliability Result
Nature Will reports that Natura-1 connects language and vision to 20-DoF dexterous hands within a 100-millisecond observation-to-action loop, corresponding to roughly ten such updates per second.
A Rust orchestration layer overlaps observation streaming, computation and action transfer, while fused CUDA kernels reduce latency, according to the company. These implementation details are relevant to making a large model usable on physical hardware. The stated timing describes the reported observation-to-action loop, however, not the operating frequency of every joint controller or motor across the robot’s control hierarchy.
Dexterous hands add another layer of difficulty. Additional controllable movements can support more varied grasps and object manipulation, but they also introduce more choices about hand posture and contact. A successful action must account for how an object behaves as the fingers move as well as where the arm should go.
Alongside instruction following, Nature Will claims recovery behavior and improvised tool use. Those capabilities could be valuable in real manipulation, which rarely proceeds exactly as planned.
Response speed and selected recovery examples do not establish deployment reliability. An outside evaluation would need to measure repeated attempts, failure conditions and behavior when recovery is unsuccessful. Neither the model’s size nor its reported control-loop latency supplies that evidence by itself.
Six Long-Horizon Demonstrations Used Extra Tuning
Nature Will showcases sustained activities including folding deformable materials, brush calligraphy and switching tools during multi-stage tasks. Its page also states that six long-horizon tasks used additional tuning on top of Natura-1, while the other reported results share the same weights.
This is the clearest qualification in the release. Those tuned demonstrations should be assessed separately from the base model’s claimed zero-shot capabilities.
Task-specific tuning is not inherently a weakness. Builders may reasonably adapt a general foundation model to a workflow that matters commercially. Successful specialized adaptation, though, provides different evidence from executing a new workflow without that adaptation.
The disclosure leaves two propositions to assess. Can shared base weights handle unfamiliar instructions, objects and combinations of skills? And can the foundation support longer, more specialized workflows after additional training?
Nature Will provides first-party examples for both. The strongest extended demonstrations do not, on their own, establish that the untuned base model can perform arbitrary multi-minute tasks.
The Benchmark Changes Both the Model and the Hardware
Nature Will reports benchmark scores of 53.4 for Natura-1 Base and 59.1 for an Agentic variant on its dexterous-hand platform. These are company-reported scores, not independently verified success rates.
The comparison has a hardware limitation. Nature Will says it evaluated π0.5 and a coding-based multimodal-model policy on its ROKAE AR5 dual-arm platform with Wuji hands, but neither baseline produced a scored result. It then evaluated the baselines using an ARX AC One dual-arm platform with grippers, reporting scores of 13.5 and 16.7 respectively.
According to the company, π0.5 received gripper demonstrations covering the same general tasks as Natura-1’s dexterous-hand training set. That improves the relevance of the comparison without making the systems equivalent.
Changing the end-effector changes which grasps and interactions are physically available. The reported score gap cannot cleanly isolate the advantage of Natura-1’s model architecture, training corpus or policy; it reflects comparisons between different combinations of software and hardware.
The baselines’ failure to score on the dexterous setup is informative about those implementations under Nature Will’s conditions. It does not establish that those model families cannot support dexterous manipulation. Independent testing on matched hardware would provide substantially stronger evidence than the current cross-platform comparison.
Builders Still Cannot Reproduce the Release
For developers, the immediate limitation is access. The announcement does not disclose downloadable model weights, a public inference API, released code or commercial access terms.
Outside teams therefore cannot test the claimed generalization, inspect the implementation or measure behavior on their own robot setups. RobotWorld’s October 13 listing corroborates the announcement’s timing, but repeats release information rather than independently validating performance.
Natura-1’s most important potential contribution is a shared dexterous policy that reuses human experience across unfamiliar instructions and objects. The scale of its training and the complexity of its hands make that proposition worth examining.
Repeatable evaluations would strengthen the case: matched hardware comparisons, clearly separated tuned and untuned results, repeated-trial outcomes and outside testing of unfamiliar tasks. Another successful video alone would not resolve those questions. Until such evaluation becomes possible, Natura-1 remains a consequential first-party demonstration of a general-control approach, not a verified foundation other builders can yet rely on.
