Nature Will launches 50-billion-parameter Natura-1 model for dexterous robot hands

Nature Will, a Chinese AI startup, announced the release of Natura-1 on October 13, describing it as its first general foundation model for the physical world. The 50-billion-parameter mixture-of-experts model was pretrained on more than one million hours of first-person human video and then fine-tuned with limited robot manipulation data to control 20-degree-of-freedom dexterous robot hands. Details are in the company’s release on Pandaily.

The firm says the curated corpus contains one million hours of language-conditioned action trajectories, screened for sensor anomalies and annotated from overall goals to individual steps. Dense 3D hand poses were recovered from monocular and stereo video, stored in a tens-of-petabytes data lakehouse, and trained on hundreds of GPUs with over 50 % FLOPs utilization. At inference, a Rust orchestration layer and fused CUDA kernels maintain a 100-millisecond observation-to-action loop. Demonstrations include folding thin fabrics, writing calligraphy with a soft brush, and multi-stage tool switching.

Our read: Natura-1 shows a pragmatic path to physical intelligence by leveraging massive human video rather than robot-collected data, but the performance claims rest on internal benchmarks that have not been independently verified. The 100 ms latency and reported scores suggest feasibility for certain B2B automation scenarios, yet scaling to diverse industrial tasks will require broader validation and cost analysis.