Beijing's Nature Will Trains a 50-Billion-Parameter Model on 1 Million Hours of First-Person Human Video to Drive Robot Hands

A Beijing research lab called Nature Will has released Natura-1, an AI model that learns manual skills from first-person recordings of people and uses them to control robot hands. The lab was founded in January by Ning Ding, a tenure-track assistant professor in Tsinghua University's Department of Electronic Engineering who earned his doctorate in computer science there in 2023 with research in natural language processing and machine learning.

Natura-1 is a 50-billion-parameter mixture-of-experts model pretrained on more than one million hours of what the company calls egocentric human experience: video recorded from a person's own point of view. Nature Will says it built an internal pipeline that converts those recordings into labelled hand movements, then adapted the model with a comparatively small amount of robot data. The model takes language and vision as input and drives hands with 20 degrees of freedom within a 100-millisecond control loop.

The published demonstrations are unusually specific. In one of six longer workflows, a pair of robot hands packs a shipping box: laying in raffia, a plush toy and a book, folding down the springy flaps, taping them shut with one hand holding the flaps while the other pulls the tape, then applying a label. Nature Will notes that these six workflows received additional tuning on top of Natura-1, while its other results come from a single set of model weights. Shorter clips show the model folding thin materials, writing calligraphy and switching tools mid-task, and adjusting its grip in response to new visual and contact feedback. In one case the robot, unable to reach a ball of yarn, used a rod threaded through a basket to pull the basket closer; in another it pushed a playing card to the edge of a box before lifting the overhanging part. The company says those intermediate steps were never explicitly instructed and did not appear in its demonstrations.

The headline comparison pits dexterous hands against ordinary grippers. Nature Will evaluated Natura-1 on its own benchmark covering twelve capability dimensions, using tasks whose scenes, objects and action combinations do not appear in the model's robot training data. On a two-arm robot fitted with standard grippers, Physical Intelligence's pi-0.5 scored 13.5 and OpenAI's GPT-6-Astra, set up to write control code for the robot, scored 16.7. Natura-1, driving 20-degree-of-freedom hands, scored 53.4, and 59.1 in an agentic mode. On the lab's own two-arm test robot built from ROKAE arms and Wuji hands, neither rival produced a score at all: pi-0.5 failed even after being fine-tuned on Nature Will's hand data, and GPT-6-Astra did not finish within the 60-minute limit per task. The company acknowledges the comparison figures therefore come from different hardware, and all of the results are self-reported and have not been independently verified.

The approach fits an emerging consensus in robotics rather than breaking from it. NVIDIA's GR00T N1, released in March 2025 as an open foundation model for humanoid robots, was trained on data that included first-person human video alongside real and simulated robot movements. In December 2025, Physical Intelligence reported that adding first-person human video during fine-tuning roughly doubled performance across a set of four generalization scenarios. On the same day Nature Will released Natura-1, Physical Intelligence opened a waitlisted beta of an API that lets outside developers fine-tune its latest model on managed compute and run the results on their own robots.

Ding's thesis is blunt. "We believe general models that can handle every task must exist, and that those models must utilize dexterous hands," he said in the company's announcement. In its technical report, Nature Will calls Natura-1 "early evidence" that general models can work with dexterous hands, and says further progress depends on better hand and sensing hardware, more pretraining data, and models that transfer to robots with less adaptation.

The bet is easy to state and hard to test: that hands plus human video are the scaling path to general-purpose manipulation, in the same way internet text was the scaling path for language. What is missing, as with most releases in this wave, is third-party evaluation. Until someone reruns Natura-1's benchmark outside the company that wrote it, the 53.4 against 13.5 remains a company claim rather than an industry result.