Ricoh develops physical AI technology that learns human movements from video

Ricoh announced that it had developed physical AI technology that learns actions as a common representation from videos of people and robots performing tasks. The company said it used a support program from AWS Japan and confirmed the technology’s effectiveness through simulations in a virtual environment.

The technology’s key feature is that it can learn actions such as “grasping,” “moving forward” and “transferring an object” in the same way from videos of human tasks and robot movements. Ricoh says the approach does not depend on the mechanics or body structure of individual robots. For developers seeking to have robots with different structures perform tasks, the technology offers a basis for understanding how to use video data and what assumptions to make when evaluating it.