Natural Will released Natura-1, a 50-billion-parameter robotics model. It was first pretrained on one million hours of human first-person data, then post-trained on a small amount of robot data to adapt it to real-world dexterous manipulation.
The system announced by the team accepts language and visual inputs and controls a 20-degree-of-freedom dexterous hand, with less than 100 milliseconds of latency from observation to action output.
In the team’s tests, increasing the pretraining data from around 44,000 hours to one million hours reduced the mean squared error of action predictions on a real-robot validation set by 67.7%. The test used existing observation data to predict actions, with no additional training during evaluation, and measured action prediction error.
The system announced by the team accepts language and visual inputs and controls a 20-degree-of-freedom dexterous hand, with less than 100 milliseconds of latency from observation to action output.
In the team’s tests, increasing the pretraining data from around 44,000 hours to one million hours reduced the mean squared error of action predictions on a real-robot validation set by 67.7%. The test used existing observation data to predict actions, with no additional training during evaluation, and measured action prediction error.