Nature Will unveils Natura-1, a 50B MoE model pretrained on 1M hours of first-person human video to control 20-DoF robot hands (reports)

Verified (fact)

According to Pandaily, China's Nature Will has unveiled Natura-1, a foundation model for the physical world. It is a mixture-of-experts (MoE) model with 50 billion parameters. It was pretrained on 1 million hours of first-person video filmed from a human point of view, then adapted with a small amount of robot data to control a robot hand with 20 degrees of freedom (DoF). The report does not say how the model will be released or under what terms.

Why it matters (interpretation)

Robot data is hard and expensive to collect. Pretraining dexterous hand manipulation on large-scale human video is drawing attention as a way to ease this data bottleneck.

Sources

• Press Nature Will Unveils Natura-1, a 50B Model Trained on Human Video to Drive 20-DoF Dexterous Hands Pandaily