Nature Will, a physical AI lab focused on building physical artificial general intelligence, announced on October 16, 2026, the launch of Natura-1, a 50-billion-parameter physical general foundation model. The model is designed for dexterous hand control and zero-shot instruction following in advanced robotics applications. Pre-trained on more than 1 million hours of egocentric human data, it connects human intent to physical action while adapting to changing task conditions and sustaining extended interactions.
Architecture and Training
Natura-1 is built as a mixture-of-experts model with 50 billion parameters and was trained jointly on language, images, video, and action. The system operates on a 100 millisecond control loop, which unifies language, vision, and high-degree-of-freedom action for language-directed behavior and physical control. This design allows the model to generalize across tasks while maintaining fine physical control for dexterous hands.
Dexterous Manipulation and Instruction Following
The model can follow complex natural language instructions and reason through spatial constraints, multi-object relationships, and negative conditions. It supports zero-shot instruction following, allowing it to perform tasks without task-specific fine-tuning. Nature Will said Natura-1 is its first physical foundation model capable of general-purpose dexterous manipulation using real hardware.
Real Hardware and Autonomous Workflows
Natura-1 autonomously executes sustained, continuous workflows and can handle long-horizon execution across extended interactions. By combining instruction following with physical control, the model brings the versatility of the human hand within reach of a system that translates human intent into action. Nature Will designed the model to scale to any application where fine or delicate manipulations are required as part of the creation process.
Company Vision and Research Focus
Ning Ding, Chief Executive Officer of Nature Will, said that general models capable of handling every task must exist and that those models must utilize dexterous hands. He described Natura-1 as a first step toward models that scale from human experience and connect language-directed behavior with physical control. The company has assembled a full-stack physical AI team with expertise spanning data acquisition, foundation model pretraining, infrastructure, simulation, evaluation, post-training, and reinforcement learning.
Data, Infrastructure, and Evaluation
Nature Will's research and development stack covers data acquisition and governance, foundation model pretraining, infrastructure, simulation and real-world evaluation, post-training, and reinforcement learning. The company is led by Founder and Chief Executive Officer Ning Ding and includes researchers, core contributors to flagship foundation models, international student supercomputing competition champions, and senior robotics hardware engineers. This structure supports the development of embodied agents that can achieve self-awareness and continuous learning within physically native environments.
Broader Implications for Physical AI
With rapid advancements in general AI and large language models, AI systems developed for physical applications represent the next frontier in technological progress. Natura-1 demonstrates that generalization and fine physical control can be learned together in a foundation model for dexterous hands. The model could support use cases in manufacturing, assembly, and other precision-driven tasks that require delicate manipulation.
The launch of Natura-1 highlights the growing effort to bring foundation model capabilities into physical environments and dexterous manipulation. By training on large scale egocentric human data and operating on real hardware, Nature Will aims to close the loop between perception, prediction, reasoning, and action. The announcement reflects continued progress toward physical artificial general intelligence that can turn human intent into reliable real-world behavior.