Carnegie Mellon Develops Pixel Humanoid for Under $15,000

PITTSBURGH -- Scientists at Carnegie Mellon University have created Pixel, a wheeled humanoid robot intended to learn household tasks by observing humans perform them, with hardware costs estimated below $15,000.

The research group outlined their project Nov. 2. Their approach combines the development of hardware, data gathering, motion control, and artificial intelligence training processes into one, and does not consider them as individual components. The estimate accounts only for the robot's hardware cost, and not its commercial selling price.

Pixel consists of two robotic arms, compliant grippers, and a wheeled platform. Its body allows the robot to mimic human motion and reach, which allows transferring of physical tasks demonstration to robot tasks easier.

The experts used different approaches for training data collection, among which remote robot control, and recording human motions without the need for a robot. Both of these demonstrations can be used simultaneously in the same AI training process. This includes full-body movement coordination in addition to AI models interpreting the instructions and visual data to complete multi-stage tasks.

The report noted that on an average, AI control policies that learned from human demonstrations had a task completion efficiency of 82% in these activities. This is the extent to which these tasks were completed, not how many tasks could be completed out of these.

Pixel project remains in the research stage. There have been no announcements about its availability to customers.