Nature Will, a Chinese startup working on what it calls general physical intelligence, released Natura-1 on October 13, describing it as its first general foundation model for the physical world. The model takes language and vision as input and drives 20-degree-of-freedom dexterous robot hands.
Natura-1 is a 50-billion-parameter mixture-of-experts model pretrained on more than one million hours of egocentric, first-person video of human activity. The company argues that robots cannot yet supply physical experience at that scale, because a robot can only collect data in settings where it already works, while people already handle tools and objects in almost every environment. After pretraining on human data, the team adapts the model to robot hands with a limited amount of robot manipulation data.
According to the company, the curated corpus amounts to one million hours of language-conditioned action trajectories, built from footage that was deduplicated, screened for sensor anomalies and annotated hierarchically from overall goals down to individual steps. The team recovers dense 3D hand poses from monocular and stereo video. It built its own data lakehouse holding tens of petabytes and an in-house distributed training engine that it says sustains over 50% model FLOPs utilization on hundreds of GPUs. At inference, a Rust orchestration layer and fused CUDA kernels keep a strict 100-millisecond loop from observation to action on physical hardware.
Nature Will says Natura-1 can follow complex natural-language instructions zero-shot, including spatial constraints, relationships between several objects and negative conditions, and can complete long, continuous workflows. Demonstrations include folding thin, deformable materials, writing calligraphy with a soft brush and switching tools across multi-stage tasks. The company notes that six long-horizon tasks used additional tuning, while all other results share the same weights.
On its own benchmark of twelve dexterous manipulation dimensions, run on dual-arm robots with dexterous hands, the company reports scores of 53.4 for the base model and 59.1 for an agentic setup. It says the two baselines it chose could not produce a score with dexterous hands, and scored 13.5 and 16.7 when tested on a gripper platform. These results are from the company's own evaluation and have not been independently reproduced.
The company describes Natura-1 as early evidence that a single model can learn dexterous skills from human experience, and says each part of its stack is designed to scale further.