Nature Will - 自然意志

Nature Will
自然意志

Beijing physical AI lab founded 2026 whose Natura-1, a 50B-parameter model trained on 1M hours of human video, drives 20-DoF dexterous robot hands.

Visit Website

Company Overview

Nature Will is a Beijing physical AI research company founded in January 2026 that trains general-purpose foundation models for robot manipulation. Chinese press identifies the founder as Ning Ding, an assistant professor in the Department of Electronic Engineering at Tsinghua University. The company states that its aim is physical artificial general intelligence: models that close the loop between perception, prediction, reasoning and action in the real world, with continuous learning in physical environments.

Its first public product is Natura-1, described in technical report NW-TR-2026-001 published on 13 October 2026. Natura-1 is a single 50-billion-parameter mixture-of-experts model that takes language and vision as input and drives 20-degree-of-freedom dexterous robot hands within a 100-millisecond observation-to-action loop. It is pretrained on more than one million hours of egocentric human video, curated into language-conditioned hand-action trajectories, and is then post-trained on a limited amount of robot manipulation data to transfer that knowledge to a physical robot hand. The company argues that robots cannot yet generate physical experience at this scale, so human video is the main data source, and that anthropomorphic hands give the closest mapping from human to robot motion.

The company built its training stack in-house. Its data system manages tens of petabytes and processes hundreds of thousands of hours of egocentric video per week, with de-duplication, anomaly screening, recovery of 3D hand poses from monocular and stereo video, and hierarchical task annotation. Training uses a distributed engine with custom attention kernels that reports over 50 percent model FLOPs utilisation on hundreds of GPUs, and inference runs through a Rust orchestration layer with fused CUDA kernels to hold the 100 ms loop.

Evaluation was done on a dual-arm platform built from ROKAE AR5 arms and Wuji dexterous hands. On the company's 12-dimension dexterous manipulation benchmark, Natura-1 scored 53.4 in base mode and 59.1 in agentic mode with 20-DoF hands. On the same tasks, a fine-tuned pi0.5 policy and GPT-6-Astra used as a coding policy scored 13.5 and 16.7 with parallel grippers on an ARX AC One platform, and neither produced a score on the dexterous-hand platform. Published demonstrations include tray loading and carrying, wiping and disposing of a tissue, adding sugar and stirring, preparing a pot, folding deformable materials, brush calligraphy and making wontons, with the longest continuous task running more than three minutes. Six long-horizon tasks used extra tuning; all other results share one set of weights.

Nature Will does not make robot hardware and has not published model weights, an API, pricing, customers, headcount or funding. Its channels are X, YouTube, Bilibili, WeChat and Xiaohongshu.

Capabilities & Activities

Primary type & automation activities this supplier delivers:

Applications & Industries

Product Categories

Products & Solutions

Natura-1

50B-parameter model pretrained on human video that drives 20-DoF dexterous robot hands from language and vision.

Partnership & Notable clients

Associating Events

Listed in RobotToday Supplier Discovery. Verified Profile

More related news

  1. 1

Contact Nature Will

WEBSITE

https://www.naturewill.ai

HEADQUARTERS

Beijing (no street address published)
Beijing

China

Company Facts

Founded

2026

Primary Role

Software/Algorithm

Company Size

-

Primary Region

Asia-Pacific

Annual Sales

-

Funding Stage

-

Funding Total

-