ZhenFund-Backed Nature's Will Unveils First Physical General Foundation Model Natura-1

Today, NatureWill unveiled Natura-1, its first general-purpose foundation model for the physical world.

Natura-1 is a 50-billion-parameter Mixture-of-Experts (MoE) model, pre-trained on one million hours of first-person human experience data. It takes language and vision as inputs and can drive a 20-degree-of-freedom dexterous hand within a 100-millisecond control loop.

Natura-1 can zero-shot follow complex natural language instructions, generalize to unseen scenarios and tasks, and autonomously complete long, continuous workflows — such as folding thin, easily deformable materials, writing calligraphy with fluid brushstrokes, or switching tools mid-stream during multi-stage tasks.

Since NatureWill's founding in April 2026, ZhenFund has been there every step of the way, watching the team advance its exploration of general physical intelligence to today's Natura-1.

Founder Ning Ding entered Tsinghua University's Department of Computer Science and Technology for his PhD in 2018 and later stayed on as faculty. The team is a young group with strong research capabilities and extensive technical achievements.

NatureWill's goal is general physical intelligence: building foundation models that learn directly from human experience, master general dexterous manipulation, and generalize across the physical world.

To this end, NatureWill developed full-stack infrastructure in-house: a data lakehouse that efficiently manages dozens of petabytes of data. Its data pipeline covers the full workflow of cleaning, annotation, and distribution control. Self-developed low-level operators and a distributed training engine jointly optimize sample dispatch, cache prefetching, and GPU computation.

Natura-1 is the first model trained on this technology stack, with every component of the system designed for further scaling.

Real-World Performance

NatureWill evaluated Natura-1's capabilities in the real world. Its self-built benchmark spans twelve dimensions across three groups.

"Dexterous Control" covers pose control and contact, bimanual coordination, in-hand manipulation, and tool use. "Instruction Following" tests the model's reasoning toward goal achievement and its generalization to unfamiliar scenes and spatial layouts, tools and objects, actions, and task combinations. "Long-Horizon Execution" evaluates the model's planning, memory, and recovery abilities during extended workflows.

The entire benchmark was evaluated by a single base model, and this model had never seen these evaluation tasks, their scenes, objects, or action combinations during robot training.

Because there are no strong general-purpose dexterous manipulation models to serve as fair baselines, NatureWill selected π0.5 and GPT-6-Astra for comparison. But even after fine-tuning π0.5 on the same robot data and giving GPT-6-Astra 60 minutes of execution time, neither produced any scorable results on the same dexterous hand platform.

To enable a relatively fairer comparison, NatureWill used the ARX AC One bimanual gripper platform, which both models were more familiar with. π0.5 was fine-tuned on a gripper-manipulation version with the same real-world training data.

In the end, π0.5 and GPT-6-Astra scored 13.5% and 16.7% respectively on the gripper. Natura-1, on this 20-DOF dexterous hand, achieved composite scores of 53.4% (Base) and 59.1% (Agentic).

Natura-1 (Agentic) achieved the highest or tied-highest score across all 12 dimensions.

1.1 Long-Horizon Dexterous Tasks

With additional fine-tuning, Natura-1 demonstrates autonomous execution lasting several minutes without human intervention. Within the same workflow, it handles different materials and contact conditions, chains multiple steps, and continuously advances the task.

Natura-1 adjusts its grasp and motion based on continuously updated observations, executing fine manipulations with new tools, objects, and scenes. The high-degree-of-freedom bimanual setup and the model's generalization capabilities work together in actual action.

1.2 Zero-Shot Instruction Following

Language defines the global goal; real-world feedback guides action. Natura-1 can follow unfamiliar instructions in a zero-shot manner. New instructions often simultaneously bring unfamiliar spatial relationships, object combinations, and manipulation objectives.

The model needs to understand constraints in the instructions, map them to the scene before it, and then drive the dexterous hands to complete corresponding actions.

Faced with actions, objects, and scenes never seen in training data, Natura-1 organizes its bimanual execution steadily. When objects move, contact conditions change, or execution is disrupted, it adjusts its manipulation to faithfully complete the user's instruction.

1.3 The Starting Point of Emergence

In everyday interactions, Natura-1 has also demonstrated some behaviors beyond imagination.

During a playing-card grasping test, the cards were extremely thin. Drawing on physical intuition, the model invented an entirely new grasping approach: it pushed the card over the edge of the box, creating an overhanging portion for fingers to slip under, then coordinated both hands to lift it.

