A team of scientists from Stanford University has unveiled a new humanoid robot called Humanplus, capable of replicating complex sequences of human movements in real-time. The project is open-source, allowing anyone to build their own robot using the provided data, code, and list of necessary components.
According to the creators, one of the key advantages of humanoid robots is the ability to utilize extensive data on human movements for training purposes. If sufficient data is not available, it can be gathered through a camera and actor. However, there is currently no system that can quickly adapt human movements to robotic mechanisms.
Initially, the team compiled 40 hours of video recordings depicting various human tasks and used NVIDIA simulation to train the model. This data enabled the robot to analyze and replicate human movements in real-time from the camera feed.
Humanplus has the unique ability to memorize movements from the camera and reproduce them independently after about 40 repetitions. The robot can recall movements with a success rate of 60-100%, but its performance is currently limited by computing resources.
The Humanplus project is open to the public, with scientists publishing the work text, source code, dataset, and a list of components required for robot assembly. The Unitree H1 robot serves as the base model, which has been modified for this project. The estimated cost of the components is around $108,000, excluding training equipment costs.
Coinciding with the Humanplus announcement, entrepreneur Elon Musk revealed plans to deploy over 1,000 Optimus humanoid robots at Tesla factories by year-end. Musk aims to start mass sales of Optimus next year, envisioning Tesla evolving into a $25 trillion business with the humanoid robot as a cornerstone product.