Nature Will unveils Natura-1 robot model for dexterous hands
Chinese startup Nature Will released its Natura-1 robot model on October 13, pairing language and vision inputs with control of dexterous hands with 20 degrees of freedom, according to Pandaily. The company describes the 50 billion parameter mixture of experts model as its first general foundation model for the physical world, pretrained on more than one million hours of first person human video.
The training strategy addresses a constraint Nature Will identifies in robot learning: robots can collect experience only in settings where they already function, while human activity covers a much wider range of tools, objects and environments. After pretraining on human footage, the team adapts Natura-1 to robot hands using what it describes as a limited amount of robot manipulation data.
According to the company, its curated corpus contains one million hours of language conditioned action trajectories. The footage was deduplicated, screened for sensor anomalies and annotated at multiple levels, from overall goals to individual steps. The team also recovers dense 3D hand poses from monocular and stereo video, giving the training pipeline a representation of hand motion rather than relying on images alone.
Dexterous tasks and the benchmark comparison
Nature Will says Natura-1 can follow complex natural language instructions without task specific training, including spatial constraints, relationships among several objects and negative conditions. Demonstrations include folding thin, deformable materials, writing calligraphy with a soft brush and switching tools during tasks with multiple stages. Six long horizon tasks received additional tuning, the company says, while the other results use the same model weights.
On its own benchmark covering twelve dimensions of dexterous manipulation, Nature Will reports a score of 53.4 for the base model and 59.1 for an agentic setup, tested on dual arm robots with dexterous hands. The two baselines it selected could not produce scores with dexterous hands, according to the company; on a gripper platform, they scored 13.5 and 16.7. The different hardware makes those figures an uneven comparison rather than a direct measure of relative model performance. The results have not been independently reproduced.
The supporting infrastructure is substantial by the company’s account: a data lakehouse holding tens of petabytes and an in house distributed training engine that it says sustains more than 50% model FLOPs utilization across hundreds of GPUs. For physical execution, Nature Will says a Rust orchestration layer and fused CUDA kernels maintain a 100 millisecond loop from observation to action on robot hardware.
Source: pandaily.com