When asked to retrieve an out-of-reach ball of yarn, the model used a rod already threaded through a basket to pull the basket closer before taking the object.

In another task below, Natura-1 completed something never seen in training: pressing the spacebar on a laptop to wake it up.

The instructions never specified these intermediate steps, and real-world demonstrations never provided such operation paths. The model organized new solutions for the goal based on the physical regularities before it.

Such emergent behavior may be an early sign of more general physical intelligence: large-scale human experience may allow the model to learn implicit understanding of physical mechanisms and form novel action strategies in unfamiliar situations.

Scaling, Scaling, Scaling

The value of human experience lies both in scale and in breadth.

Natura-1's one million hours of first-person data covers the vast majority of daily life and production scenarios. Internet video is richly diverse, with accumulated experience continuously broadening the world the model has seen.

In NatureWill's data pipeline, over 20 low-level operators continuously execute deduplication, quality screening, and distribution control. They recover hand and full-body motion from monocular and binocular video, decompose global goals into subtasks and atomic steps, and form video-language-action multi-level trajectories.

A data lakehouse of dozens of petabytes supports the entire process. The pipeline can produce 350,000 hours of training data per week, scaling linearly with compute resources, allowing real-world experience to continuously flow into the training process.

Scalable infrastructure must satisfy two requirements simultaneously: during training, enabling the model to continuously consume multi-modal trajectories at high throughput; during inference, responding to the real world within 100 milliseconds.

The former requires sustained efficiency in distributed computation, while the latter demands precise control over latency across the entire execution path.

To this end, NatureWill rebuilt the training and inference technology stack through hardware-software co-design, from data organization and distributed execution to GPU computation.

On the training side, sample dispatch and cache prefetching reduce padding waste and idle waiting. Under measured production configurations, model FLOPs utilization (MFU) exceeds 50%, with GPU utilization reaching 98%. On the inference side, Rust coordinates scheduling and sessions, reusing resident GPU state caches.

Self-developed attention operators achieved approximately 1,200 TFLOP/s throughput in forward and backward benchmark tests on average, exceeding the performance of a tuned FlexAttention baseline by 6x. Ultimately, end-to-end inference latency from observation to action is strictly controlled within 100 milliseconds.

Natura-1's architecture revolves around one question: how to extract more intelligence from every piece of experience?

Language, images, video, and action are placed into the same 50-billion-parameter MoE model. Instruction understanding, world modeling, and action generation are jointly learned by the same sequence model.

In its architectural design, Natura-1 follows the first principle of scaling potential. Its model architecture is almost that of a standard end-to-end MoE Transformer, with all inductive biases restricted to the organization of multi-modal data within sequences and the causal relationships of the attention mechanism.

This allows the model's optimization potential for training and inference efficiency to rival frontier LLMs and video generation models.

NatureWill treats video frames and action poses at the same moment as two different projections of the same physical state. Though high-quality action labels are scarce, video runs through every trajectory.

Natura-1 uses past video as memory and future frames as world-modeling targets, learning video and action through joint denoising and flow matching — gaining supervision beyond action labels from abundant video.

Predicting the next step also requires understanding human intent. Fitting the joint distribution of vision and action poses does not necessarily produce this understanding.

NatureWill uses fine-grained language supervision so that different intentions correspond to different video and action predictions, making "what to do next" entirely dependent on the goal proposed by the human. A small amount of robot data is then used for post-training to adapt capabilities gained from human experience to specific robot platforms.

NatureWill integrates data pipeline, model architecture, and infrastructure design, aiming to transform more human experience into more precise action prediction.

As training progressed from approximately 44,000 hours of experience to one million hours, prediction error decreased noticeably on both human data held out from training and robot validation data.

These curves provide empirical validation for an integrated scaling strategy. With continued investment in data and compute, NatureWill aims to expand the scale and breadth of its capabilities and advance toward general-purpose physical foundation models.

Natura-1 provides initial confirmation that dexterous manipulation execution and open-world generalization can converge within a single unified general system.

This means human-level fine manipulation capability need not be developed separately for each scenario, but can emerge from a foundation model as human experience scales.

Humans shaped this world with their hands. Now, machines too are beginning to learn from that experience, understand human intent, and change the world with dexterous hands.

Click "Read More" for the full technical blog: https://www.naturewill.com/blog/natura-1